Skip to Content
ResearchTechnologyRAG with pgvector

RAG with pgvector

Summary

We use PostgreSQL with the pgvector extension for vector similarity search. Embeddings from OpenAI text-embedding-3-small (1536 dimensions). HNSW index for fast approximate nearest neighbor search.

Chunking Strategy

  • L1 — API metadata (name, version, description, team, tags)
  • L2 — Endpoint chunks (HTTP method, path, summary, params, responses)
  • L3/L4 (v2) — Schema chunks, example chunks

Alternatives Considered

  • Managed vector stores (Pinecone, Weaviate) — Higher cost ($40–100/month); we chose pgvector for ~$5–15/month with existing PostgreSQL
  • Ollama — Used for local dev; OpenAI in production for consistency

Implications

  • RAG data lives in isolated platform_ai schema
  • platform-backend proxies to platform-ai; never owns RAG entities
Last updated on
Technology Research