Retrieval-Augmented Self-Recall — Part 2: Hybrid RAG on Nothing but Postgres

Short summary

Part 2 of the RE-call series argues that Postgres with pgvector is the right choice for agent memory, not a dedicated vector database, because agent memory is relational, transactional, and small-scale. The retrieval pipeline runs four stages: dense cosine search, full-text search, Reciprocal Rank Fusion (k=60) to combine them, and optional cross-encoder re-ranking. The key insight is that dense and sparse retrieval fail in opposite directions, so fusing them yields both concept-matching and keyword precision.

• •Postgres + pgvector replaces a dedicated vector DB for agent memory — relational metadata, transactional consistency, no sync layer
• •Four-stage pipeline: dense search, full-text search, RRF fusion (k=60), optional cross-encoder re-ranking
• •Dense and sparse retrieval fail in opposite directions; combining them covers both concept-matching and exact-token precision

Generated with AI, which can make mistakes.