Skip to content

Storage: PostgreSQL (pgvector + FTS) and Neo4j

  • PostgreSQL

    Single source for chunks, embeddings, and FTS.

  • pgvector

    High-dimensional vector similarity for dense retrieval.

  • Neo4j

    Entity/relationship graph and optional vector index on chunks.

Get started Configuration API

One Postgres, Many Corpora

Partition by repo_id/corpus_id in schema to keep corpus isolation at the data layer.

TSConfig

indexing.postgres_ts_config resolves the text search config (simple or a stemmer language) based on tokenizer settings.

Neo4j Isolation

neo4j_database_mode=per_corpus avoids expensive cross-corpus filters but requires Enterprise Edition.

PostgreSQL Client Responsibilities

  • Upsert chunk rows with metadata and content
  • Upsert embeddings for pgvector similarity
  • Maintain FTS (tsvector) index
  • Execute vector and sparse searches with corpus scope

Neo4j Client Responsibilities

  • Ensure database exists (per mode)
  • Ensure vector index on Chunk nodes when enabled
  • Upsert entities and relationships
  • Expand from seed hits to related chunks
flowchart LR
    CH["Chunks"] --> PG["PostgreSQL"]
    EMB["Embeddings"] --> PG
    ENT["Entities"] --> NEO["Neo4j"]
    REL["Relationships"] --> NEO

Configuration Hooks

Section Field(s) Meaning
indexing postgres_url DSN for Postgres
graph_storage neo4j_uri, neo4j_user, neo4j_password Neo4j connectivity
graph_storage neo4j_database_mode, neo4j_database_prefix DB isolation strategy
graph_indexing store_chunk_embeddings, chunk_vector_index_name Neo4j vector search on chunks 
# Resolve Neo4j database name for a corpus (1)!
from server.models.tribrid_config_model import GraphStorageConfig
print(GraphStorageConfig().resolve_database("dev_corpus"))

# Connectivity checks are via readiness
curl -sS http://localhost:8000/ready | jq .

// Client: no direct DB use — rely on API readiness
  1. Uses prefix + sanitized corpus id in per_corpus mode
Vector Similarity

Set graph_indexing.vector_similarity_function to cosine or euclidean based on your embedding model norms.