# Flowing — documentation, marketing, community notes & FAQ (full text) > Desktop academic writing environment that keeps manuscript context and verifiable passages from your local PDF library in the AI writing loop. Machine-readable export of public documentation, comparisons, topic guides, blog posts, community notes, and FAQ. Screenshots are omitted; see linked pages for figures. - **Summary:** https://flowing.works/llms.txt - **Chinese summary:** https://flowing.works/llms-zh.txt - **This file:** https://flowing.works/llms-full.txt - **Documentation index:** https://flowing.works/en/docs ## Independent feature and coverage Ness Labs selected Flowing as a **Featured Tool** and published a founder interview in its Tools for Thought series about evidence-grounded AI writing. Source: https://nesslabs.com/flowing-featured-tool ## Stable definition of Flowing Flowing is not a general chatbot, reference manager, or paper discovery engine. It is a desktop academic writing environment that keeps manuscript context and verifiable passages from the researcher's own PDF library in the AI writing loop. ### Two kinds of grounding - **Reference-evidence grounding:** whether the source passage supports the claim, mechanism, causal verb, and degree of certainty; this prevents a plausible mechanism from being presented as a supported mechanism. - **Manuscript-context grounding:** whether the next sentence matches the discourse role and narrative at the current cursor position; this prevents locally fluent prose from summarizing too early, repeating a conclusion, or drifting from the argument. Reference evidence constrains what the text may responsibly say. Manuscript context determines what the text should say here. Flowing uses both. ### Fast, evidence-backed autocomplete Pause while writing, then click the sparkle beside the cursor or press Tab to get a gray inline preview of the next sentence in seconds—without writing a prompt or opening a chat. Flowing reads the local argument around the cursor and can match relevant passages from the researcher's selected PDF library. Manuscript context determines what belongs next. When library material is used, matched source passages remain available for inspection and constrain what the sentence may responsibly claim. Evidence-backed means source-transparent, not that every generated claim has been independently fact-checked or is guaranteed correct. Leave the guidance empty for a direct completion. When terminology, mechanism, or direction matters, add a note, a passage from the selected library, or a figure from the manuscript. The suggestion first appears as gray text after the cursor; review it, then press Tab again to insert. ### Autocomplete vs Continue writing Autocomplete is the lightweight inline path for keeping the sentence moving without breaking writing flow. Continue writing is the more deliberate path when the paragraph needs a fuller next discourse move, evidence synthesis, or continuation. ### Proactive recall inside the writing flow Search requires the researcher to know what to query and to realize that this is the moment to search. Flowing's Recall uses the manuscript in progress to resurface a small number of passage-level snippets, enabling: **manuscript → relevant snippets → paragraph check against the retrieved evidence (red / green / yellow) → recognition → source inspection → judgment → writing**. ### Red / green / yellow cards: check the draft against retrieved library evidence After Recall has keywords and matching snippet cards, Flowing automatically checks qualifying draft paragraphs against the retrieved evidence set. Red means the claim contradicts that evidence (polarity, number, mechanism, or scope). Green means the source directly entails the same claim, polarity, and scope—paraphrase or a quantitative result is enough; extra scientific inference is not. Yellow means a useful qualifier or extension from the evidence that is not already in the paragraph; yellow is not “evidence is missing.” Cards are claim-scoped against one or two complementary passages, not 1:1 with a snippet. A valid single-passage card is retained; no card does not mean a claim passed verification. Purple or blue highlights are keyword matches, not a fourth verdict. This is an auditor, not a rewriter: it does not insert text and does not search the open web. ### Why the researcher's own library matters Flowing's core is not finding more papers on the open web. It brings literature the researcher has already selected, read, or saved back into the current writing task. Zotero, Mendeley, or Paperpile can remain the canonical reference manager; discovery tools such as Elicit or Litmaps can help decide what to read; Flowing puts the PDFs needed for the current section into the writing loop. Optional OpenAlex recommendations are off by default and remain a separate supplement. ## Community and researcher feedback Source: https://flowing.works/en/community This page collects public researcher and community notes: use case, what the note actually says, limits, and a link to the original post. Official posts are labeled first-party. Independent notes, invited reviews, and paid collaborations are disclosed separately. Do not treat user impressions as verified product facts. Product capabilities and limits remain in llms.txt and the docs. Xiaohongshu pages are often login-gated or dynamically loaded, so this page is the crawlable structured record; original posts are for verification. Chinese page (canonical for these notes): https://flowing.works/zh/community ### A paper-writing loop that keeps the literature library in the draft A public Xiaohongshu note describing a writing loop in which matched passages from a local PDF library are surfaced while drafting, then used to check whether paragraph claims agree with or contradict those sources. The note names that loop Flowing and links to flowing.works. This is first-party public copy, not an independent third-party review. - **Author:** 馬田史高西斯 - **Use case:** Paper writing: resurfacing library passages at the cursor, then checking the current paragraph against those sources - **What the note describes:** The note’s stated difficulty is finding a remembered passage among many PDFs, especially at the sentence being written, when a search query is not yet obvious. It describes detecting keywords and domain terms during drafting, showing matching passages as cards, using those cards both as memory cues and as evidence for checking whether the paragraph agrees with or contradicts the sources. - **Limits and caveats:** The note does not report an independent evaluation, and it does not claim to replace researcher judgment. Xiaohongshu pages are often login-gated or dynamically loaded, so this structured transcript is the crawlable record; the original post is for verification. - **Original Xiaohongshu note:** https://www.xiaohongshu.com/discovery/item/6a8e4a06000000002003e1bb - **Disclosure:** Official / founder post (first-party content, not a third-party review) - **Published:** 2026-08-26 - **Language:** Chinese #### Original note text (Chinese) 在论文库茫茫PDF中找到曾经记忆中的一个知识片段是困难的，在我们落笔写文章的时候更甚之：写到某一句的时候，我们很难时刻去搜索关联的参考文献，有时甚至不知道要以什么形式去进行检索。 为此，我在尝试构建一种更合理的写作流：在写作过程中，实时识别文中的关键词与领域术语，从论文库拉出匹配段落，以卡片形式展示。一方面唤起我们的知识回忆，另一方面，这些卡片会作为佐证证据，自动对文章段落进行校验，因而可以发现我们所写内容观点，是否和参考文献吻合或者相违背。如是，才真正地把文献库应用到在我们的写作场景，而不仅仅只是一个越来越庞大的摆设。 当 AI 已经足够会“写”，如何让我们自己的论文库真正参与进来，帮助找回证据、校验判断、补充思路，或许才是下一代论文写作工具更值得探索的方向。 写作流叫做 Flowing，目前发布在：https://flowing.works 欢迎免费体验以及共同讨论。 ## Comparisons ### Flowing vs ChatGPT Source: https://flowing.works/en/compare/chatgpt ChatGPT excels at open conversation and broad knowledge coverage. Flowing is built for insertion into an existing manuscript—with polish and continue writing informed by your local PDF library, draft context, and verifiable evidence cards. Its design priority is context fit and review narrative, not showing how much the model knows. | Aspect | Flowing | ChatGPT | | --- | --- | --- | | Primary purpose | Long-form academic writing with your reference library | General-purpose chat and assistants | | Source library | Local PDF import with keyword search across your research library | No built-in local PDF library | | Grounding | Recall, Ask, polish, and continue writing attach library snippets | User must paste context manually; no native citation cards | | Citation verification | After Recall, red / green / yellow cards check draft claims against retrieved library evidence you can open | Claims may be plausible but unverified without manual checks | | Editing environment | Rich text + LaTeX with live PDF preview | Chat interface; export is manual | | Continue writing | Adds the next discourse move your paragraph needs—few representative studies, then synthesis | Often expands into a mini-survey—more papers, mechanisms, and scope drift | | Polish / revision | Improves discourse function—argument flow, claim calibration, accuracy—with library-backed cards | Often rewrites for elegance without respecting local role (e.g., bridge vs summary) | | Context fit | Stronger when you need the next sentence in your draft—not a standalone survey paragraph | Broader coverage can read well alone but disrupt the paragraph you are building | | Best for | Literature reviews, theses, source-faithful drafts | Brainstorming, summaries, general Q&A | **Continuation, not generation** Generic AI tends to show everything it knows—five papers, every mechanism, topics that drift from your section. Flowing’s continue writing stays on your thread: a few representative examples, a synthesis line, then back to your argument. That is what top-tier review prose expects. **Polish as revision, not paraphrase** A good suggestion makes your academic argument easier to follow—not merely more elegant. With local evidence, neighboring paragraphs, and accuracy review, Flowing’s polish targets discourse function: claim calibration, scope, and how evidence is organized—not synonym swapping. **When ChatGPT is still the right tool** Use ChatGPT for quick explanations, coding help, or a first draft when you will reorganize later. Use Flowing when the sentence must land in the exact place in your manuscript—with sources you can open on the card. ### Flowing vs Zotero Source: https://flowing.works/en/compare/zotero Zotero organizes citations and bibliographies brilliantly. Flowing is not a Zotero replacement—it is where you draft long-form text with AI anchored to PDF passages you import. Many researchers use both: Zotero for the canonical library and bibliography, Flowing for the writing loop. | Aspect | Flowing | Zotero | | --- | --- | --- | | Primary job | Draft and revise long-form academic text with grounded AI | Collect, organize, and cite references; Word/LibreOffice integration | | PDF handling | Import PDFs into a searchable library for recall and AI context | Store attachments; full-text search via plugins or linked files | | AI writing | Native recall, Ask with context, polish, and continue writing with evidence cards | Not a writing-first AI environment; plugins vary by setup | | Citation output | Focus on drafting with verifiable in-text evidence from your library | Industry-standard bibliography styles and citation insertion | | Where it runs | Desktop app (macOS, Windows) with integrated editor | Desktop connector + browser extension; syncs library metadata | | Best together | Import the PDFs you need for the chapter you are writing today | Keep the canonical library and final reference list for submission | **Complementary, not either-or** Zotero answers “what did I read and how do I cite it?” Flowing answers “how do I draft this section with passages I can verify?” Many workflows keep Zotero for bibliography management and Flowing for the writing loop. **Fewer ghost citations while drafting** When AI suggests a claim, Flowing attaches library snippets you can open immediately—reducing the “citation looks real but the paper doesn’t say that” problem during drafting, before you finalize references in Zotero. **When Zotero alone is enough** If you only need reference storage, PDF annotation, and formatted bibliographies without AI-assisted drafting, Zotero may be sufficient. Add Flowing when you want source-grounded writing assistance in the same workspace as your manuscript. ### Flowing vs Litmaps Source: https://flowing.works/en/compare/litmaps Litmaps helps you discover and visualize how papers connect. Flowing brings passages from selected PDFs into the manuscript workflow so you can recall evidence, ask questions, revise, and continue writing without leaving the draft. | Aspect | Flowing | Litmaps | | --- | --- | --- | | Core strength | Write, recall, and revise with library-grounded AI | Map citation networks and discover related papers | | Output | Manuscript paragraphs with evidence cards | Graphs, timelines, and reading lists for exploration | | Library model | Local PDF import optimized for search and AI attachment | Connected papers via databases and citation links | | AI role | Ask with context, polish, and continue writing on your draft | Discovery and visualization; not a full writing environment | | Verification | Open the cited passage on the evidence card | Trace relationships between papers; verify claims in your draft separately | | Typical workflow | Prepare library → write section → recall → synthesize → polish | Seed paper → expand map → prioritize reading → export reading list | **Discovery vs drafting** Litmaps shines when you need to see the landscape—what cites what, what you might have missed. Flowing shines when you are ready to write and need every sentence to stay accountable to sources you have actually imported. **From map to manuscript** A practical split: use Litmaps (or similar tools) to decide what to read next; import the PDFs you need into Flowing and draft the literature review with keyword recall and Ask with context. **When Litmaps is the better fit** Early-stage scoping, systematic mapping, and visual exploration of a field. Choose Flowing when the bottleneck is writing coherent, cited prose—not finding the next paper to open. ### Flowing vs Obsidian Source: