KnowledgeForgeAsk your documents. Get answers that cite the page.
KnowledgeForge reads the PDFs, Office files, scans, e-books and recordings you already have, files them, and answers from them in plain words — every claim cited, every citation one click from the page it came from. It runs on your own computer. Nothing is uploaded.
Download for Windows See the editionsAnyone whose knowledge is in documents that no search box reads.
- Engineers and researchers with shelves of datasheets, standards, papers and manuals, who need the answer and the page it is on.
- Professionals whose work lives in contracts, reports, specs and scans — and who cannot send those files to a cloud AI.
- Students and writers who want to ask a hundred papers a question and get an answer that cites them.
- Households and labs that want one catalog on one always-on machine, open to everyone in a browser.
A research assistant that has read your whole shelf.
Point KnowledgeForge at your documents once. It reads them, files them and remembers where everything is — then answers from them, and shows you exactly where.
In every copy, free
Reads your documents properly
PDFs with their layout, tables and scanned pages; Word, PowerPoint and Excel; web pages, Markdown, CSV, e-books, e-mail and pictures of text. Drop in files, whole folders or zip archives. Keep them where they are, or copy them into the library.
Chat with citations
Ask in plain words. The answer is written from your documents, every claim cited — one click opens the page. Ask the whole library or a single document. Follow-ups keep the thread, and when your library doesn't cover a question, it says so.
Search that finds the passage
Not just the file: the paragraph that answers you, ranked by a second model that reads each candidate against your question. The scan you forgot you had, found by what it says, not what it was called.
A catalog that files itself
Every document gets a summary, a category, its subject and key topics, written by the local AI — so a drawer of unnamed PDFs becomes a library you can browse. Read any document beside its passages, open it in Word or your PDF reader, or download the original.
AI matched to your hardware
KnowledgeForge looks at your computer and picks the local model that suits it — Qwen3 from 1.7B on a laptop to 30B on a graphics card with memory to spare. NVIDIA, Vulkan or the processor alone. No setup, no API keys.
As many libraries as you need
One for work, one for a project, one for a hobby. Each library is a folder you choose — back it up, copy it, move it — and switching takes a click.
With Pro 14 days free in every copy
Research that shows its work
Give it a topic. It finds sources on the web and among scholarly papers, reads them, writes a report with a citation on every claim — then checks the report against those same sources, claim by claim. How the check works.
Recordings, cited by the moment
Lectures, meetings, interviews and podcasts transcribed on your computer, nothing sent anywhere. Ask them like any document: the answer cites the moment it was said, and the player opens there.
Your library in your editor
Claude Code, Cursor, Windsurf and Claude Desktop can search your library, ask it, read from it and have it research for you — through its MCP server, without your documents leaving your computer. Hold every editor to one corpus if you like.
Watched folders
A scanner's output, a downloads folder, a shared drive: new, changed, moved and deleted files followed by themselves, on network drives too. The library keeps itself up to date.
Your own models
Give chat, summaries or research to Ollama or LM Studio on another machine, or to OpenAI, Anthropic or any OpenAI-compatible service — beside the built-in model. The app always says what leaves your computer.
Documents kept up to date as you edit
Edit a document in Word, LibreOffice or your editor. KnowledgeForge notices, marks it changed, and indexes it again in one click — or by itself when you save. Corpora of any size, CHM and older Office formats, and faster ingestion tuned for your machine come with Pro too.
A report you can trust, because it was checked against its own sources.
Most AI research tools write fluently and cite loosely. KnowledgeForge writes the report, then pulls it apart into individual claims and checks each one against the text of the page it cites.
- The arithmetic is recomputed. If the report says 1,000,000,000 × 128 × 4 bytes is 256 GB, the check works it out, finds 512 GB, and says so.
- Every claim is matched against its source. One the page does not support is marked unsupported — as is one with no citation at all.
- Speculation dressed as fact is caught. "Typically", "industry standard", "a rule of thumb": a number resting on one of those is flagged as high risk.
- Contradictions are put side by side, and a vendor's documentation outweighs a forum post.
- Gaps are stated, not filled. What cannot be proven sends the run searching again; what still cannot be found is written down as a gap rather than invented.
Sources come from the web — DuckDuckGo, or a SearXNG you run yourself — and from OpenAlex, arXiv, Europe PMC and Semantic Scholar for papers. Research the web, your library, or both. Export the report to Word, PDF or Markdown with its citations as footnotes, keep the pages worth keeping with where they came from, or file the report itself in your library.
