I built this because I kept hitting a wall with standard RAG implementations. My agents could retrieve text chunks based on similarity, but they failed completely at understanding relationships between concepts that weren't explicitly stated in the same paragraph.
I didn't want to maintain a separate Neo4j instance, so I built a hybrid solution on top of Postgres (using pgvector).
The Architecture: It separates ingestion from processing. I call it "Sleep Cycles".
Ingest: Data is pushed and vectorized immediately for fast retrieval.
Consolidate: An async worker (Redis/BullMQ) processes the new data to extract entities and build a graph structure, linking new info to existing nodes.
This allows the system to answer multi-hop questions that pure vector search misses.
I just released a GitHub Action to automate syncing repository documentation, as manual context loading was a friction point.
I didn't want to maintain a separate Neo4j instance, so I built a hybrid solution on top of Postgres (using pgvector).
The Architecture: It separates ingestion from processing. I call it "Sleep Cycles".
Ingest: Data is pushed and vectorized immediately for fast retrieval.
Consolidate: An async worker (Redis/BullMQ) processes the new data to extract entities and build a graph structure, linking new info to existing nodes.
This allows the system to answer multi-hop questions that pure vector search misses.
I just released a GitHub Action to automate syncing repository documentation, as manual context loading was a friction point.