https://flowing.works/en/compare/obsidian Obsidian excels at linking notes, markdown, and personal knowledge graphs. Flowing is not a note-app replacement—it is where you draft citable manuscripts with AI suggestions anchored to PDFs in your research library. Many researchers capture ideas in Obsidian, then write source-faithful prose in Flowing. | Aspect | Flowing | Obsidian | | --- | --- | --- | | Primary job | Draft and revise long-form academic manuscripts with library-grounded AI | Personal knowledge management, note linking, and markdown vaults | | PDF library | Import reference PDFs; keyword recall and evidence cards across your library | Markdown notes and attachments; academic PDF workflows rely on plugins | | AI writing | Native recall, Ask with context, polish, and continue writing with evidence cards | Community plugins and Copilot-style extensions; no built-in evidence cards from a PDF library | | Evidence verification | Open the cited passage on the card before accepting any suggestion | Notes and links; verify claims against source PDFs manually | | Editing environment | Rich text + LaTeX with live PDF preview for submission-ready drafts | Markdown-first editor; LaTeX export via plugins or external tools | | Best together | Import the PDFs you need for the chapter you are writing today | Capture reading notes, ideas, and links in your vault between writing sessions | **Notes vs manuscripts** Obsidian answers “how do I connect what I know?” Flowing answers “how do I write this section with passages I can verify?” Many researchers keep Obsidian for ideation and reading notes, then draft citable prose in Flowing. **Evidence cards, not just links** Backlinks and wikilinks help you navigate notes. Flowing’s Recall returns excerpt cards and page thumbnails; evidence-bearing polish and continuation results expose the library passages used, so you can inspect claims before they enter your manuscript. **When Obsidian alone is enough** If you only need a linked note vault, daily journaling, or lightweight markdown without AI-assisted academic drafting, Obsidian may be sufficient. Add Flowing when the bottleneck is writing source-faithful literature reviews and thesis chapters. ### Flowing vs Notion Source: https://flowing.works/en/compare/notion Notion is a flexible workspace for docs, wikis, and team projects—with AI across blocks. Flowing is not a Notion replacement: it is purpose-built to keep an active manuscript and passages from imported research PDFs in the same AI writing loop. | Aspect | Flowing | Notion | | --- | --- | --- | | Primary job | Academic writing with verifiable library evidence | General workspace—docs, databases, wikis, project management | | Source library | Local PDF import with keyword search across your research library | Cloud pages and databases; no native local PDF research library | | AI grounding | Recall, Ask, polish, and continue writing attach library snippet evidence cards | Notion AI works on page content; no built-in PDF passage verification | | Citation verification | Evidence cards link to imported passages on your machine | AI output may be plausible but unverified against your actual sources | | Editing environment | Rich text + LaTeX with live PDF preview | Block-based editor; LaTeX and academic formatting require workarounds | | Data model | PDFs stay local; only attached context sent for AI requests | Cloud-hosted workspace with team sharing | **Workspace vs writing instrument** Notion shines when you need shared docs, project boards, and flexible databases. Flowing shines when you need the next sentence in your literature review—with representative evidence from papers you actually imported, not a generic summary. **Evidence-based workflow, not empty chat** Notion AI can rewrite blocks and answer questions about your pages. Flowing’s difference is structural: Recall returns library passage cards, while evidence-bearing polish and continuation results show the passages you can audit before accepting. **When Notion is the better fit** Team wikis, lab onboarding docs, grant admin trackers, and lightweight notes. Choose Flowing when you are drafting a thesis chapter or review article where reviewers will check your sources. ### Flowing vs Mendeley Source: https://flowing.works/en/compare/mendeley Mendeley (Elsevier) organizes references, PDFs, and annotations for researchers. Flowing is not a Mendeley replacement—it is where you draft with AI anchored to passages from PDFs you import. Many researchers keep Mendeley for library management and Flowing for the writing loop. | Aspect | Flowing | Mendeley | | --- | --- | --- | | Primary job | Draft and revise long-form academic text with grounded AI | Reference management, PDF storage, annotations, and citation plugins | | PDF handling | Import PDFs into a searchable library for recall and AI context | Store and annotate PDFs; sync library across devices | | AI writing | Native recall, Ask with context, polish, and continue writing with evidence cards | Not a writing-first AI environment; focus on reference workflows | | Citation output | Focus on drafting with verifiable in-text evidence from your library | Citation styles and Word/LibreOffice integration via Mendeley Cite | | Where it runs | Desktop app (macOS, Windows) with integrated editor | Desktop app + web library; Elsevier ecosystem integration | | Best together | Import the PDFs you need for the section you are writing today | Keep the canonical library and final reference list for submission | **Complementary, not either-or** Mendeley answers “what did I read and how do I cite it?” Flowing answers “how do I draft this section with passages I can verify?” Many workflows keep Mendeley for reference storage and Flowing for source-grounded writing. **From annotation to evidence-backed prose** Mendeley highlights help you remember what mattered in a paper. Flowing turns those same PDFs into active context—keyword recall, Ask with context, and continue writing that attach excerpt cards you can open before accepting text. **When Mendeley alone is enough** If you only need reference storage, PDF annotation, and formatted bibliographies without AI-assisted drafting, Mendeley may be sufficient. Add Flowing when you want evidence-backed writing assistance in the same workspace as your manuscript. ### Flowing vs Elicit Source: https://flowing.works/en/compare/elicit Elicit helps researchers find papers, extract claims, and answer questions across the literature. Flowing serves the next stage: bring passages from selected PDFs and the manuscript itself into the writing loop for recall, revision, and continuation. | Aspect | Flowing | Elicit | | --- | --- | --- | | Primary job | Draft and revise long-form academic manuscripts with library-grounded AI | Discover papers, extract data, and answer research questions across literature | | Source model | Your imported local PDF library—passages you choose and can verify on the card | Broad literature search and extraction; not a local manuscript workspace | | Output | Manuscript paragraphs with evidence cards you open before accepting | Tables, summaries, and answers for scoping—not integrated long-form drafting | | AI writing | Recall, Ask with context, polish, and continue writing in one editor | Research Q&A and extraction; writing happens elsewhere | | Verification | Evidence cards link to imported passages on your machine | Citations to discovered papers; verify claims in your draft separately | | Best together | Import the PDFs you need for the section you are writing today | Scope a field, shortlist papers, then export a reading list | **Discovery vs drafting** Elicit shines when you need to find what the field says—systematic extraction, comparison tables, and quick answers across many papers. Flowing shines when you are ready to write and need every sentence accountable to sources you have actually imported. **From extraction to evidence-backed prose** A practical split: use Elicit (or similar tools) to decide which papers matter and what they claim; import those PDFs into Flowing and draft the literature review with keyword recall, Ask with context, and continue writing that attach excerpt cards. **When Elicit is the better fit** Early-stage scoping, rapid evidence tables, and answering “what do we know about X?” across hundreds of abstracts. Choose Flowing when the bottleneck is writing coherent, cited prose in your manuscript—not finding the next paper to read. ### Flowing vs Paperpile Source: https://flowing.works/en/compare/paperpile Paperpile organizes references, PDFs, and citations—with strong Google Docs and Word integration. Flowing is not a Paperpile replacement—it is where you draft with AI anchored to passages from PDFs you import. Many researchers keep Paperpile for bibliography management and Flowing for the writing loop. | Aspect | Flowing | Paperpile | | --- | --- | --- | | Primary job | Draft and revise long-form academic text with grounded AI | Reference management, PDF library, and citation insertion | | PDF handling | Import PDFs into a searchable library for recall and AI context | Store PDFs in cloud library; annotate and cite in Google Docs or Word | | AI writing | Native recall, Ask with context, polish, and continue writing with evidence cards | Not a writing-first AI environment; focus on references and citations | | Citation output | Focus on drafting with verifiable in-text evidence from your library | Industry-standard bibliography styles and in-document citation plugins | | Where it runs | Desktop app (macOS, Windows) with integrated editor | Web app + browser extension; Google Docs and Word add-ons | | Best together | Import the PDFs you need for the chapter you are writing today | Keep the canonical library and final reference list for submission | **Complementary, not either-or** Paperpile answers “what did I read and how do I cite it in Google Docs?” Flowing answers “how do I draft this section with passages I can verify?” Many workflows keep Paperpile for bibliography management and Flowing for the writing loop. **Fewer ghost citations while drafting** When AI suggests a claim, Flowing attaches library snippets you can open immediately—reducing the “citation looks real but the paper doesn’t say that” problem during drafting, before you finalize references in Paperpile. **When Paperpile alone is enough** If you only need reference storage, PDF annotation, and formatted bibliographies in Google Docs without AI-assisted drafting, Paperpile may be sufficient. Add Flowing when you want evidence-backed writing assistance in a dedicated academic editor. ## Topic guides ### Literature review with Flowing Source: https://flowing.works/en/topics/literature-review A literature review is not a list of summaries—it is an argument about what the field shows, where it disagrees, and what gap you address. Flowing keeps that argument anchored to PDFs you import. When you need an evidence paragraph next—not a field survey—continue writing with local library support tends to read like Nature Reviews: a few representative studies, tight synthesis, and no scope drift. - Import your core reference PDFs into a local library—search and AI features revolve around papers you choose. - Keyword-grounded recall surfaces highly relevant passages while you draft each section. - Ask with context lets you synthesize across attached draft and library passages in a few clicks. - Evidence-backed continue writing adds what the paragraph needs now—not every paper the model has seen. - Basis-backed polish revises discourse function (evidence introduction, comparison, synthesis)—not just wording. - Rich text and LaTeX editing with live PDF preview keeps the review manuscript in one workspace. #### A workflow built for synthesis Start by importing the PDFs that matter for your review chapter—not necessarily your entire Zotero vault, but the papers you will actually cite in this section. Flowing indexes them locally so keyword recall can match terms in your draft to passages across those files. As you write, use recall to see candidate evidence cards, attach the best passages as context chips, and ask focused questions: How do these three studies disagree? What limitation do they share? Flowing’s answers draw on the passages you attached—not the open internet. #### Continuation vs. knowledge dumping A common failure mode with generic AI: the next paragraph is academically accurate and eloquent—but it covers too much, cites too many papers, and wanders from the section you are writing (e.g., from laser gain media into photovoltaics). Reviewers call that an annotated bibliography, not narrative. Flowing’s continue writing is optimized for insertion: maintain the author’s rhythm, add only the evidence the current move requires, and return to a synthesis line. That restraint matches how top-tier reviews read—unlike dumping every related paper into one paragraph. #### Staying faithful while moving fast Literature reviews fail when speed trades off against accuracy—unsupported claims, misread findings, or citations that look real but do not exist. Flowing reduces that risk by linking AI suggestions to library snippets you can open on the card before accepting text into your manuscript. Basis-backed polish and evidence-backed continue writing extend the same principle: rewrites and new sentences should strengthen your argument with references you can check, not invent. #### Who this is for Thesis literature review chapters, survey articles, grant background sections, and any long-form synthesis where reviewers will check your sources. If you only need a quick unstructured summary of one paper, a general chat tool may suffice; if you need a citable chapter, Flowing is the better fit. ### Avoid fake citations with source-grounded AI Source: https://flowing.works/en/topics/fake-citations Large language models can generate citations that look perfect—correct journal, realistic author list, convincing year—and still be wrong. Flowing does not eliminate human judgment, but it shifts AI output toward passages from PDFs you imported, so you verify claims against real text. - Fake citations (hallucinated references) are