- Supported
Eight channels of 64 bits at 4800 MT/s give 307 GB/s of peak bandwidth 3
Arithmetic recomputed: 8 × 64 ÷ 8 × 4800 × 10⁶ = 307.2 GB/s. Matches the source. - Arithmetic
A billion 128-dimension float vectors take 256 GB 5
Recomputed: 1,000,000,000 × 128 × 4 bytes = 512 GB, not 256 GB. Rewritten. - High risk
DDR5-4800 is typically the choice for this class of board
"Typically" rests on no source. Searched again: two more pages fetched, none settles it. Stated as a gap. - Background
Bandwidth scales with the number of channels populated
Foundational knowledge, labelled so; counted beside the score, not inside it.
From a folder of documents to answers, in three steps.
-
Install and let it set itself up
A small installer, no administrator rights. On first run a setup screen downloads what your computer needs — showing every step — while a short wizard and tour show you around.
-
Add your documents
Drop in files, folders or archives. The processing queue shows each document being read, indexed and summarized, page by page. Search works before the first summary is written.
-
Ask, and check the source
Chat for an answer in plain words, or search the catalog for the passages themselves. Either way, the page it came from is one click away.
Free to keep. Pro when you need more.
Every copy of the app starts with 14 days of Pro, whole — no key, no card, no account. After that it is Free, unless you buy a key, and everything you made during the trial stays searchable, readable and exportable.
Free Available now
For one person on their own computer, with no time limit.
- Chat and search with page citations
- A catalog that files and summarizes itself
- Local AI chosen for your hardware
- Several libraries, of up to 50 documents each
- One corpus of up to 10 documents, and one research run
- Updates, whenever you choose to install them
Pro On sale soon
Everything in Free, without its limits, and further reach. One payment, one key, two computers, each on its own.
- As many documents and corpora as your computer holds
- The Research Assistant, with the integrity check
- Audio recordings, cited by the moment they were said
- Your library inside Claude Code, Cursor and other editors
- Watched folders, on network drives too
- Your own models: Ollama, LM Studio, OpenAI, Anthropic
- Documents kept up to date as you edit them
- CHM help files and older .doc, .ppt, .xls and .rtf files
- Faster ingestion, tuned for your computer
- Larger models on your NVIDIA or AMD graphics card
Try it first: every copy of the app is Pro for its first 14 days, with no key and no card.
Home-Lab Server On sale soonEarly-adopter price
KnowledgeForge on an always-on machine of your own, reached in a browser from every computer at home or in the lab. What the server adds.
- The Pro features, shared by everyone in the house, with accounts and roles
- Larger models on the server's NVIDIA or AMD graphics card
- Audio and video transcribed on the server, cited by the moment
- Watched folders on the server's disk or on a network share (SMB)
- Deep research that reads many more sources and writes the long report
- Corpus Studio: datasets written from your documents, and fine-tuning on the server's GPU
- One key covers the server and five desktops that connect to it, each with everything Pro has
- More desktops at $39.99 each, added to the same key: one payment, no subscription
Try it first: every server is fully working for its first 14 days; after that, without a key, what is on it stays readable and exportable. One installer for any Linux machine (x86_64) that runs Docker: download it there and run it with sudo.
Early-adopter price while the desktop bridge is being built, for the server and for each desktop added to its key; covers every 1.x update, the bridge included.
Enterprise By contract
A shared knowledge base for an organization, on its own hardware.
- Models fine-tuned on your own documents, so answers speak your field's language
- Custom chat bots that answer from the libraries you choose
- Shared team libraries, with roles and permissions per library
- Audit traceability: a tamper-evident log of activity
- GitLab sync
Enterprise is licensed by contract, sized to your organization. Tell us about it and we will write back.
Licensed by contract
| Free | Pro | Home-Lab Server | Enterprise | |
|---|---|---|---|---|
| Price | $0 | On sale soon | On sale soon | By contract |
| Covers | Your computer | 2 computers; more with another key | The server and 5 desktops; more at $39.99 each | A team, on its own hardware |
| Payment | None | Once | Once | By contract |
| The desktop app connects to your server | — | — | ✓ | ✓ |
| Chat and search with page citations | ✓ | ✓ | ✓ | ✓ |
| A catalog that files itself; local AI chosen for the hardware | ✓ | ✓ | ✓ | ✓ |
| Documents in a library | 50 | No limit | No limit | No limit |
| Corpora | 1, of 10 documents | No limit | No limit, each person's own | No limit |
| Research runs, with the integrity check | 1 | No limit | No limit, plus deep and architecture scopes | No limit |
| Audio recordings, cited by the moment | — | ✓ | ✓ and video | ✓ |
| Your library in Claude Code, Cursor, Windsurf, Claude Desktop | — | ✓ | ✓ for every user, with API tokens | ✓ |
| Watched folders | — | ✓ network drives too | ✓ the server's disk or an SMB share | ✓ |
| Your own models: Ollama, LM Studio, OpenAI, Anthropic | — | ✓ | ✓ plus vLLM and Azure OpenAI | ✓ |
| Documents kept up to date as you edit them | — | ✓ | ✓ | ✓ |
| CHM and older .doc, .ppt, .xls, .rtf | — | ✓ | ✓ | ✓ |
| Faster ingestion, tuned for the machine | — | ✓ | ✓ | ✓ |
| Accounts and roles; everyone's catalog in a browser | — | — | ✓ | ✓ |
| Larger models on a GPU, NVIDIA or AMD | — | ✓ | ✓ | ✓ |
| Corpus Studio and fine-tuning on the server's GPU | — | — | ✓ | ✓ |
| Custom chat bots that answer from the libraries you choose | — | — | — | ✓ |
| Audit traceability: a tamper-evident log of activity | — | — | — | ✓ |
Every edition is one payment, not a subscription. A key covers every 1.x release; 2.0 is the first paid upgrade, at a discount for anyone who holds a 1.x key. Enterprise is licensed by contract, and parts of it are still being built: write to us, and we will say what your contract covers and when. Leave your address under "Tell me about new releases" to hear when a new version is out.
