↓ Skip to main content
Yonk-Labs
Yonk-Labs
  1. Projects/

pg-raggraph

·2 mins
Python Postgresql Graphrag Pgvector Rag Ai Recursive-Cte
Table of Contents

pg-raggraph is a GraphRAG library that runs entirely in PostgreSQL. pgvector handles the vector side, adjacency tables and recursive CTEs handle the graph side, and tsvector provides BM25 keyword search — all in one ACID-compliant database. No graph extension, no second system to back up, no data sync between two stores.

Why pg-raggraph?
#

  • Most GraphRAG workloads don’t need a graph database — vector similarity, 1-3 hop traversal, and a result merge are the entire hot path. Recursive CTEs over an adjacency table do the traversal in 30-100ms, which is the hop budget GraphRAG actually uses.
  • One database, one backup story — if your data already lives in Postgres, pip install pg-raggraph and start ingesting against the same instance your application talks to. No ETL out and back.
  • Deploys where you actually run — works on a stock managed Postgres (e.g. AWS RDS). No need for a graph extension that your cloud provider doesn’t offer.
  • Honest about where it loses — benchmarked across six corpora (Postgres docs, NTSB reports, sales call notes, MuSiQue multi-hop QA, versioned Python docs, PubMed abstracts), with the cases where graph augmentation doesn’t pay off documented rather than hidden.

What it’s for
#

  • GraphRAG retrieval — entity/relationship extraction plus graph-augmented retrieval for multi-document reasoning, without standing up Neo4j or Apache AGE.
  • Agent memory — the relational/episodic, time-anchored kind of memory that needs a graph traversal, not just a vector match.
  • Multi-hop QA — walk from a query’s seed entities to related documents that share entities, in one SQL query plan.

Architecture
#

Documents → Chunks → Entities → Relationships  (+ provenance junction tables)
            pgvector (HNSW)   recursive CTEs   tsvector BM25 + pg_trgm (GIN)

Six tables — documents, chunks, entities, relationships, plus entity_chunks and relationship_chunks provenance links back to source. Indexes: HNSW for vectors, GIN for tsvector + fuzzy matching, btree for adjacency joins.

Links#