a known risk when AI writes without access to your actual sources. - Flowing attaches library snippets to recall, Ask, polish, and continue-writing suggestions. - After Recall, qualifying draft paragraphs are checked against the retrieved evidence set with red / green / yellow cards. - Red contradicts the retrieved evidence; green means the source directly entails the same claim, polarity, and scope; yellow is a useful qualifier not already in the paragraph. Same-topic overlap is not support, and missing support is not a yellow card. - Evidence cards let you open the cited passage in your imported PDF before accepting text. - Your PDF library stays local; only attached context for a request is sent for AI processing. - Human review remains essential—Flowing makes verification faster, not optional. #### What “fake citation” means in practice The failure mode is not always a completely invented paper. Sometimes the paper exists but the model misstates the finding, quotes a number that is not in the PDF, or blends two studies into one claim. Generic chat makes these errors easy because there is no built-in link between the sentence and a specific passage you can audit. Researchers on forums often call these “ghost citations”—entries that look fine in the draft but collapse when you search your library or open the PDF. #### How Flowing changes the verification loop When you use keyword recall or Ask with context, Flowing works from passages in your imported library. Suggestions arrive as cards tied to those snippets. Before you accept a rewrite or a continuation, you can jump to the source text and confirm the claim matches what the authors actually wrote. After Recall has snippet cards, qualifying paragraphs in the draft itself are also checked. Red, green, and yellow cards mark whether a claim contradicts the retrieved evidence, is directly entailed at the same polarity and scope, or still lacks a useful qualifier the evidence supplies. That is a tighter test than “the paper exists” or “the paper is on the same topic.” That loop—suggest, link, verify, accept—is slower than blind copy-paste from chat, but far faster than hunting down a hallucinated reference after your supervisor marks up the draft. #### Good habits still matter Source-grounded AI is not a substitute for your own reading. Import the papers you will rely on, attach the right context chips, and keep final bibliography management in the tool you trust (many researchers pair Flowing with Zotero). Use Flowing when citations must survive scrutiny; use open chat when you are brainstorming without commitments to specific sources. ### Source-grounded AI writing Source: https://flowing.works/en/topics/source-grounded-writing Flowing combines two kinds of grounding that solve different problems. Reference-evidence grounding ties claims to inspectable text from selected PDFs—and, after Recall, red / green / yellow cards check the draft itself against the retrieved evidence set. Manuscript-context grounding keeps a revision or continuation aligned with the discourse role and narrative at the current cursor position. - Flowing is a desktop writing app—not a browser chat window with no library. - Prepare → Write → References → Recall → Ask with context → Polish → Continue writing is the core loop. - Keyword recall matches draft terms to passages across your imported PDFs. - After Recall, qualifying paragraphs are checked against the retrieved evidence: red contradicts, green marks direct entailment of the same claim, yellow flags a useful qualifier. - Context chips attach draft or library passages to each AI request. - Reference evidence and manuscript context are distinct inputs; Flowing keeps both in the writing loop. - Continue generates new discourse; polish optimizes existing discourse—they are not the same task. - Polish targets discourse function (argument flow, claim calibration)—not synonym swapping. #### Two kinds of grounding Reference-evidence grounding asks whether a source passage supports the claim, mechanism, causal verb, and certainty level. Manuscript-context grounding asks whether the next sentence performs the right discourse role at this exact cursor position. The first constrains what the text may responsibly say; the second determines what the text should say here. Grounding does not make interpretation automatic or human review optional. Recall exposes matched passages, Ask works from context chips you attach, and evidence-bearing polish or continuation results show passages to inspect. Optional OpenAlex recommendations are a separate, off-by-default discovery aid. #### Red, green, and yellow cards on the draft After Recall has highlighted keywords and matching snippet cards, Flowing automatically checks qualifying paragraphs against the retrieved library evidence. A left-gutter mark opens “Check paragraph against library evidence”: the paragraph on the left, cards on the right. This is not the same as evidence panels on polish or continue-writing suggestions—those attach sources to an AI rewrite. Here the object under review is the sentence already in the draft. Cards are claim-scoped against the retrieved evidence set (one or two complementary passages), not 1:1 with a snippet. Red means the claim contradicts that evidence in polarity, number, mechanism, or scope. Green means the source directly entails the same claim, polarity, and scope—paraphrase or a quantitative result is enough; extra scientific inference is not. The underline is the most complete contiguous claim that is fully entailed. Yellow is a useful qualifier or extension that is not already in the paragraph; it is not “the evidence is missing.” Purple highlights mark keyword matches; they are not a fourth verdict. The system does not rewrite the paragraph or search the open web. #### Two principles behind polish and continue Continue writing generates new discourse—e.g., introduce representative studies after a claim, then synthesize. Polish optimizes discourse you already wrote—organizing evidence, calibrating claims, fixing scope drift. Deep polish with accuracy review is closer to academic revision than “make it sound better.” Flowing’s prompt design follows: prefer improving a paragraph’s discourse function over improving its wording. A successful suggestion makes your argument easier to follow—not merely more elegant. #### The six-step workflow Prepare by importing reference PDFs. Write in rich text or LaTeX with live preview. Recall surfaces relevant library passages for keywords in your draft. Ask with context for synthesis or clarification on attached passages. Evidence-grounded revision improves paragraphs with attached sources. Continue writing extends your thread with cited evidence—not boilerplate. Each step keeps the manuscript and your sources in one workspace, so you spend less time switching between PDF reader, chat tab, and editor. #### When to choose source-grounded writing Choose this approach when the deliverable will be reviewed for accuracy—journal submissions, theses, grant proposals, and literature reviews. Skip it when you only need unstructured brainstorming with no citation obligations. Flowing is optimized for the first case. See our literature review and fake citations guides for field-specific workflows. ### Polish, revision, and continue writing Source: https://flowing.works/en/topics/polish-and-continue With your PDF library, draft text, neighboring paragraphs, and library-backed accuracy review, Flowing’s polish and continue writing are designed around a manuscript-specific goal: fitting the next sentence to your narrative, not merely improving vocabulary. - Continue generates new discourse. Polish optimizes existing discourse—they are different tasks. - Continue may add representative evidence; polish organizes evidence already in play—it should not invent a new study out of scope. - Prefer improving discourse function over improving wording: make the argument easier to follow, not merely more elegant. - Generic chat can sound polished in isolation—the same summary sentence can be perfect at the end of a section and wrong in the middle. - Local library snippets + draft context + content review = tighter review narrative and fewer ghost claims. #### Where the advantage shows up The gap is clearest in continue writing: Flowing tends to add a few representative studies and a synthesis line that follows your thread. Generic chat more often stacks extra papers, mechanisms, and adjacent topics—accurate prose that still does not belong at that cursor position. That is intentional. Flowing is optimized for continuation into an existing manuscript, not for generating a standalone survey paragraph you still have to cut down. #### Polish as academic revision Nature authors say “revise this paragraph,” not “polish my paragraph.” Flowing’s evidence-grounded revision—and deeper accuracy review—targets claim calibration, terminology conflicts, and scope drift: revision moves, not paraphrase. In a literature review, local discourse moves matter: evidence introduction (For example…), comparison (In contrast…), synthesis (Collectively…), transition (Building on…). The right revision depends on which move you are in—not just which sentence is selected. #### Library as field writing convention Your imported corpus is more than a terminology database. Across many reviews in the same field, conventions emerge—how studies are grouped, when to use Collectively versus These studies indicate, how many examples belong in one paragraph. Flowing draws on that corpus when you attach library context, nudging output toward how the field actually writes—not generic LLM prose. Pair that with evidence cards you can open before accepting text, and polish/continue become accountable to sources—not performances of how much the model knows. #### A three-layer mental model Grammar correction sits at the bottom. Academic revision (local polish with library and accuracy review) sits in the middle. Continue writing—generating the next discourse unit—sits on top. Flowing focuses on the top two layers for serious drafts, where generic chat most often breaks narrative rhythm. If your goal is the next sentence in your thesis or review—not a brainstorm you will rewrite from scratch—start from local evidence and let continue/polish respect your thread. ## Blog ### The Bottleneck Has Moved: What Matters When AI Can Already Write Source: https://flowing.works/en/blog/the-bottleneck-has-moved Fluent academic prose is becoming cheap. The harder problem is deciding what should be written in the first place. Fluent academic prose is becoming cheap. The harder problem is deciding what should be written in the first place. Not long ago, one of the hardest things for an AI writing tool to do was simply produce a paragraph that sounded as if it belonged in a paper. That problem has mostly changed. Give a strong model a paragraph from a manuscript today and it can usually continue in competent academic English. The terminology will often be right. The transition will be smooth. The sentence may even sound exactly like something the author could have written. And yet it can still be the wrong sentence. We ran into this repeatedly while looking at continuations from a chemistry manuscript on metal-halide perovskites. At one point, the manuscript was discussing Mn²⁺ incorporation and its effect on photoluminescence. The argument had already narrowed to a fairly specific question: how Mn²⁺ incorporation might affect halide-vacancy-related defect states, and whether those local changes could help explain the optical behaviour observed in the system. ChatGPT continued in a broader direction, discussing cavity engineering, defect passivation and compositional optimisation. There was nothing obviously bad about the continuation. All of those are legitimate ideas in the broader perovskite literature. Read on its own, the paragraph sounded technically competent and would not have looked particularly strange in a general Discussion section. The problem was that it answered a different question. The manuscript was no longer asking, broadly, what approaches can improve optical performance in perovskites? It was already trying to understand what Mn²⁺ was doing in this particular system. Moving back toward general optimisation strategies widened the discussion precisely when the paper needed to stay narrow. This is the kind of failure I find more interesting than a grammatical mistake. The model has not misunderstood the vocabulary of the field. It has misunderstood why this paragraph exists here, and where the argument is trying to go next. A sentence can be fluent, technically plausible and individually defensible, yet still pull the manuscript in the wrong direction. #### When plausible language becomes abundant Academic writing is often described as if the difficult part were turning ideas into sentences. Sometimes it is. But once a researcher is deep inside a paper, the harder decisions usually happen before the sentence appears. A paragraph about an experimental result can move in several reasonable directions. The author might explain a mechanism, compare the result with earlier work, introduce a limitation, shift from efficiency to stability, or decide that the paragraph is already finished. A language model can write all of those continuations. Its ability to generate them does not tell us which one belongs there. Generative AI has dramatically increased the supply of plausible language. A researcher no longer has to struggle to produce three possible continuations; a model can produce ten before they have finished thinking through the first one. More choice is useful only up to a point. After that, the work becomes selection. The researcher still has to judge whether the proposed direction fits the role of the paragraph, whether the evidence is strong enough, whether a mechanism is being overstated, and whether an apparently relevant paper really supports what the new sentence claims. The irony is that better prose can make this harder. If an AI produces an awkward sentence, we inspect it. If it produces a polished sentence using the right vocabulary