axon-rag 0.4.4

Your documents, answerable. On your hardware.

Drop in PDFs, code, spreadsheets, or URLs — ask anything, get cited answers from a local LLM.
Nothing leaves your machine.

🤔 Why Axon?

Most RAG tools make you choose between cloud power and data privacy. Axon is local-first — full capability with zero egress when you run on Ollama or vLLM; cloud providers (OpenAI, Gemini, Grok, GitHub Copilot, Ollama Cloud) stay optional.

• 🔒 Private by default — local inference via Ollama or vLLM is the recommended path; cloud providers (OpenAI, Gemini, Grok, GitHub Copilot, Ollama Cloud) remain opt-in. No telemetry. Strict offline / air-gap mode shuts every outbound call off entirely.
• 📄 Ingest anything — 54 file formats (PDF, DOCX, Jupyter, code, images, URLs) in one command. SHA-256 dedup skips unchanged files.
• 🤖 Works in your tools — @axon in Copilot Chat, MCP for Claude Code / Codex / Gemini CLI / Cursor, Graph panel in VS Code or your browser.
• 🤝 Built for teams — share your knowledge base with signed, revocable read-only keys. Sealed (AES-256-GCM encrypted) sharing works safely through OneDrive, Dropbox, and Google Drive. Per-user permissions, full audit trail, no extra infrastructure. Quick setup →
• 🕸️ See your knowledge as a graph — interactive 3D entity-relationship graph. Embedded webview in VS Code; opens in your browser everywhere else. Click any node to inspect its supporting chunks and source excerpt.
• 🔬 Production-grade retrieval — hybrid search, reranking, HyDE, multi-query expansion, and automatic web fallback. Zero manual tuning.

✨ Capabilities

⚡ Quick Start

pip install "axon-rag[starter]"   # Python 3.10+. Sealed sharing + extra loaders. Web GUI ships with axon-api.
axon                              # First run auto-launches the setup wizard, then drops into the REPL.

That's it. The wizard configures your LLM provider, embedding model, and retrieval defaults; subsequent runs go straight to the REPL.

See it run — first-run wizard, ingest, and a cited query

If something doesn't look right:

axon --doctor                     # Health checks: Python, Ollama, model pulled, store writable.
axon update                       # Check PyPI and upgrade the package + VS Code extension together.

Local inference uses Ollama or vLLM (self-hosted). Cloud providers (OpenAI, Gemini, Grok, GitHub Copilot, Ollama Cloud) work via API keys.

→ Setup guide for VS Code, MCP, and cloud providers →

🔐 Sealed Sharing Quick Start

Share an encrypted knowledge base through OneDrive, Dropbox, or Google Drive. Cloud providers see only ciphertext.

pip install "axon-rag[sealed]"   # install sealed extra on both machines

Owner (5 steps)

axon --store-init "/path/to/OneDrive/AxonStore"  # 1. point store at sync folder
axon --store-bootstrap "your-passphrase"          # 2. bootstrap master key (once per machine)
axon --project-new research                       # 3. create project + ingest
axon --project research --ingest /docs
axon --project-seal research                      # 4. encrypt in place (≈1 s per 100 MB)
axon --share-generate research alice              # 5. print SEALED1:... string — send to grantee

Grantee (3 steps)

axon --store-init "/path/to/OneDrive/AxonStore"   # 1. same shared folder
axon --share-redeem "SEALED1:..."                  # 2. redeem — DEK stored in OS keyring
axon --project mounts/owner_research "question"   # 3. query; Axon decrypts to temp, wipes on exit

→ Full Sharing Guide — OneDrive setup, revocation, headless/Docker grantees, filesystem compatibility matrix.

🚀 Entry Points

Browser UI: start axon-api and open http://localhost:8420/gui/. No extra command or dependency needed — the GUI ships with the server.

Deprecation: axon-ui (Streamlit) is deprecated and will be removed in a future release. It is no longer part of [starter] or [all]; install it explicitly with pip install "axon-rag[ui]" if you still need it.

Single-instance routing. If an axon-api server is already running, the axon CLI detects it (a quick /health/ready probe) and routes store-mutating commands — --ingest and project create / delete / switch — through that server instead of opening a second in-process copy of the same store. This avoids two processes racing on the vector-store files, skips reloading the embedding model on every CLI call, and keeps a single writer for both the store and the logs. Pass --local to force an in-process brain, or point the CLI at a specific server with the AXON_API_BASE environment variable (or the api_host / api_port config fields).

Relatedly, axon-api itself refuses to start a second server on a store another axon-api is already serving (one writer per store) — set AXON_ALLOW_MULTIPLE_SERVERS=1 if you really want two.

🔌 VS Code + GitHub Copilot

Install the bundled VSIX to unlock the @axon chat participant, Knowledge Graph panel, Code Graph panel, and Governance dashboard — directly inside VS Code alongside Copilot.

Extensions panel  →  "..."  →  Install from VSIX...
→  run `axon-ext`  (or install from VSIX manually)

Or connect via MCP for Copilot agent mode — point .vscode/mcp.json at axon-mcp and all 51 tools appear in the agent hammer menu automatically.

The VS Code extension surfaces 39 LM tools to Copilot Chat, covering core RAG operations, sealed-store security, sharing, and governance.

Full setup guide →

🐍 Use Axon from Your Python Agent

Drop-in retrievers for LangChain and LlamaIndex agents — no REST round-trips, no extra process. Both wrap the same AxonBrain.search_raw() codepath the REST and REPL surfaces use, so hybrid search, reranking, HyDE, multi-query, and the GraphRAG budget apply automatically.

# pip install "axon-rag[langchain]"
from axon import AxonBrain, AxonConfig
from axon.integrations.langchain import AxonRetriever

brain = AxonBrain(AxonConfig.from_yaml("config.yaml"))
retriever = AxonRetriever(brain=brain, top_k=5)

docs = retriever.invoke("what does the project do?")  # list[Document]

# pip install "axon-rag[llama-index]"
from axon.integrations.llama_index import AxonLlamaRetriever

retriever = AxonLlamaRetriever(brain=brain, top_k=5)
nodes = retriever.retrieve("what does the project do?")  # list[NodeWithScore]

Per-call overrides (e.g. force HyDE for one question): retriever.with_overrides({"hyde": True}).invoke(query). From async code, await retriever.aretrieve(query, hyde=True, top_k=8) accepts the same flags as kwargs without rebuilding the retriever (v0.4.1).

📚 Documentation

Getting started

Reference

Deep dives

🔒 Security

Ingestion is sandboxed to a configurable base directory (RAG_INGEST_BASE). Requests outside it are rejected with 403. See SECURITY.md.

📄 License

MIT — see LICENSE.