EdgeQuake
Graph-RAG in Rust: one binary, one Postgres, any LLM — cloud or fully local.
Turn documents into a knowledge graph and get answers with sources, not just similar chunks.
Quick start · Pick your LLM · How it works · Honest numbers · Docs
Quick start
Needs Docker. Nothing else to install: no Rust, no Node.js, no build.
curl -fsSL https://raw.githubusercontent.com/raphaelmansuy/edgequake/edgequake-main/quickstart.sh | sh
Then open http://localhost:3000. The wizard detects local servers (Ollama, LM Studio, oMLX, MLX-LM, llama.cpp) or takes an OpenAI / Anthropic key. Ports bind to 127.0.0.1 only.
# Headless / CI
quickstart.sh --yes --provider ollama
quickstart.sh --yes --provider omlx --base-url http://host.docker.internal:9050
Ask your first question (after dropping a PDF into the UI, or via the API):
curl -X POST http://localhost:8080/api/v1/documents/upload -F "[email protected]"
curl -X POST http://localhost:8080/api/v1/query \
-H "Content-Type: application/json" \
-d '{"query": "What are the main concepts?", "mode": "hybrid"}'
Heads-up: the Ollama path pulls chat and embedding models (multiple GB the first time). Use a cloud key if you want the fastest first answer.
Pin a version: EDGEQUAKE_VERSION=0.38.4 sh quickstart.sh. More options: Docker deployment options · Enable login.
Pick your LLM
EdgeQuake is not tied to one vendor. Configure a Connection (URL + key, encrypted at rest) in Settings → Connections, or by environment variable.
| You want | Use | Guide |
|---|
| Best quality, no hardware | OpenAI, Anthropic, Gemini, Mistral | providers |
| Private, runs on your Mac/PC | Ollama, LM Studio, oMLX, MLX-LM, llama.cpp, vLLM-MLX | Ollama · oMLX |
| Cloud LLM + local embeddings | Hybrid (EDGEQUAKE_EMBEDDING_PROVIDER) | model roles |
| Your own gateway | Any OpenAI- or Anthropic-shaped server | generic |
Each workspace can assign a different model to every role (extract, query, keyword, summary, embedding, vision). POST /api/v1/providers/test and edgequake doctor tell you what is actually reachable instead of guessing.
Connections, doctor, and encrypted keys land with v0.33.0 (schema 169, already on main). See upgrade-to-0.33.0.
Why Graph-RAG
Vector search finds similar text. It struggles when the answer lives in the relationships between things: "who reports to whom", "which contract affects which product", "what themes run across these 200 papers".
EdgeQuake implements the LightRAG algorithm in Rust. Documents become entities and relationships; at query time it combines vector search with graph traversal.
How it works
Document ─► chunks ─► LLM entity + relation extraction ─► knowledge graph
│ │
pgvector (embeddings) + Apache AGE (graph)
│ │
Question ─► keywords ─► vector search + graph walk ─► LLM ─► answer with sources
| Mode | Best for |
|---|
naive | Keyword-like lookups |
local | Specific entities and their neighbours |
global | Themes and high-level questions |
hybrid (default) | Balanced, comprehensive answers |
mix | Weighted vector + graph blend |
bypass | Direct LLM, no retrieval |
Details: feature tour · LightRAG deep dive · architecture.
Honest numbers
We publish what we measured, including where we tie. Benchmark: GraphRAG-Bench medical-mid, EdgeQuake vs LightRAG, same corpus and models (make bench, n=200).
| Metric | EdgeQuake | LightRAG (Python) | Verdict |
|---|
| Answer accuracy (Acc) | 0.792 | 0.786 | Statistical tie (95% CI includes 0) |
| Cold query p50 | 4447 ms | 4359 ms | ≈ tied (1.02×) |
| Warm query p50 (LLM + embed cache) | 82 ms | 993 ms | EdgeQuake faster when cached |
| Evidence recall | 0.932 | 0.949 | LightRAG ahead |
So the pitch is not "more accurate than LightRAG". It is the same retrieval quality, productionised: a single Rust binary, Postgres as the only datastore, multi-tenant isolation, auth, a web UI, MCP, SDKs, and explicit, tested upgrades. Methodology and caveats: EQ vs LightRAG · vs GraphRAG · vs traditional RAG.
Built for production
- One datastore for the graph. PostgreSQL 16/17/18 with pgvector. No separate vector DB or graph DB to run. PDF bytes, uploads, page images, and extracted markdown can move to a private bucket. Text and chunks stay in Postgres. The default is still the database. See originals.
- Safe upgrades. The API never migrates your database.
edgequake migrate is explicit, progressive, and replay-tested from every published release (SPEC-150, upgrading).
- Secure defaults. Quickstart binds to loopback; secrets are generated; connection keys are AES-256-GCM encrypted and write-only (SECURITY.md, SPEC-163).
- Multi-tenant. Fail-closed workspace isolation for query, delete, and recovery. Built-in auth, SSO (Keycloak/OIDC), and audit logging.
- PDF vision pipeline. Text mode by default (pdfium embedded); vision mode for tables, scans, and complex layouts, with automatic fallback.
- Honest health.
/health probes real providers; edgequake doctor checks DB, schema, roles, and security posture.
- Multi-arch images.
linux/amd64 + linux/arm64 on GHCR, every release.
Integrations
Documentation
Contributing
Contributions are welcome, from typo fixes to new providers. Start with CONTRIBUTING.md; for a local build run make install && make dev (see AGENTS.md and the pre-delivery checklist).
Acknowledgments
EdgeQuake implements the LightRAG algorithm by Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. Also inspired by Microsoft's GraphRAG.
License
Apache License 2.0, see LICENSE. Copyright 2024-2026 Raphaël MANSUY
Star History