and a familiar scientific explanation, our attention can shift very quickly from “Is this justified?” to “Does this read well?” The Mn²⁺ example illustrates the problem. A statement about defect passivation can be scientifically sensible while still failing in several ways: the experiment may have been performed under different conditions, the source may support the general mechanism but not the strength of the claim, or the sentence may simply take the manuscript somewhere the author does not want it to go. These are not really problems of language. #### A thousand PDFs do not give you a thousand papers’ worth of working memory There is another reason generation is becoming the wrong place to focus all our attention. Researchers already possess far more potentially useful information than they can actively remember. A literature library grows in a strange way. At first, almost every paper is familiar. After a few years there may be hundreds or thousands of PDFs. Many were read carefully at some point. Some contain exactly the experiment, caveat or comparison needed for the paragraph being written today. But the researcher no longer remembers enough to retrieve them cleanly. The memory is often incomplete: there was a paper that showed something similar under continuous excitation. Or: I remember a figure where the dopant seemed to occupy a particular site. Or simply: I’ve seen this mechanism somewhere before. At this stage the normal search box becomes surprisingly demanding. It asks the researcher to convert a partial memory into the right terminology before the system can help. That is frequently not how memory works. Recognition often arrives before language. You can fail to remember a paper’s title, author and keywords, then recognise the relevant paragraph almost instantly when you see it again. This is why a small passage from a paper can sometimes be more useful than a perfectly ranked list of ten titles. A snippet is not only a search result. It can be a cue. Sometimes the reaction is straightforward: yes, that’s the paper I meant. But the more interesting case is when the resurfaced passage changes the direction of thought. Perhaps you were looking for evidence about efficiency and a passage reminds you that the same study reported a stability effect you had forgotten. Perhaps a mechanism described in another material system suggests a comparison you had not planned to make. Perhaps the paper turns out to be less supportive than you remembered, and you decide not to write the sentence at all. None of these outcomes looks particularly impressive in a conventional AI demo. No paragraph may have been generated. But something useful has happened: knowledge that had fallen out of active memory has returned to the work. Large literature collections are usually treated as storage and retrieval problems. We organise papers into folders, tag them, index full text and improve search. Those things matter. But once the collection becomes large enough, another problem appears. The question is no longer only whether the information exists somewhere in the library. It is whether the right fragment can come back into the researcher’s attention at the moment when it becomes useful. Search begins with a need that has already been articulated. Recall does not always have to. It can begin with the manuscript. #### Context is useful only when it becomes selective A common response to AI limitations is to add more context. Give the model the full paper. Give it twenty papers. Increase the context window. Add notes, previous conversations and eventually the whole literature library. There are obvious benefits to this, but a research library exposes the limit of the idea. A thousand PDFs already contain an enormous amount of context. Almost all of it is irrelevant to the paragraph currently being written. The interesting problem is not how much information the system can hold. It is how well it can decide what deserves to become active. Imagine a researcher writing three sentences about a particular experimental observation. Somewhere in the library are two passages from papers read eighteen months ago that bear directly on the mechanism, another passage that challenges the obvious interpretation, and hundreds of papers sharing the same broad keywords while contributing nothing useful to this particular moment. Putting all of them into a larger context window does not solve the problem. Finding the few fragments worth looking at might. This is where the distinction between generation and decision support starts to matter. A useful system could surface those fragments while the researcher is still deciding what the paragraph means. They do not have to dictate the next sentence. One of them may support it. Another may weaken it. Another may persuade the researcher to abandon the sentence altogether. Academic tools have traditionally treated the manuscript and the literature library as two different places. You write in one and search the other when necessary. AI gives us an opportunity to connect them more closely, but simply connecting them is not enough. What matters is what crosses that boundary, and when. A paper title is sometimes enough. Often it is not. A paragraph, a figure caption, or a few lines around a result may contain the thing that actually changes what the researcher thinks next. #### Once generation is cheap, the workflow itself becomes the interesting problem This is the part I find increasingly important. If AI can already produce the sentence, perhaps the next question is not how to make it generate even more. It is how to reorganise what happens before generation. Today, a fairly normal research-writing workflow still looks something like this: write → remember something vaguely → leave the manuscript → search the library → try several queries → open PDFs → locate the passage → return to the manuscript → ask the AI → continue writing. Each individual step is manageable. The friction comes from having to reconstruct the connection between the manuscript and the literature every time. A different workflow becomes possible if the manuscript itself is treated as a signal. The paragraph already tells us quite a lot: what the researcher is discussing now, what has just been established, which concepts are active, and often what kind of evidence might matter next. That means the researcher should not always have to stop writing and formulate a search query before the literature can become useful. A small number of relevant passages could resurface from the researcher’s own library while the argument is still taking shape. Not fifty recommended papers. Not the entire library compressed into a context window. Just a few fragments with a real chance of mattering here. One might recover a paper the researcher was already trying to remember. Another might reveal that the remembered evidence was weaker than expected. A third might introduce a comparison the researcher had not been looking for at all. At that point the workflow begins to look less like prompt → generation, and more like manuscript → relevant fragments → recognition → evidence → judgment → writing. The AI still has an important role at the end of that chain. It can help articulate what the researcher has decided, propose a continuation, or improve the expression. But generation no longer has to be the first interesting thing the system does. Sometimes the most useful intervention happens before a sentence exists. #### What becomes valuable then? I would not measure the next generation of academic AI mainly by how much text it can produce. I would pay more attention to whether it can bring forward material that matters to the local argument, rather than merely the broad topic of the manuscript. Whether a suggestion can be traced back to the original evidence without ten minutes of detective work. Whether literature that has fallen out of active memory can become useful again. And whether the system leaves the final judgment where it belongs: with the researcher deciding whether two experiments are comparable, whether a mechanism is justified, whether the evidence is strong enough, and whether the manuscript should move in that direction at all. These capabilities do not always result in more text. Sometimes the result is a deleted sentence. Sometimes it is ten minutes spent rereading an old paper. Sometimes a resurfaced passage changes the framing of an entire paragraph. Measured in words per minute, none of that looks especially productive. But research has never worked particularly well under that metric. AI is already becoming very good at producing the sentence. What interests me more now is whether it can make everything immediately around that sentence easier: remembering what we have read, seeing the evidence that matters, and deciding what is actually worth writing next. ### Research Writing in the Age of AI: Beware the Bubble of Beautiful but Empty Prose Source: https://flowing.works/en/blog/evidence-before-fluency AI has become very good at writing sentences that look right. The harder problem is that some of these sentences are simply too convincing. AI has become very good at writing sentences that look right. Grammar is rarely the problem anymore. The terminology often sounds appropriate. Give a model a paragraph from a manuscript and ask it to continue, and it usually knows when to introduce a mechanism, when to qualify a result, and when to connect one observation to another. The harder problem is that some of these sentences are simply too convincing. They read well enough that we stop asking a more basic question: how much of this sentence is actually supported by the evidence? Suppose you are writing about Mn²⁺ incorporation and defect passivation in a perovskite system. The experimental results are already on the page. What you need next is an explanation for the improvement in photoluminescence. An AI model might naturally produce something like: “Mn²⁺ incorporation improves photoluminescence by suppressing defect-mediated non-radiative recombination.” It is clean. The terminology is plausible. The causal chain is easy to follow. It would not look out of place in a Discussion section. Then you go back to the paper that is supposed to support it. What the paper actually shows may be narrower: Mn²⁺ incorporation is accompanied by a reduction in vacancy-related defect states, while photoluminescence also improves. The authors may suggest that these changes are related, but the experiment does not necessarily establish the full causal chain implied by the generated sentence. The evidence may support: the improvement may be associated with reduced defect states. It may not support: the improvement is caused by defect suppression. Only a few words have changed. In research writing, those few words can carry most of the scientific responsibility. #### What AI is best at can also hide what is still unresolved Before AI-assisted writing became routine, many problems announced themselves through bad prose. If a sentence was difficult to write, you stopped. If the mechanism was unclear, the paragraph often sounded unclear too. If two results did not fit together, the awkwardness sometimes forced you to look at the data or literature again. Now a model can remove that awkwardness in seconds. That is useful. But it can also remove some of the signals that used to tell us that the thinking was unfinished. A vague interpretation can quickly become a polished paragraph. The transitions work. The terminology is correct. The logic feels complete. There may even be a citation attached. Yet linguistic completeness is not the same as scientific completeness. Sometimes we have simply made an unresolved problem sound resolved. This is a different kind of empty prose from the old cliché of academic writing full of decorative language and little substance. AI-generated emptiness can be much harder to spot. It may contain field-specific terminology, a coherent argument, and a real reference. What it lacks is not the appearance of substance, but enough evidence and judgment underneath the language. #### A real citation does not mean the sentence is supported More academic AI tools now attach citations to generated claims, and that is clearly better than unsupported generation. But a citation is a very coarse unit. A paper can be relevant to a sentence without supporting that sentence as written. The part that actually matters may be one experiment in the Results section, a figure, a sentence in the Discussion, or even the conditions described in a caption. Within the same paper, several things may coexist: what the experiment directly observed, what the authors inferred from those observations, what they proposed as one possible explanation, and what remained unresolved. When all of that is compressed into a smooth sentence, those boundaries are easy to lose. The citation can be real. The paper can be relevant. The topic can match perfectly. And the sentence can still say more than the source allows. So “does this sentence have a citation?” is only the beginning. A more useful test is: if the source passage were placed in front of me right now, would I still be comfortable writing the sentence this strongly? #### Some sentences should not exist yet The default AI-writing workflow usually begins with generation. Generate the sentence. Read it. If it seems plausible, verify it or add a citation afterwards. For many tasks, that order works well. Grammar correction, shortening, tone adjustment, title variants, or turning an already settled idea into a cleaner abstract are mostly problems of expression. The researcher already knows, roughly, what they want to say. AI helps say it better. Scientific claims are different. Here the unresolved question may be whether the statement should be made at all. Is this correlation or causation? Does this experiment establish the mechanism? Are the conditions in the two papers actually comparable? Should the verb be “demonstrates,” or only “suggests”? Does the conclusion apply to this system, or only to the one in the source paper? When the complete sentence appears first, there