GraphRAG (Bun + TypeScript)

This project mirrors the Python functionality using a clean architecture in Bun/TypeScript.

Structure

• src/config/env.ts: Environment loading and validation
• src/infrastructure/llm/openai.ts: Chat and embedding clients
• src/infrastructure/graph/neo4j.ts: Neo4j driver and graph I/O
• src/infrastructure/vector/qdrant.ts: Qdrant client, collection ensure, upsert, search
• src/application/extract.ts: LLM JSON extraction → nodes/relationships
• src/application/retriever.ts: Vector search → Neo4j subgraph
• src/application/graphrag.ts: Context formatting and answer generation
• src/index.ts: CLI entry mimicking Python flow

Requirements

• Bun 1.2+
• Neo4j Aura (or any Neo4j instance with TLS)
• Qdrant (Cloud or self-hosted)
• OpenRouter (for chat) and an embeddings provider (OpenAI or compatible)

Environment

Create a .env in the project root:

# Neo4j
NEO4J_URI=neo4j+s://qdrant.databases.neo4j.io
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=your_password

# Qdrant
QDRANT_URL=https://your-qdrant-url
QDRANT_KEY=your_qdrant_key

# Chat LLM (OpenRouter)
OPENROUTER_URL=https://openrouter.ai/api/v1
OPENROUTER_API_KEY=your_openrouter_key

# Embeddings LLM
EMBEDDING_URL=https://api.openai.com/v1
EMBEDDING_API_KEY=your_embedding_key

Notes:

• The embeddings model is text-embedding-3-small (dimension 1536) to match Qdrant collection size.
• NEO4J_URI supports neo4j+s:// for TLS.

Install & Run

bun install
bun run typecheck
bun start

The script will:

1. Ensure Qdrant collection exists (1536, Cosine)
2. Extract a graph via LLM (JSON)
3. Ingest nodes/relationships into Neo4j
4. Embed paragraphs and upsert into Qdrant
5. Retrieve top-K related entities and subgraph from Neo4j
6. Generate a final answer with GraphRAG

Troubleshooting

• Ensure all envs are set and valid. The app will fail fast if any are missing/invalid.
• Confirm your Qdrant URL/API key and collection dimensions align with the embedding size (1536).
• In Neo4j Browser, you can verify:

MATCH (n:Entity) RETURN count(n);
MATCH (:Entity)-[r]-() RETURN count(r);
MATCH (n:Entity) RETURN n.id, n.name LIMIT 25;