Agentic Graph RAG

Plain vector RAG answers questions about passages but loses the relationships between things: who founded what, which organizations collaborate, how two entities connect. Graph-only retrieval needs a schema and clean entities, and the same real-world entity arrives under many spellings across documents. Deciding which retrieval strategy fits a given question is itself a judgement, and users need to see why an answer was produced, not just the answer.

The platform is an API gateway in front of four domain services (ingestion, ontology, graph construction, retrieval), all FastAPI, with a shared Python library of models, adapters, and utilities volume-mounted into each container. Services are stateless; documents, schemas, and job state live in Redis-backed stores.

Ingestion parses five document formats, chunks with a choice of three strategies, embeds chunks, and stores vectors in Qdrant. It can optionally chain into ontology extraction and graph construction as a background pipeline. Graph construction runs as an async job: entities and relationships are extracted by a model, resolved in two stages (fuzzy name match, then embedding similarity) against both the current batch and what is already in the graph, and merged into Neo4j with constraints that prevent duplicate nodes.

Retrieval is a LangGraph state machine: understand the query, route to a strategy, run vector, graph, or hybrid search, then synthesize. Every step appends to a reasoning trail returned with the answer. Tools include vector search, graph traversal, entity search, hybrid search, shortest path, and natural-language to Cypher or Gremlin generation.

Graph access goes through an abstract adapter so the backend can be swapped by configuration; Neo4j is fully implemented and an AWS Neptune adapter is interface-ready. Prometheus metrics are exposed from the gateway, and a D3-based frontend provides upload, ontology management, graph visualization, and a query interface.