is a subtle change in the direction of reasoning. Instead of asking “What does the evidence allow me to conclude?”, we begin asking “I already have a sentence. Which paper can support it?” That reversal is easy to miss, especially when the sentence is very good. #### Put the evidence in front first, and the sentence often changes Return to the Mn²⁺ example. Imagine that before writing the explanation, you first see two short passages from papers already in your library. One reports Mn²⁺ incorporation together with a reduction in vacancy-related defect states. Another is much more cautious, saying that this change may contribute to reduced non-radiative loss. Now the sentence is likely to come out differently: “The improved photoluminescence may be associated with the reduction of vacancy-related defect states following Mn²⁺ incorporation.” It is less decisive. Perhaps less elegant. But it is closer to what the evidence actually permits. Sometimes reading the source leads to a stronger outcome: the mechanism should not be written at all. The two observations may occur together without the experiment being able to show that one caused the other. In that case, the best result is not a more carefully hedged AI sentence. It is deleting the sentence. Research writing contains many moments like this. Sometimes the best continuation is no continuation. #### Literature can stop a sentence, not just support one There is another possibility. You begin by looking for evidence explaining why photoluminescence improved. While reading the source, you notice that the authors also discuss stability under continuous excitation. Now the question in your own paragraph starts to shift. Perhaps the more interesting issue is not only why emission improved, but whether the same intervention affected both efficiency and stability. The literature has not helped you finish the sentence you intended to write. It has interrupted it. That matters because literature retrieval is often designed as if its purpose were to justify claims that already exist. But research does not always move in that order. A figure can suggest a comparison you had not considered. An old mechanism can make your current explanation look too narrow. A forgotten experimental condition can weaken what looked like a clean conclusion. Sometimes the most useful piece of evidence is the one that makes you stop. #### The pause after seeing the evidence is worth keeping There is plenty of friction in research writing that should disappear. Searching folders. Opening the same PDFs repeatedly. Trying five versions of a keyword. Finding a paper you know you have read. Scrolling for the page where a result was reported. AI can reduce much of that, and it should. But once the relevant evidence is actually on screen, there is another kind of friction that I would rather keep. The few seconds spent asking: Did this paper really demonstrate that? Are the experimental conditions comparable? Is the author reporting an observation or interpreting one? Am I saying more than the evidence says? Those pauses are not inefficiency. A surprising amount of scientific rigor lives inside them. The useful role for AI is to reduce the mechanical work required to reach the evidence, not to eliminate the judgment that begins once the evidence is there. #### Some academic tools are beginning to move in this direction A few newer tools are already experimenting with ways to bring sources closer to the writing process. Jenni allows researchers to bring PDFs and reference libraries into AI-assisted writing and citation workflows, so generated text can be grounded in sources rather than produced in isolation. Flowing takes a somewhat different approach. It connects the current manuscript with a researcher’s own literature library, resurfacing relevant source passages as the researcher writes and allowing those passages to inform later polishing or continuation. The implementations differ, but the underlying direction is interesting. The next useful step for academic AI may not be to produce an even more fluent sentence. It may be to make the relevant evidence easier to see before the sentence becomes settled. Then the researcher can decide whether the claim should be strong, cautious, reframed, or removed entirely. #### The harder part is still the claim AI can help with the wording afterwards. The harder part is still deciding how much the evidence actually allows us to say. ### Research Writing: How Our Memory Gets Buried Under Thousands of PDFs Source: https://flowing.works/en/blog/the-problem-with-a-thousand-pdfs A research library is only useful if the right part of it can come back into view when you need it. A research library is only useful if the right part of it can come back into view when you need it. After a few years of research, something strange starts to happen to a personal literature library. The number of papers keeps growing. The number you can readily recall does not. When you first enter a field, the library may contain only a few dozen papers. You roughly remember what each one is about, where a particular figure came from, which experiment used which method. Finding something is rarely difficult because much of the literature is still sitting somewhere in working memory. A few years later, there may be hundreds or thousands of PDFs. Some were read carefully. Some were saved for a single figure. Some were annotated or discussed in a group meeting. Others stayed in the library because, at the time, you had the vague feeling that this might be useful someday. The papers are still there. What fades is your connection to them. When writing, the problem is often not “I’ve never seen anything about this.” It is more like “I’m sure I’ve seen something like this somewhere.” Or: “There was a paper that did something similar.” Sometimes all that remains is: “I remember a figure that might be relevant to this mechanism.” You do not remember the title, the author, the year, or even a reliable keyword. You remember a shape. #### Search works best once memory has become language Suppose you are writing about defect passivation in a perovskite system. You vaguely remember a paper in which doping reduced defect states and affected non-radiative recombination. You try: dopant defect passivation perovskite. Too broad. Then: Mn vacancy passivation perovskite. Still not right. You open one paper. Wrong material system. Another discusses Mn²⁺, but mainly in the context of luminescence tuning rather than the mechanism you remember. You try “defect suppression.” Then “halide vacancy.” Perhaps an author name you think you remember. This kind of back-and-forth is familiar to anyone with a large literature library. What consumes attention is not always the quality of the search engine. Often it is the work of translating a half-formed memory into a query the search system can understand. Human memory does not always arrive as language. Recognition comes before recall can be put into words. You may spend ten minutes unable to remember the title of a paper, then see three or four lines from the middle of it and know immediately: yes. That’s the one. A short passage can sometimes do what a list of titles cannot. It brings back not only the paper, but the reason you cared about it: the figure before it, the experimental condition, perhaps even the judgment you made when you first read it. #### What gets lost is not only the paper A literature library also accumulates things that never became proper notes. A method you wanted to revisit. A figure that suggested a control experiment. A paragraph in the Discussion that had little to do with your project at the time, but seemed worth remembering. A result that bothered you, although you never followed the thought. Months later, the paper itself may have disappeared from active memory while something faint remains: I think I’ve seen a similar explanation before. This is why a large library can feel both rich and strangely inaccessible. The files are intact. The thoughts attached to them are not. It also explains why someone can have a thousand papers in a library and still return, while writing, to the same familiar few dozen. The rest are not necessarily less useful. They are simply harder to bring back at the moment they matter. #### Sometimes you don’t know that you should be searching There is a harder case than forgetting the right keyword. Sometimes you are not looking for anything at all. You are writing. Perhaps the Results section is moving into Discussion. You are explaining an improvement in photoluminescence and naturally following the argument toward defect passivation. Then you come across an old passage mentioning that the same intervention also affected stability under continuous excitation. The paragraph may now need to go somewhere else. What began as “Why did the emission improve?” may turn into “Did this intervention affect both efficiency and stability? Is that relationship more interesting than the original framing?” The paper has not answered the question you were asking. It has changed the question. This happens often enough in real research that I am wary of treating literature retrieval only as a way to support claims that have already been formed. A forgotten condition can weaken a conclusion. A figure can suggest a comparison. A mechanism from another system can make the current explanation look too narrow. Sometimes the literature is useful precisely because it interrupts the direction you were already taking. #### A different way for the library to enter the writing process This is also why I am not convinced that the answer to a large literature library is simply more recommendations. Researchers rarely suffer from a shortage of papers. Sometimes what matters more is three forgotten lines from a paper they already read six months ago. Imagine keeping the manuscript in front of you while a small number of passages from your own library surface alongside it. Not fifty recommended papers, and not the entire library compressed into a context window. Just a few snippets with a plausible connection to the paragraph you are currently writing. One gets ignored. Another confirms that the experimental conditions really are comparable. A third looks familiar enough to make you stop and open the original PDF. The manuscript already provides a surprisingly rich signal: what you are discussing, what has just been established, whether the paragraph is explaining a mechanism or making a comparison, and what kind of evidence may matter next. In some situations, that is more informative than a keyword you invent after leaving the writing flow. The usual sequence is something like: remember something vaguely → stop writing → invent keywords → search → open PDFs → find the passage → return to the manuscript. I would rather have the option of: keep writing → a few relevant fragments resurface → recognise / remember → judge → inspect the source if needed → decide what to write next. Search still matters. If you know the author, title, method or material system, a search box is exactly what you want. This other mode is useful when the memory is incomplete — or when you have not yet realised that an old paper may matter at all. There are already tools moving toward parts of this workflow. Jenni lets researchers bring PDFs and sources from tools such as Zotero or Mendeley into the writing process for citation, questions and AI-assisted drafting. Flowing takes a more passage-oriented approach, resurfacing snippets from a researcher’s own library according to the manuscript context while they write. What interests me is less which implementation wins than the shift in the workflow itself. #### The library should not wait for the right keyword A personal literature library should not have to wait until I remember the right keyword before it becomes useful. Sometimes the most valuable thing it can do is return a fragment I had forgotten — at exactly the point where it changes what I was about to write next. ### The Missing Piece in AI Academic Writing Isn't the Model. It's Your Local Research Library. Source: https://flowing.works/en/blog/missing-piece-local-research-library Thinking about a brand new approach to evidence-grounded AI writing. Thinking about a brand new approach to evidence-grounded AI writing. GPT-5. Claude. Gemini. DeepSeek. Every few months there’s another benchmark. Another leaderboard. Another “best writing AI.” But after spending the last two years building tools for researchers, I became convinced that this wasn’t the real question. The real question is: What should AI write from? #### Researchers rarely start from an empty page When researchers write papers, they rarely start from an empty page. They constantly move between ideas they’ve already collected: a paragraph from a review paper, an experimental setup they bookmarked months ago, a sentence they highlighted last semester, an image buried somewhere in a PDF. Those fragments are where writing actually happens. But most AI writing tools aren’t built around that moment. Their default interaction starts from a blank prompt—so the researcher’s own library never enters the loop unless someone pastes it in by hand. #### Today’s workflow leaves the library outside the loop Today’s workflow usually looks like this: Prompt → LLM → Generated paragraph. The model knows everything. Except the papers that actually matter. Researchers end up copying paragraphs from PDFs into ChatGPT over and over again. Not because they enjoy it. Because the interaction gives the AI no lasting access to their paper library. #### The overlooked problem isn’t writing. It’s retrieval. Or more precisely—continuous retrieval while writing. When writing a manuscript, you rarely know exactly which paper you need. Often you only remember: “I saw something similar a few months ago...” or “There was a really good explanation somewhere...” The problem isn’t that the paper doesn’t exist. It’s buried somewhere inside hundreds or even thousands of PDFs—and hard to find again when you need it. #### We built Flowing around a different idea Before building Flowing, I assumed literature