All sales are final. Try KnowledgeForge on your own documents and hardware before you buy — the app is Pro for its first 14 days, and a server fully working for its first 14. Once a key is issued we do not refund, credit or exchange it. This does not affect rights your local consumer law gives you and does not allow to be waived. The license agreement, section 5, has the full terms.
One catalog for the whole house, or the whole lab.
The same KnowledgeForge, on a machine of your own that never sleeps: a GPU box in the closet, a NAS, an old workstation. Everyone signs in from a browser. Everyone's documents are in one catalog; everyone's conversations, corpora and reports are their own.
Bigger models, always on
The server's graphics card runs a larger model than a laptop can, for everyone at once — vLLM on NVIDIA, llama.cpp on AMD. Without a GPU it still runs, and can give each job to a provider instead.
Accounts and roles
Administrators, editors and viewers. Nobody sees another person's conversations, not even an administrator. Each person gets their own API tokens for editors and scripts, with what each may do.
Video as well as audio
Meetings, lectures and recorded calls — MP4, MKV, MOV and WebM as well as MP3, WAV and the rest — transcribed on the server as they arrive. An answer cites the moment, and plays it.
Fed from your shares
Watch a folder on the server's disk, an NFS mount, or an SMB share on your NAS — the server reads it over the network with nothing mounted. New, changed, moved and deleted files are followed by themselves.
Corpus Studio and training
Turn the catalog into training data — harvested, deduplicated, weighted, with synthetic question-and-answer pairs written from your documents — and fine-tune a model on the server's GPU. Serve the result from the same server.
One installer, kept safe
One file installs it on any Linux machine that runs Docker, sets up its own TLS, and schedules backups. Nothing but the web port is reachable from outside. Upgrading is running the newer installer.
The server installer is published on launch day. Leave your address under "Tell me about new releases" above to be told.
Your documents stay on your computer.
Not sent for search, not for summaries, not for answers. The models that read, search and write all run on your own machine — unless you choose, with Pro, to give a job to an AI provider of your own, and then the app says where it goes.
What stays here
Your documents, the index and catalog made from them, your questions, the answers and your conversations. No analytics, no telemetry, no advertising.
When it goes online
To download its components and models when it sets up; once, at its first start, to record when its trial began — the computer's name and a hash that identifies it, nothing about your documents; when you press a license button; and when you check for updates — and, with Pro, for a job you gave an AI provider you chose, or a research run's searches. After setup it works offline.
Updates on your terms
Nothing installs itself. You check, you install, you restart when it suits you — every update signed, so the app only accepts ours.
The full privacy policy and license agreement ship inside the app, in Settings › Legal.
Runs on the computer you have.
A graphics card makes it faster and lets it run a larger model, but a laptop's processor is enough to start.
- System
- Windows 10 or 11, or Linux, 64-bitLinux with glibc 2.35 or newer: Ubuntu 22.04, Debian 12 and later. macOS is coming in November.
- Graphics
- OptionalNVIDIA cards run the largest models; other GPUs through Vulkan; the processor alone works.
- Memory
- 16 GB or more recommendedfor the local AI; search needs less.
- Disk
- About 8 to 28 GBfor the components and models it downloads at setup, depending on your hardware — plus your documents.
- Home-Lab Server
- Linux with Dockerx86_64, 16 GB of memory or more, 20 GB of disk; 60 GB or more with an NVIDIA or AMD GPU for its model. Any browser opens it.
Good to know.
How is this different from uploading my files to a chatbot?
Three ways. Your files never leave your computer: the models that read, search and answer run here. The answer is built from your documents and cites each claim, so you can open the page and check it rather than take it on trust. And it is a library, not a chat window: hundreds of documents, filed and summarized, that you can search by what they say and ask month after month.
