Vector extensions

AI pipelines integrates with two high-performance PostgreSQL extensions for vector search: VectorChord for dense semantic (embedding-based) search, and VectorChord-BM25 for sparse keyword (BM25) search. Both are configured as part of the pipeline’s embedding (knowledge base) step and complement the standard pgvector capabilities built into EDB Postgres AI.

Extension	Search type	Best for
Enable VectorChord (optional)	Dense vector (ANN)	Semantic similarity at billion-scale; lower latency than HNSW/IVFFlat at high dimensions.
VectorChord-BM25	Sparse keyword (BM25)	Exact keyword relevance ranking; complements dense search in hybrid retrieval.

Together, these extensions form the foundation for hybrid search pipelines — combining semantic meaning (VectorChord) with keyword precision (VectorChord-BM25) to produce higher-quality retrieval results.

For full extension documentation including installation and release notes, see:

• Enable VectorChord (optional)

• VectorChord-BM25

• VectorChord-BM25
  • When to use VectorChord-BM25
  • How it works
  • Using VectorChord-BM25 alongside pipelines
    • Step 1: Set up the tokenizer
    • Step 2: Create the BM25 table
    • Step 3: Populate and index
    • Step 4: Query with BM25
  • Combining BM25 with a dense knowledge base
  • Relationship to hybrid search
  • Further reading
• VectorChord
  • When to use VectorChord
  • Index types
    • chord_hnsw
    • chord_vchordq
  • Using VectorChord with pipelines
    • Configuring a pipeline with chord_hnsw
    • Configuring a pipeline with chord_vchordq
    • Multi-step pipeline with chunking and VectorChord
  • Choosing between index types
  • Further reading