retrieval was already a solved problem. Zotero/Mendeley manages papers well. Connected Papers helps discover papers. Semantic Scholar helps search papers. Then I realized something surprising: almost none of them actually help while you’re writing—at least not in a natural way. So, instead of asking researchers to search their literature... what if the literature searched for them? As you write, Flowing continuously analyzes the current manuscript and retrieves highly relevant passages from your own library. Not entire papers. Not another search results page. Passages from your library—shown as Snippet Cards beside the editor. Those snippets stay available while you write. You can preview the original paragraph via its page thumbnail, or open the source PDF at that passage in one click. It sounds like a small interaction. In practice, it fundamentally changes how AI writes. #### Better evidence leads to better AI This also changes AI writing itself. Most AI assistants generate text primarily from their internal knowledge. Flowing instead lets AI draft from those snippets—passages retrieved from your paper library, not from general knowledge alone. When you continue a manuscript, the Snippet Cards don’t just sit beside the editor. They become the basis for the next sentences, so the continuation stays closer to your sources, terminology, and line of argument. One of our earliest users—a chemistry professor—recently compared Flowing and ChatGPT while continuing a manuscript he was actively writing. Both systems used strong language models. The difference wasn’t the model. It was the context. Because Flowing retrieved relevant passages from his paper library and used them as the basis for continuation, the generated text stayed much closer to his research direction, terminology, and upcoming discussion. If you’re interested, you can read the complete comparison here → AI Continuation Comparison on a Real Chemistry Manuscript. #### Where this leads Large language models will keep improving. That’s inevitable. For academic writing, the more important question is quieter: How do we make the right passages available at the moment of writing—for both you and the AI? That’s where the next generation of academic writing may begin—and it’s the problem Flowing is built around: keeping your research library inside the writing loop—so you can recall the right passages as you write, while AI writes from that evidence. If you’ve felt the same gap while drafting, we’d love for you to try Flowing. ### Beyond ChatGPT: What the Next Generation of AI-Assisted Research Writing Could Look Like Source: https://flowing.works/en/blog/ai-paper-continuation-comparison An in-depth comparison of how Flowing and ChatGPT continue the same research manuscript across four writing scenarios. This is a guest post by Hao Chou, a researcher working at the intersection of materials chemistry and optoelectronics. Lately, I’ve been revising a manuscript on rare-earth-doped perovskite micro/nano lasers. A friend of mine has been building a desktop academic writing tool called Flowing and asked if I’d be willing to test its continuation feature. Since I already keep ChatGPT open whenever I’m writing papers, I decided to compare them side by side. The setup was simple: I placed both tools at the same point in the manuscript, gave them the same instruction, and evaluated only their first response. This was not a controlled model benchmark. It was a comparison of the two tools as I would realistically use them in my everyday writing workflow. After trying this across four different writing scenarios, I gradually realized that the interesting part wasn’t which continuation sounded more polished. It was which one was more likely to generate the sentence that actually belonged in my manuscript. Here’s what I found. #### Cursor 1 — A literature review needs continuity, not a premature summary The first comparison came from a literature review paragraph. The paragraph had already begun discussing representative approaches for reducing lasing thresholds, and several more studies were about to follow. At this cursor, I wasn’t looking for another overview of the field. I simply needed a sentence that could connect naturally to the remaining literature. Flowing stayed within the local narrative of the paragraph and smoothly transitioned toward the following studies. ChatGPT generated a broader summary covering cavity engineering, defect passivation, compositional optimization, and several other directions. Taken by itself, I actually liked the sentence. The problem wasn’t correctness. It simply arrived too early: the remaining studies hadn’t been introduced yet, so summarizing the field at this point interrupted the rhythm of the literature review. One thing that stood out to me was Flowing’s Evidence Card. Nearly every retrieved paper revolved around the same local topic, which may have helped the continuation stay aligned with the paragraph rather than drift toward a broader overview. In the end, I kept Flowing’s continuation almost unchanged, while moving ChatGPT’s sentence to the end of the section, where it worked well as an overall summary. #### Cursor 2 — A mechanism section should extend the reasoning, not restate the conclusion The second comparison came from the mechanism discussion. The previous paragraph had already introduced the energy-level diagram. The next sentence wasn’t supposed to repeat that Ce³⁺ reduces the lasing threshold—it was supposed to explain why. Flowing continued from intermediate energy states to carrier relaxation and eventually population inversion, extending the mechanistic reasoning step by step. ChatGPT focused more on reinforcing conclusions that had already appeared, such as reduced non-radiative loss, improved carrier accumulation, and lower thresholds. Both continuations were perfectly reasonable. The difference was simply their emphasis: one developed the mechanism further, the other reinforced what had already been established. #### Cursor 3 — After characterization: bridge to function, not a field overview The third comparison came after I had summarized SEM, EDS, and XPS results showing regular geometry, smooth surfaces, and high crystal quality. The paragraph was about to explain why that mattered for laser microcavities. Flowing connected those morphological attributes to scattering losses, optical confinement, and Fabry–Pérot microcavities—almost writing the bridge into the next paragraph for me. ChatGPT produced an excellent overview of MHP micro/nano laser advantages, applications, and citations. If the goal had simply been to introduce the field, I might even have preferred it. But at that cursor, I needed the paragraph to continue from the characterization results in front of me, not expand the background. Looking back, I suspect this was also influenced by how the two systems used context. Flowing combines the current paragraph with evidence retrieved from related papers. Here, the retrieved passages stayed closer to crystal quality and optical confinement rather than pulling the discussion toward a generic field introduction. #### Cursor 4 — Sometimes, the manuscript itself is the context that matters most The last comparison changed my perspective the most. This paragraph had already confirmed laser oscillation. ChatGPT continued by discussing threshold behavior, linewidth narrowing, cavity Q factor, optical feedback, and light confinement. Everything was consistent with the local paragraph. Flowing went one step further. It explicitly connected the discussion back to the Fabry–Pérot cavity that had already been established earlier in my manuscript, and linked it with the threshold analysis and cavity Q factor discussed in the following figures. My first reaction was that this seemed slightly bold—the current paragraph never mentioned the Fabry–Pérot cavity. Then I remembered that I had already established it several pages earlier while discussing Figure 3c. Flowing wasn’t inventing a new conclusion. It was reusing information that already existed in my manuscript because it dynamically incorporated context from the surrounding manuscript when generating the continuation. In my usual browser-based ChatGPT workflow, I would not realistically copy Figure 3c, its caption, and the surrounding discussion into every prompt. That was the moment I realized something: for scientific continuation, the most valuable context isn’t always another fifty reference papers. Very often, it’s simply the manuscript you’ve already written. #### Final thoughts I still use ChatGPT every day—for brainstorming, exploring unfamiliar concepts, and improving wording. It remains one of the strongest general-purpose AI tools I use. But once a manuscript is already taking shape and I need to continue from a specific sentence, Flowing is now the tool I tend to try first. From my perspective, that is what makes Flowing different. It is designed to keep AI grounded in both your local literature library and the manuscript you are actively writing. In these four examples, that context often made the difference between a continuation that merely sounded good and one that actually fit. ## Documentation ### Flowing documentation Source: https://flowing.works/en/docs Learn the full writing-with-context workflow—from your first PDF import through References, polish, and continuation. Flowing is a desktop app for **writing with context**—not empty chat. Your reference PDFs stay on your machine and power recall, Q&A, polish, and continuation as you draft. ### Three core capabilities, one writing workflow - **Snippet Card · Recall the source:** Recall surfaces PDF passages relevant to your draft as cards with a direct link to the original page. - **Evidence Card · Check the claim:** Red marks contradiction, green direct support, yellow a useful qualifier. Checks cover retrieved passages only; they do not rewrite your draft or guarantee completeness. - **Autocomplete · Keep writing:** Click the sparkle or press Tab to preview the next sentence in context. Inspect sources when used, then press Tab again to insert. ### Step-by-step guides Follow these chapters in order, or jump to the step you need. Each guide includes screenshots from the app. - 01. **Prepare** — Create a library and import reference PDFs. (https://flowing.works/en/docs/prepare) - 02. **Manual** — Keyword search across your library—free on every plan. (https://flowing.works/en/docs/manual) - 03. **Write** — Rich text, LaTeX, and PDF preview—synced in real time. (https://flowing.works/en/docs/write) - 04. **Autocomplete** — Preview the next sentence at the cursor, review it, then press Tab to insert. (https://flowing.works/en/docs/autocomplete) - 05. **Recall** — Recall library passages while drafting; optional web paper recommendations are off by default. (https://flowing.works/en/docs/recall) - 06. **Verify** — Check draft paragraphs against recalled library evidence—red, green, and yellow cards. (https://flowing.works/en/docs/evidence-verification) - 07. **Ask with context** — Attach passages and chat with AI. (https://flowing.works/en/docs/ask-with-context) - 08. **Polish** — Grammar and content polish with library context. (https://flowing.works/en/docs/polish) - 09. **Continue writing** — Generate new text after the cursor—library-backed. (https://flowing.works/en/docs/continue-writing) - 10. **References** — Browse bibliography cards, jump citations, and insert from your library with @. (https://flowing.works/en/docs/references) > **Note:** Manual keyword search across your library does **not** consume monthly tokens on any plan—you can search freely while writing. ### Prepare your library Source: https://flowing.works/en/docs/prepare Create a library, import reference PDFs, and grow it with open-access web recommendations when you need more sources. Importing your library is the foundation for *Recall*, *Ask with context*, and *evidence-grounded revision*. Even a focused set of core papers is enough to experience the full loop. ### Create a library and import PDFs 1. **Open Libraries** — From the top navigation, open Libraries—this is where your reference PDFs live. 2. **Click New Library** — Give your library a name (for example, default or a project name). 3. **Select PDF files or a folder** — Flowing imports every PDF it finds and begins parsing text for search and AI features. 4. **Wait for parsing** — Large files may take several minutes. Progress shows which PDF is being chunked (for example, 2/7). _[Figure: New Library dialog with library name and PDF import. Create a library and point Flowing at a folder of PDF papers.]_ _[Figure: Import progress bar while chunking PDFs. Parsing progress while text is extracted from each PDF.]_ > **Tip:** Click *Run in background* to close the dialog and keep working—the library continues importing while you write or explore. ### Import and background parsing _[Figure: Run in background button during PDF import. Use Run in background while large libraries finish parsing.]_ While importing, Flowing splits each PDF into searchable passages and builds metadata (title, citations, dates). Large libraries may take several minutes—that is normal. _[Figure: Library list showing import complete. When parsing finishes, the library is ready for search.]_ ### Auto-recommend open-access papers from the web When your library is still light on PDFs, turn on *Auto-recommend papers from the web* at the top of the Recommended panel. Flowing reads the paragraph you are working on and searches OpenAlex for **open-access papers tightly tied to what you are writing**—not a generic keyword dump. Each suggestion runs through a propose-and-verify pipeline: an LLM drafts a small set of candidate papers with a concrete *Why relevant* rationale and citation role, then OpenAlex confirms the work exists and has an OA PDF. Rationales that only repeat topic keywords are rejected—what you see should explain how a paper supports a specific mechanism, measurement, or claim in your draft. _[Figure: Recommended panel with web paper cards and Why relevant explanations. Web recommendations show title, metadata, matched tags, and a Why relevant line linked to your paragraph.]