Is it really free?
Yes. Free has no time limit, no account and no ads. Every copy starts with 14 days of Pro, whole, so you can try everything on your own documents. After that, Free keeps libraries of up to 50 documents, one corpus of up to 10 and one research run. Whatever you made during the trial stays: a library that has more than 50 documents keeps them all searchable, and takes new ones once it has fewer. A free license is optional: one click in Settings › License registers your computer, which is how we count the people using KnowledgeForge.
What makes the research different from asking a chatbot to research?
The report is checked after it is written. Each claim is matched against the text of the page it cites, arithmetic is recomputed, speculation is flagged, and what cannot be proven is searched for again and then stated as a gap rather than filled in. The Integrity tab shows the working for every claim, and the score says how much of the report its sources actually support. The pages it reads are staged, not indexed: nothing joins your library until you keep it.
What files can it read?
PDFs, including scanned pages and tables; Word, PowerPoint and Excel files; web pages, Markdown, text, CSV and AsciiDoc; EPUB e-books; e-mail saved as .eml; pictures, whose text is read by OCR; and zip and tar archives of any of these. Pro adds CHM help files and the older .doc, .ppt, .xls and .rtf formats, converted on your computer by 7-Zip or LibreOffice, and audio recordings — MP3, WAV, M4A, AAC, FLAC, Ogg and Opus — transcribed on your computer, each passage cited by the moment it was said. The Home-Lab Server transcribes video too.
Does it send my documents to a cloud AI?
Not unless you ask it to. The language model that writes summaries and answers runs on your computer, and so do the models that read and search your documents; nothing about your documents is sent anywhere. With Pro you can choose to give a job — chat, say — to Ollama or LM Studio on another machine, or to a cloud service such as OpenAI or Anthropic. The app then names the host, and sends it only what that job needs: for chat, the question and the passages found for it.
Does it need an internet connection?
Only to set itself up: the first run downloads the components and models your computer needs, and records when its trial began. After that it works offline, and goes online again only when you press a license button, check for updates, or run a web research.
What does a key cover, and can I move it?
A Pro key works on two computers at a time, each on its own: Free and Pro do not connect to a server. For a third computer, buy another Pro key. A Home-Lab Server key covers the server and five desktops that connect to it, each with everything Pro has; more desktops can be added to the same key at $39.99 each. To move a key, press Deactivate this machine on the old computer, then activate it on the new one. Once activated, the license is checked on your computer, offline. All sales are final, so try it first: that is what the 14 days are for.
Is it a subscription?
No. Every edition is one payment, and the key covers every 1.x release — the updates, and for the Home-Lab Server the desktop bridge as it ships. When 2.0 comes it is a paid upgrade, at a discount for anyone who holds a 1.x key; nothing you have stops working if you do not take it, and a 1.x key keeps running 1.x. Desktops added to a Home-Lab Server key are one payment too.
What support comes with it?
Anything wrong with the product — a bug, an install that fails, an update that breaks something — is ours to fix, at no charge: report it or write to us, whichever edition you have, Free included. Help beyond that — setting KnowledgeForge up for you, choosing hardware or a model, configuring your NAS, network or server — is a support ticket at a flat price, asked for by email. Enterprise has its own support in its contract.
Can I run it on my NAS or a server?
That is the Home-Lab Server: one installer for any Linux machine that runs Docker, with an NVIDIA or AMD graphics card or none. A NAS that runs Docker qualifies. TrueNAS and Unraid templates are planned; until then the installer works on either from a shell.
Windows says “Windows protected your PC”. Is that safe?
KnowledgeForge is new, and Microsoft Defender SmartScreen warns about any app that has not yet been downloaded by many people — even one that is signed, as this one is. Click More info: the publisher should read Duvitech LLC. If it does, click Run anyway. If it reads anything else, don't run it, and tell us. Or install it from the Microsoft Store, which asks no such question; it is the same app.
How do updates work?
Never behind your back. In System Status, press Check for updates; if there is one, install it and restart when you are ready — work in the queue picks up again afterwards. You can also have the app check each time it starts.
Is there a version for Linux or Mac?
Linux, yes. The AppImage runs as it is: make it executable (chmod +x) and start it. It needs FUSE 2, the libfuse2 package, which Ubuntu 24.04 and later call libfuse2t64. On Debian and Ubuntu there is also a .deb: sudo apt install ./KnowledgeForge_…_amd64.deb. Both need glibc 2.35 or newer. A macOS version is coming in November.
Start with the documents you already have.
Free for Windows and Linux, 64-bit, with 14 days of Pro to start. On Windows it installs for your user account, with no administrator rights; on Linux the AppImage runs as it is.
DownloadQuestions, or interested in Pro, Home-Lab Server or Enterprise? Write to knowledgeforge@duvitech.com.