Graph-guided retrieval engine for multi-hop reasoning over personal knowledge.
Kairos keeps semantic text recall and structural graph traversal as separate primitives. The current system stores source text and vectors in PostgreSQL/pgvector, writes graph structure to Neo4j, and uses GDS PPR to expand dense passage seeds.
ONNX Runtime and DJL keep embeddings on the JVM for chunks, concepts, and queries.
Gemini Flash extracts triples during ingestion so the graph grows with the source material.
The retrieval layer reads dense candidates and graph structure without merging those concerns into one database model.
Dense retrieval and graph traversal run together because cosine similarity alone misses multi-hop questions.
pgvector for dense search and Neo4j for graph traversal
all-MiniLM-L6-v2 vectors shared by chunks, concepts, and queries
Personalized PageRank seeded by dense passage candidates and hydrated through PostgreSQL
Kairos is a protected Spring Boot monolith around a context engine core. AI indexing feeds PostgreSQL/pgvector and Neo4j separately, while online retrieval combines dense anchors with graph expansion.
Dense passage recall gives useful anchors, but current retrieval still needs graph expansion for multi-hop context.
Kairos separates semantic storage and graph structure, then combines both signals at retrieval time.
The main architecture is a Spring Boot monolith with identity, protected source APIs, a context engine core, AI indexing adapters, and two persistence surfaces. Ingestion feeds both stores; retrieval combines pgvector anchors with Neo4j GDS expansion before PostgreSQL hydrates the final chunk payloads.
Dual-store retrieval over one generalized database
One generalized store forces either graph traversal or dense similarity search into a weaker execution model. Kairos keeps pgvector for vector search and Neo4j for graph traversal because HippoRAG 2 needs both primitives natively.
HippoRAG 2 over cosine-only retrieval
Cosine-only retrieval misses questions that require a chain of concepts spread across different passages. HippoRAG 2 seeds Personalized PageRank from semantic anchors, so relevance can move through the graph before the final ranker chooses context.
ONNX Runtime on the JVM over a Python sidecar
A Python sidecar adds another runtime, another deployment unit, and another failure boundary to every embedding call. ONNX Runtime plus DJL tokenizers keeps chunk, node, and query embeddings inside the JVM with one semantic space.
LLM-powered OpenIE over manual relation entry
Manual relation entry does not scale with passive knowledge accumulation and leaves the graph stale unless the user curates it constantly. Gemini Flash extracts triples during ingestion so the graph grows with the content itself.