_ 1. **Turn on the toggle** — Enable Auto-recommend papers from the web (off by default). The first time, confirm the privacy notice. Optionally include paragraph body text for stronger matches—also off by default. 2. **Keep writing** — As keywords stabilize in a section, Flowing recommends 0–3 verified OA papers per qualifying paragraph, skipping titles already in your library. 3. **Read Why relevant** — Use the rationale under each card to judge whether the paper truly backs your line of argument—not just shares vocabulary. 4. **Add to library** — Import papers you want to keep. They join the same local library used for snippet recall and context attach. > **Note:** Web recommendations use tokens on paid tiers and send keywords—and optional paragraph text only if you opt in—to Flowing’s recommendation service. OpenAlex metadata is used for verification; links are never invented. ### Your data stays local Flowing is a **desktop app**: your PDFs, drafts, and search index stay on **your machine** by default. Only the snippets needed for AI features are sent to model services when you use them. > **Note:** Library size can range from **dozens to thousands** of PDFs. Start with papers most relevant to your current project. ### Manual search Source: https://flowing.works/en/docs/manual Search your library by keyword on demand—exact and phrase matching, free on every plan. The *Manual* tab lets you search your library by keyword whenever you want—**without using monthly tokens**. Use it to pull passages on demand while writing or reviewing sources. ### Open the Manual tab _[Figure: Manual tab in the right panel on the writing page. Switch to Manual on the writing page to search your library by keyword.]_ 1. **Go to the writing page** — Open a document from Documents in the top navigation. 2. **Select Manual** — In the right panel, click the Manual tab (next to Recommended and Polishing). 3. **Enter keywords** — Type one or more keywords or a short phrase. Results appear as snippet cards from your library. ### Search tips _[Figure: Manual keyword search results as snippet cards. Snippet cards show matching text and a page thumbnail for each hit.]_ Manual search uses **keyword and phrase matching** across your library—not semantic similarity. Multi-word queries match phrases in order; punctuation and spacing are normalized so common variants still hit. > **Tip:** Try distinctive terms from your field (method names, materials, abbreviations). Shorter, specific phrases often work better than long sentences. ### Preview and jump to source _[Figure: Hovering a manual search result thumbnail for full-page preview. Hover a thumbnail to preview the source page; click to jump back to the PDF.]_ Each result card includes excerpt text and a page thumbnail. Hover the thumbnail for a full-page preview with the match highlighted—same as in *Recall*, but driven by your own search query. > **Note:** Manual search is unlimited on Free and above. Automatic snippet recommendations in the Recommended tab use tokens on paid tiers. ### Write Source: https://flowing.works/en/docs/write Three writing views—rich text, LaTeX, and PDF preview. Rich text and LaTeX stay in sync in real time. The writing pane has three modes: **rich text**, **LaTeX**, and **PDF preview**. Rich text and LaTeX edit the same document and *stay in sync in real time*—edits in one view update the other immediately. ### Rich text mode Rich text mode with in-draft keyword highlights and the Recommended panel. _[Figure: Rich text editor with scientific content and keyword highlights. Rich text mode for WYSIWYG drafting.]_ Use the toolbar for bold, italic, links, equations (Σ), tables, and structural blocks. _[Figure: Editor formatting toolbar. Formatting toolbar: styles, equations, tables, and preview.]_ ### LaTeX mode Edit the underlying .tex source when you need precise control over structure, macros, and layout. This is the same document as rich text—**both views stay in sync in real time** as you type. _[Figure: LaTeX source editor alongside the draft. LaTeX mode for direct .tex editing—synced with rich text.]_ ### PDF preview See the compiled PDF alongside your draft. Built-in Tectonic compiles your document—no TeX Live install required. The preview matches the exported PDF and updates shortly after you pause typing. _[Figure: Live PDF preview of the compiled draft. PDF preview shows the compiled output as you write.]_ ### Import and export On the documents list, import an existing manuscript from **Word (.docx)** or **PDF**, then keep writing in rich text or LaTeX. Export the draft as Word or PDF when you need a file to share or submit. ### Evidence-backed autocomplete Source: https://flowing.works/en/docs/autocomplete Pause at the cursor, click the sparkle or press Tab, and get a gray inline preview in seconds; guide it with a passage or figure from your selected PDF library when needed. Pause while writing, then click the sparkle beside the cursor or press Tab. Flowing generates the next sentence without requiring a prompt or opening a chat. When direction matters, add a note, a passage from your selected PDF library, or a figure from the manuscript. ### Generate without leaving the draft After a short pause, the sparkle appears beside the cursor. Click it or press Tab to generate. Flowing reads the nearby manuscript context, can match relevant passages from the selected library, and shows a gray inline preview at the insertion point. _[Figure: Generate without leaving the draft. The interaction stays beside the cursor: no chat window, copy-paste loop, or manual search is required.]_ ### Guide the continuation when needed Hover the sparkle to add an optional note, selected passage, or image. Leave everything empty for a direct completion, or add material when you need to steer terminology, mechanism, or direction. _[Figure: Guide the continuation when needed. Optional context gives you control without turning every completion into a prompt-writing task.]_ ### Review, inspect evidence, then insert The suggestion appears as gray text after the cursor and is not inserted immediately. Review whether it fits the argument, inspect matched source passages when library material is used, then press Tab again to insert. You remain responsible for checking the scientific claim and citation fit. _[Figure: Review, inspect evidence, then insert. Fast does not mean opaque: manuscript context determines what belongs next, while source passages constrain what the sentence can responsibly claim.]_ > **Note:** Evidence-backed means the suggestion can expose supporting library passages for inspection. It does not mean every generated claim has been independently verified or is guaranteed correct. ### Recall Source: https://flowing.works/en/docs/recall Surface relevant library snippets as you write. Click a thumbnail for a full-page preview; click the card corner or highlighted keywords to open the source PDF. Recall is *finding library passages that match keywords in your draft*. Flowing spots keywords and domain terms as you write, pulls matching passages from your library, and shows them as cards. ### Recommended tab As you write, the *Recommended* tab fills with terms found in your draft (for example, PLQY, XRD, material names). Each row shows how many snippets and documents matched. Click a term to expand snippet cards. _[Figure: Snippet result cards with text excerpts. Snippet cards show excerpt text plus a page thumbnail.]_ ### Thumbnail preview Click the thumbnail on a snippet card to open a **full-page preview** of that source page—the matched passage is highlighted in context, with figures and layout intact. _[Figure: Clicking a snippet thumbnail to open full-page preview. Click a card thumbnail to inspect the full source page.]_ ### Jump to the source PDF To open the original PDF at that passage, click the **top-right corner of the card** or any **highlighted keyword** in the excerpt—not the thumbnail itself. _[Figure: Full PDF page preview with highlighted passage. Full-page preview with the matched passage highlighted in context.]_ ### Check the paragraph against library evidence Once Recommended has keywords *and* matching snippet cards, Flowing automatically checks qualifying draft paragraphs against the retrieved evidence. A left-gutter mark opens “Check paragraph against library evidence”: the paragraph on the left, red / green / yellow cards on the right. The next chapter walks through the left-gutter mark and the “Check paragraph against library evidence” dialog, with screenshots of the three card kinds. > **Note:** Automatic snippet recommendations while writing use tokens on paid tiers. *Manual* keyword search does not. ### Evidence verification Source: https://flowing.works/en/docs/evidence-verification After Recall has keywords and matching snippet cards, Flowing automatically checks qualifying draft paragraphs against the retrieved library evidence. A left-gutter mark opens red / green / yellow cards—an auditor, not a rewriter. Evidence verification is the step after *Recall*: once recommended search has highlighted keywords and matching library snippets, Flowing checks qualifying draft paragraphs against that retrieved evidence set. It does not search the open web, and it does not rewrite the paragraph. ### When paragraphs are checked 1. **Wait for Recall snippets** — Open the Recommended tab and keep writing until keywords appear with matching snippet cards. Verification starts only after both are present. 2. **Write a qualifying paragraph** — The paragraph must contain a highlighted keyword that has snippet cards. Short or mostly-formula blocks are skipped. In rich text, a longer prose paragraph qualifies; in LaTeX, a blank-line block of at least three non-empty lines of prose qualifies. 3. **Leave the paragraph** — While the caret stays inside a paragraph you are still editing, Flowing holds that paragraph—it will not verify it or change its mark. Move the caret out (or wait until the paragraph no longer qualifies) to let the check run. ### The left-gutter mark A colored bar and count appear in the left margin of a checked paragraph. The bar is segmented by card kind (*red* / *yellow* / *green*). A spinner means the check is still running. A slate mark with 0 still opens the dialog, so you can see the empty state. _[Figure: Left-gutter evidence mark beside a draft paragraph, with Recall snippet keywords in the sidebar. The left-gutter bar and count mark a paragraph that has been checked against library evidence.]_ ### Open the evidence dialog Click the mark to open “**Check paragraph against library evidence**”. The left pane is the paragraph (keyword highlights plus claim underlines). The right pane is the cards, with the same library evidence chips as polish—click a chip to jump to the PDF. The footer may show a short paragraph summary derived only from the cards that were emitted. _[Figure: Centered dialog titled Check paragraph against library evidence, with the paragraph on the left and a supported evidence card on the right. The dialog pairs the paragraph with red / green / yellow cards. Hover the legend dots for what each kind means versus the retrieved evidence.]_ ### Red, green, and yellow cards Cards are verdicts about a claim versus the retrieved evidence set (one or two complementary passages)—not a relevance score, and not “the paper is on the same topic.” A card may rest on one or two passages; if combining two fails, a valid single-passage card is kept. *Red* — contradicted by evidence: the claim conflicts with the retrieved evidence in polarity, number, mechanism, or scope. The quoted claim must map onto the paragraph. *Green* — strongly supported: the source directly entails the same claim, polarity, and scope. Paraphrase or a quantitative result is enough; extra scientific inference is not. The underline is the most complete contiguous claim that is fully entailed—not a keyword, not a stitched list, and not widened for readability. *Yellow* — useful qualifier / extension: the evidence supplies a condition, limitation, comparison, mechanism, boundary, or scope that is relevant and not already in the paragraph. Yellow is not “the evidence does not support this claim.” Absence of support is not a card. > **Note:** Purple or blue highlights are Recall keyword or entity matches, not a fourth verdict. Cards omit bibliography-only rows, style or grammar notes, invented papers or quotes, and “insufficient evidence” cards. ### What it does not do No card does not mean a claim passed verification: relevant passages may not have been retrieved, or the paragraph may not qualify for a check. Verification only checks snippets already recalled from the libraries you selected—not the open web, and not papers that were never imported. It does not insert or change manuscript text. Unselecting a library drops that library’s cards immediately without a new request; reselecting restores the cached result. Switching documents loads that document’s cache, so re-entering the same draft with unchanged paragraph text does not re-verify. > **Tip:** Judgment stays with you. Use the cards to **strengthen, hedge, reframe, or delete** a claim before it settles—then polish or continue writing if you want a rewrite. ### Ask with context Source: https://flowing.works/en/docs/ask-with-context Type @ to search your library, attach text or figure snippets as context, then chat with AI in a continuous thread. Ask with context keeps your references attached to the conversation—*no copy-pasting* between PDFs, your draft, and a separate chat app. ### Invoke library search with @ Open *Ask with context* from the sparkle button, a text selection, or a Recall snippet card. In the message box, type **@** to open *Add context from library*. Search by title or file name, browse papers grouped by library, and pick the document to ground the conversation—the picker stays open until you attach a passage or dismiss it. _[Figure: Typing @ in the chat input to open library search. Type @ in the message box to search and pick papers from your library.]