AnyRAG

Your documents. Your machine. Answers with receipts.

Point AnyRAG at any folder of PDFs, Word, Markdown, or text — ask in plain language, get cited answers. Nothing leaves the laptop.

See the product Build something like this

Private research assistant · Local-only

RAG without the cloud tax.

AnyRAG turns a personal library into searchable collections you can question like a research assistant. Each answer cites the section and page; a click opens a source reader so you can verify the claim in context. Generalized from AskJ — same trust model, any corpus you bring.

Ask with citations

Chat over a collection, stream an answer, open the source — Scholar mode keeps claims grounded.

AnyRAG desktop interface with cited answer and source reader
Desktop or localhost web UI — same product, private by default.

How it works

Three workflows that define the product.

Documents beside a laptop showing collection ingest

01 · Create → index → ask

Link a folder. Index. Ask.

Create a collection, point at a documents folder (read-only) or import files into a managed library. Watch ingest progress, then ask. Rescan later when the folder changes — AnyRAG never writes back to a linked folder.

Cited scholar answer in AnyRAG

02 · Scholar with receipts

Neutral answers. Clickable citations.

Hybrid retrieval (keyword + semantic when embeddings exist) grounds each reply. Citations open an in-app source reader so you can expand context and read the full section — not a black-box chatbot.

Quotes mode with verified quotation marks

03 · Quotes you can trust

Verbatim answers, mechanically checked.

Quotes mode returns quotations verified against sources — marked verified or not found verbatim. Optional Persona mode speaks in a single-author voice with a clear AI-reconstruction label.

Capabilities

Collections stay isolated. Your machine stays the boundary.

Collections
Named libraries with description and author. Switch from the header — indexes, chats, and settings never mix. Delete a collection without touching a linked folder.
Ingest
PDF, Word, Markdown, HTML, plain text. Linked folders or managed imports. Smart section detection, editable titles/years, optional section summaries, live progress, Rescan for changes.
Retrieval
Hybrid BM25 + dense search with keyword-only fallback. Diversity so one long file does not dominate. Relevance gate when a question is outside the collection.
Answer modes
Scholar (cited), Quotes (verified verbatim), Persona (opt-in first-person for single-author corpora, clearly tagged).
Research UI
Streaming chat, source reader, suggested starters, quote of the moment, timeline when dates are rich enough, per-collection history, Markdown export.
Models
Embedded llama.cpp + Qwen 1.5B on desktop; Ollama or any OpenAI-compatible local server in development. ONNX encoder for embeddings; optional reranker.
Privacy
Localhost only. No API keys required for core use. Collections, indexes, and conversations under your user data directory.
Desktop packaging
Native window via pywebview; PyInstaller macOS app ships empty — you bring the documents. Bundle includes on-device model stack.

Who it’s for

Private corpora that cannot — or should not — go to a SaaS LLM.

Built local-first

No cloud backend in v1. No OCR for scanned image-only PDFs. One collection at a time. Rescan required after folder changes. Proof point for REDN’s AI Capability Pack.

Python local server · vanilla HTML/JS UI · PyInstaller + pywebview · pypdf / python-docx · hybrid BM25 + dense retrieval · llama.cpp / Ollama · ONNX encoder — AI-accelerated delivery, human-architected for private research.

This is an AI Capability Pack build. Related: AskJ.

Need private AI over your docs?

The AI Capability Pack is how we ship products like AnyRAG — AI-accelerated, human-architected, and yours to own.

Email hello@redn.asia Back to REDN