Laurent00TT/PharosRAG

pip install -e '.[dev]' (src-layout, editable install)
docker compose --env-file .env.compose up -d --scale pharos=3 (three-tier scale-out)

Usage

• •python -m pharos ask "What do we have on X?" (closed-pipeline QA with citations)
• •Admin: python -m pharos index --corpus <parsed_dir> --dest ~/rag_real (stop daemon first — single-client lock)
• •For Claude Code agentic mode: copy .mcp.json.example to .mcp.json, fill in PHAROS_API_KEY and local paths
• •python -m pharos health; no daemon? use 'pharos mcp --direct' (loads GPU model itself)

Key features

• ✓Two exits: HTTP API (pharos serve, /v1/ask: retrieve → grounding → DeepSeek → cited answer) and MCP (pharos mcp) sharing one tool contract (toolcore)
• ✓Enterprise ACL: X-API-Key → identity → tenant/principals, hard filtering at retrieval; fail-closed (unknown/absent key = 401)
• ✓Runs over 77 real documents and 7,652 chunks; every retrieved passage tagged trust: untrusted against prompt injection
• ✓Resident daemon design (MCP is a thin HTTP adapter); three-tier scale-out (inference GPU / pharos app / Qdrant server)
• ✓Ships a 12-piece Chinese-language RAG learning and interview-prep series (4,122 lines)

Best for

Teams that need a locally-hosted, access-controlled RAG knowledge base over internal documents with page-level citations, served to both humans (HTTP) and agents (MCP).

Caveats

• ⚠Documentation is in Chinese (code, CLI, and commit history are in English)
• ⚠Runs under systemd; daemon must be running for MCP mode (or use pharos mcp --direct)
• ⚠Requires DEEPSEEK_API_KEY for the /v1/ask closed-pipeline generation
• ⚠HTTPS and public exposure are explicit non-goals — tunnel in for remote access; binds 127.0.0.1 by default
• ⚠Throughput ceiling is the forward speed of a single inference GPU card (GPU serialises); replicas add concurrency, not throughput
• ⚠Single-client lock on embedded Qdrant — stop the daemon before indexing