_ ### Build context snippets After choosing a paper, **select text** in the PDF preview or **box a figure or chart region**—each selection becomes a context chip above the input (text excerpt or image snippet). Attach multiple chips from one paper or across your library before you ask. _[Figure: Selecting a text passage from a PDF to attach as context. Drag to select text—the excerpt appears as a context chip.]_ _[Figure: Boxing a chart or figure region to attach as image context. Box a figure or chart region—the capture appears as an image context chip.]_ ### Ask and follow up Send your question; the model answers using the attached chips. **Keep chatting**—follow-up messages stay in the same thread with context preserved, so you can clarify, compare sources, or dig deeper without re-attaching. _[Figure: AI chat response with attached context chips and follow-up messages. The model responds using your context chips; continue the conversation in the same thread.]_ ### Polish Source: https://flowing.works/en/docs/polish Grammar polish fixes issues on hover in the editor; content polish rewrites a paragraph with library context you can add, remove, accept, or dismiss. Flowing has two polish modes: **grammar polish** for quick in-place fixes as you hover the draft, and **content polish** for paragraph rewrites grounded in your library—with context snippets you control before accepting a suggestion. ### Grammar polish While you write, **hover** over text in the rich-text or LaTeX editor. Flowing flags grammar and style issues inline; click the fix control to apply a correction on the spot—no panel switch required. _[Figure: Hovering draft text to see an inline grammar fix. Hover in the editor to surface a grammar issue, then click to fix it.]_ ### Content polish Open the *Polishing* tab on the writing page to enter paragraph-pick mode. **Hover** a paragraph to see a dashed outline (rich text and LaTeX both work), then **click** to select it—the paragraph and its keywords appear in the right panel. _[Figure: Hovering and selecting a paragraph in content polish mode. Hover a paragraph for the dashed outline, then click to select it.]_ Flowing pre-loads up to *10 library passages* matched to the paragraph’s keywords as context chips. **Remove** any chip you do not want, or **add** more from your library before generating suggestions. _[Figure: Adding and removing library context snippet chips before polishing. Ten matched passages are attached by default—add or remove chips before you run polish.]_ ### Add your context Beyond the auto-matched chips, you can **add context by hand** before you generate. In the context area, type **@** to search your library and pick a paper—then **select text** or **box a figure region** in the preview. Each pick becomes a context chip (text excerpt or image snippet). Combine hand-picked sources with the matched passages until the context reflects what you want the rewrite to draw on. _[Figure: Adding text, images, and @context before content polish]_ ### Result card with evidence Each suggestion appears as a *result card*. The proposed rewrite sits at the top; below it, the **evidence panel** lists the library passages that ground the text—excerpt text, source title, and page thumbnails you can preview. Review the basis, then **Accept** to apply the rewrite to your paragraph or **Dismiss** to skip. Flowing never auto-replaces the whole paragraph without your confirmation. _[Figure: Content polish result card with evidence panel and Accept or Dismiss]_ > **Note:** Content polish uses monthly tokens on paid plans. Check per-feature usage in the app billing panel. ### Continue writing Source: https://flowing.works/en/docs/continue-writing Generate new text after your cursor—grounded in your library, with the same context chips as content polish. Continue writing extends your draft *forward from the cursor*. Like content polish, it gathers library passages as context chips you can add or remove—but instead of rewriting a paragraph, it **inserts new sentences after the cursor** without replacing text you already wrote. ### Place the cursor Open the *Continue* tab on the writing page. Click where the new text should begin—typically the end of a paragraph or section. Flowing uses the **preceding draft** (including the paragraph you are in) as the anchor for what to write next. You can also hover a paragraph for the dashed outline and click to focus context, same as content polish. _[Figure: Cursor placed where continuation should be inserted. Place the cursor at the insertion point before generating continuation.]_ ### Context snippets This step works like content polish: Flowing pre-loads up to *10 library passages* matched to nearby keywords as context chips. **Remove** chips you do not want or **add** more from your library, then generate continuation candidates. _[Figure: Adding and removing library context chips before continue writing. Ten matched passages by default—tune context chips before you generate.]_ ### Add your context Beyond the auto-matched chips, you can **add context by hand** before you generate. In the context area, type **@** to search your library and pick a paper—then **select text** or **box a figure region** in the preview. Each pick becomes a context chip (text excerpt or image snippet). Combine hand-picked sources with the matched passages until the context reflects what you want the continuation to draw on. _[Figure: Adding text, images, and @context before continue writing]_ ### Result card with evidence Each suggestion appears as a *result card*. The proposed continuation sits at the top; below it, the **evidence panel** lists the library passages that ground the text—excerpt text, source title, and page thumbnails you can preview. Review the basis, then **Accept** to insert the new sentences at the cursor. Adjust context or move the cursor and generate again to keep extending a section—nothing is auto-inserted without your confirmation. _[Figure: Continue writing result card with evidence panel and Accept]_ > **Note:** Continue writing uses monthly tokens on paid plans, same pool as content polish and Ask with context. ### References Source: https://flowing.works/en/docs/references Browse the bibliography in the References sidebar, jump between in-text citations and cards, and insert papers from your library with @. Flowing parses the *References / Bibliography* block in your manuscript and lists every entry in the *References* sidebar—aligned with the citation numbers in your draft. ### References sidebar Open the *References* tab in the right panel to see every bibliography entry as a card: citation number, title, author, and year. Click a card to jump to the matching entry in the manuscript’s References section. _[Figure: References sidebar listing bibliography entries next to the manuscript References section. The References tab mirrors the bibliography in your draft—click a card to jump to that entry.]_ ### Jump between citations and cards Click an in-text citation such as **[61]** to highlight the matching card in the sidebar. From a card, use **Citation 1**, **Citation 2**, … to jump back to each place that reference appears in the draft. _[Figure: In-text citation linked to the matching References sidebar card. In-text citations and sidebar cards jump to each other.]_ ### Insert a reference with @ Type **@** in the editor—or click the **@** button in the toolbar—to search your library and insert a PDF as a citation. Flowing places the numbered marker in the text and appends the full bibliographic entry, keeping the References tab in sync. _[Figure: Toolbar @ control used to insert a citation from the library. Type @ or use the toolbar control to insert a citation from your library.]_ ### Search your library The picker opens *Add context from library*. Search by title or file name across the libraries selected for this document, then choose a paper to insert. Papers already in your library are marked so you can confirm the match before citing. _[Figure: Add context from library modal for picking a PDF to cite. Search selected libraries and pick a PDF to insert as a reference.]_ > **Note:** Flowing extracts the first References / Bibliography / Works Cited block in the document. Supported formats include numbered styles (Nature, IEEE, Vancouver, ACS), BibTeX and **\bibitem**, and author–year (APA/MLA). ## FAQ Source: https://flowing.works/en/faq Answers about the product, your library, large models, and billing. ### Product & usage **Q: Do I have to import a research library?** It is not required—Ask with context and evidence-grounded revision still work without a library. We strongly recommend importing your reference PDFs early: recall by keyword, contextual Q&A, and evidence-grounded revision all revolve around your research library. Importing your library is the essential foundation for getting the most out of Flowing. **Q: How is Flowing different from ChatGPT or generic AI writing tools?** Flowing is a desktop tool for writing with context, not empty chat. Core workflow: keyword-grounded recall, Ask with context (attach draft or library passages in a few clicks), evidence-grounded revision, and evidence-backed continue writing—each suggestion linked to local library snippets you can verify on the card. Compared to generic platforms like ChatGPT, that grounding delivers stronger reliability and convergence: continuation stays on your thread with cited evidence, not open-ended filler. Context-based polish draws on the same attached passages, so rewrites stay more unified in style and tone—closer to your manuscript, with fewer unsupported claims. **Q: What do the red, green, and yellow cards mean?** After Recall has keywords and matching library snippets, Flowing can check a qualifying draft paragraph against the retrieved library evidence. This is an auditor, not a rewriter: it does not insert text. Red: the claim contradicts the retrieved evidence (polarity, number, mechanism, or scope). Green: the source directly entails the same claim, polarity, and scope—paraphrase or a quantitative result is enough; extra scientific inference is not. Same-topic overlap alone is not support. Yellow: a useful qualifier or extension from the evidence that is not already in the paragraph. Yellow is not “evidence is missing.” Cards are claim-scoped against one or two complementary passages, not 1:1 with a snippet. Purple keyword highlights mark Recall matches; they are not a fourth verdict. Cards omit bibliography-only rows, style notes, invented papers, and “insufficient evidence” rows. Judgment stays with the researcher. **Q: Which models does Flowing use? Do I need my own API keys?** Flowing currently routes through DeepSeek and Kimi (Moonshot), split by feature. Automatic snippet recall (snippet recall, proper-noun extraction, domain filtering, and related tasks) uses DeepSeek. Ask with context, evidence-grounded revision, and paragraph evidence verification use Kimi—including multimodal understanding when PDF figures are cited. You do not configure API keys yourself; Flowing’s backend proxies the calls. Default model versions may change with releases (currently DeepSeek and Kimi K2 family). Pro+ can unlock stronger model options. **Q: What editing modes are supported? Can LaTeX users use Flowing?** Flowing supports rich text and LaTeX side by side with switching anytime—equations, tables, figures, and more. Live PDF preview shows exports exactly as they will look, whether you prefer Word-style editing or LaTeX source. From the documents list, import an existing manuscript from Word (.docx) or PDF, and export the draft as Word or PDF. **Q: How many papers do I need in my library?** Anything from dozens to thousands of PDFs works. A larger library gives recall and polish more to draw on. Even a focused set of core papers is enough to experience the full keyword → context → polish loop. Start with references most relevant to your current project. **Q: Where are my papers and drafts stored? Are they uploaded to the cloud?** Flowing is a desktop app: your library, vector index, and draft files stay on your machine by default. When you use AI features (retrieval, Q&A, polish), only the necessary context snippets are sent to model services—we do not host your full PDF library on our servers. ### Plans & billing **Q: Monthly subscription vs prepaid tokens—what is the difference?** Monthly plans give a fixed token allowance each cycle; unused balance resets when the cycle rolls. Best for steady, frequent writing. Prepaid packs are one-time purchases that never expire. Best for short, intense bursts—deadline crunches or revision sprints. They stack together: each cycle spends your monthly allowance first, then prepaid covers overflow. Write almost every day? Pro/Pro+ monthly. Occasional spikes? Prepaid packs. **Q: Which plan fits me—Free, Go, Pro, or Pro+?** Free: light trial with unlimited manual keyword search and limited AI. Go: $5/mo with 1M tokens/month — enough for regular writing (about 200–300 Ask or 100–200 polishes). Pro (recommended): $10/mo with 4M tokens/month — enough to write a paper in earnest (about 800–1,200 Ask or 500–700 polishes), plus generous recall, Ask with context, and Evidence Card. Pro+: heavy research workflow, strongest models, and 16M tokens/month (4× Pro; about 3,000–5,000 Ask or 2,000–3,000 polishes). Check remaining balance in the app before upgrading. **Q: What uses tokens? Is manual search billed?** Manual keyword search does not consume monthly tokens on Free and above—you can search your library freely. Tokens apply to Automatic snippet recall (automatic snippet recommendations while writing), Ask with context, evidence-grounded revision, and other AI features sharing one monthly pool. The app shows per-feature usage and remaining balance. **Q: How do I activate, subscribe, or pay on the website?** Download the desktop app and complete device activation (invitation or free activation). For Pro/Pro+ or prepaid packs, start checkout in the app. To pay in the browser via Stripe, copy the claim code from the app, paste it on this site’s pricing page for the matching tier, and finish payment. Entitlements shown in the app are authoritative.