EdgeQuake: Graph‑RAG in Rust for Production‑Grade Retrieval

🤔 Curiosity: Why do graph‑aware RAG stacks keep winning?

In production RAG, I keep running into the same wall: vector similarity alone struggles with multi‑hop reasoning—“how does X relate to Y through Z?” EdgeQuake positions itself as a Graph‑RAG engine that keeps the structure of the document, not just its semantics. That immediately maps to how we build systems in games: relationships matter as much as facts.

---

📚 Retrieve: What EdgeQuake ships

EdgeQuake is a high‑performance Graph‑RAG framework in Rust that implements the LightRAG algorithm. The core idea is simple but powerful: extract a knowledge graph during ingestion, then query that graph in addition to vector space.

Key capabilities (from the repo)

• Knowledge Graph Extraction (entities + relationships)
• 6 Query Modes (naive → local → global → hybrid → mix → bypass)
• Rust performance (Tokio async + zero‑copy design)
• REST API + SSE streaming (OpenAPI 3.0)
• React 19 UI with Sigma.js graph visualization
• MCP integration for agent tooling

---

📚 Retrieve: How the LightRAG pipeline works

Minimal API usage (ingest + query)

# Upload a document
curl -X POST http://localhost:8080/api/v1/documents/upload \
  -F "file=@your-document.pdf"

# Query the system
curl -X POST http://localhost:8080/api/v1/query \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What are the main concepts?",
    "mode": "hybrid"
  }'

The ingestion pipeline (as described in README) is essentially:

1) Chunk documents 2) Extract entities/relationships with LLMs 3) Glean (multi‑pass extraction for recall) 4) Normalize entities (dedup) 5) Embed chunks/entities 6) Store in graph + vector backends (AGE + pgvector)

And at query time, you can pick modes depending on the question:

Mode	Best for	Behavior
Naive	keyword‑like lookups	vector similarity only
Local	specific relationships	entity neighborhood traversal
Global	themes	community summaries
Hybrid	balanced	combines local + global
Mix	controlled weighting	blend naive + graph
Bypass	general queries	LLM only

flowchart LR
  A[Document] --> B[Chunking]
  B --> C[LLM Entity Relation Extraction]
  C --> D[Gleaning Pass]
  D --> E[Entity Normalization]
  E --> F[Graph Store AGE]
  E --> G[Vector Store pgvector]
  H[User Query] --> I{Query Mode}
  I --> J[Graph Traversal]
  I --> K[Vector Search]
  J --> L[Hybrid Merge]
  K --> L
  L --> M[LLM Answer]

---

💡 Innovation: Why this matters for game‑scale RAG

1) Graph‑first retrieval matches game design

In narrative systems, relationships are the content. If we retrieve only by semantic similarity, we miss the why behind a fact. Graph‑RAG lets us answer:

• “Which factions influence this NPC?”
• “What events led to the quest chain?”

2) Multi‑mode querying is production leverage

I love that EdgeQuake exposes multiple query modes. We can map modes to UX flows:

UX Flow	Query Mode	Rationale
Quick FAQ	Naive	speed > depth
World‑lore	Global	thematic coverage
Quest chain	Local	relationships matter
GM tools	Hybrid	balanced recall

3) MCP makes it agent‑ready

EdgeQuake ships an MCP server. That’s immediate plug‑in to agent workflows without writing glue code.

---

💡 Innovation: What I’d ship first

1) Internal lore QA tool using Hybrid queries 2) Live‑ops incident debugger with graph traversal (root‑cause explanation) 3) Narrative consistency checker to validate quests against lore graph

Practical evaluation checklist

• Graph extraction quality (precision/recall)
• Query latency under load
• Memory footprint per doc
• Hybrid query gains vs naive

---

References

Code & Docs

• Repo: https://github.com/raphaelmansuy/edgequake
• LightRAG paper: https://arxiv.org/abs/2410.05779
• Docs index: https://github.com/raphaelmansuy/edgequake/tree/edgequake-main/docs

MCP

• MCP: https://modelcontextprotocol.io/