Self-host Pharos
0.3.0Local-first agentic RAG for team document libraries
- Stars
- 241
- Ports
- 8787 (pharos serve HTTP API), 8900 (inference service (GPU embeddings/reranking))
- Requires
- Qdrant
What is Pharos?
Pharos is a local-first agentic RAG system for team document libraries. It ingests multi-format documents (PDF, scans, docx, pptx, xlsx) via MinerU parsing, provides hybrid retrieval combining dense embeddings, sparse BM25, and reranking, and enforces enterprise-grade access control (ACL) with multi-tenant identities and principals. It offers two consumption modes: an HTTP API (/v1/ask) for closed-pipeline QA with citations, and an MCP exit for agentic RAG where the agent drives retrieval. The system runs as a resident daemon (via systemd or Docker Compose) that owns the GPU models and Qdrant vector store, with a three-tier architecture (inference, pharos app, Qdrant) for horizontal scaling. Documentation is primarily in Chinese with a 12-part learning series.
Should you self-host Pharos?
Pharos scores 44/100 on selfhost.directory (how scoring works) and ranks #80 of 125 in AI. It has 241 GitHub stars.
There was no stable release in the past 12 months, so check the repository's recent activity before you rely on it.
Setup is advanced: plan for several services and some ongoing maintenance. It is MIT-licensed and written mostly in Python.
Its closest alternative on selfhost.directory is ComfyUI (93/100, 143 668 stars). It scores higher overall, so compare the two before you pick. See Pharos vs ComfyUI side by side.
Features
- Agentic RAG
- Backup and restore procedures
- Citations with page-level traceability
- CPU-only test suite
- Dual HTTP + MCP exits
- Enterprise ACL
- Fail-closed security model
- Horizontal scaling (three-tier)Throughput ceiling limited by single GPU forward speed
- Hybrid Retrieval
- Multi-format document ingestion
- Observability
- RAG learning documentation
- Remote inference support
- Session-isolated deduplication
- Systemd supervision
Pharos README
More in RAG
private-gptOpen-source Claude-compatible API layer adding RAG with citations, tools, MCP and SQL queries to any local model server, plus a self-hosted workbench UI.
qurioQurio is a self-hosted, open-source RAG engine that ingests docs and code locally, then serves grounded context to AI coding assistants via MCP to prevent hallucinations.
R2RSelf-hosted agentic RAG system with a RESTful API: multimodal ingestion, hybrid search, knowledge graphs, a Deep Research agent, and user & access management.
ragSelf-hosted, open-source retrieval-augmented generation chat app on txtai and Streamlit: vector RAG plus graph path queries, with your own LLM.
SeekStormOpen-source Rust search engine with BM25 lexical and vector search fused in one hybrid query, multi-tenant REST server, web UI and MCP support. Self-hosted, free forever.
supermemorySelf-hosted AI memory and context server: Supermemory Local extracts facts, builds user profiles, and blends RAG with memory search through one API on port 6767.
surfsenseSelf-hosted, open-source NotebookLM alternative: an air-gapped desktop AI agent that researches, edits and transforms your documents with any local or hosted model.
TaskingAIAn open-source, self-hosted BaaS stack for building LLM agents: unified model APIs, RAG, built-in tools and an admin console, deployed with Docker Compose.
UtopiaSelf-hosted, open-source enterprise world model: a Rust binary plus Postgres/pgvector holding a bitemporal knowledge graph, with agentic RAG chat and an MCP server.
VaneSelf-hosted, open-source AI answering engine: searches via bundled SearxNG and answers with cited sources, using local Ollama models or OpenAI, Claude and Gemini.
vortexSelf-hosted RAG application for document chat with multiple LLM providers. Combines vector and keyword search, re-ranks results, and shows source citations.
WeKnoraTencent's open-source knowledge framework for teams: self-hosted RAG Q&A with citations, agents with sandboxed skills, and auto-generated wikis on swappable LLMs.
0 comments