[Feature]: Replace PGVector with VectorChord or ParadeDB for HybridSearch using BM25

[qdrddr]

Use Case

I want to benefit from Hybrid Search, and I want to benefit from SOTA models with dimensions higher than 2000 using PostgreSQL DB.

Problem Statement

Semantic search is inferior to Hybrid (Vector + FTS BM25)

How This Feature Would Help

Improve accuracy

Proposed Solution

implement HybridSearch
https://docs.vectorchord.ai/vectorchord/use-case/hybrid-search.html
https://docs.vectorchord.ai/vectorchord/getting-started/vectorchord-suite.html#vectorchord-suite

Alternatives Considered

No response

Priority

Important - affects my workflow

Additional Context

No response

Checklist

• I would be willing to contribute this feature

[nicoloboschi]

we already do hybrid search
see #183 (comment)

[qdrddr]

@nicoloboschi, I was talking about HybridSearch with PostgreSQL using the VectorChord or ParadeDB extensions.
Also, PGvector has limitations on dimensionality that cannot be larger than 2000. Which makes it impossible to use with SOTA Embedding models such as https://openrouter.ai/qwen/qwen3-embedding-8b (4096 dimensions), while the VectorChord extension does not have this limitation and supports up to 65535 dimensions. VectorChord does not have dimensionality limitation, is much faster than PGVector and can improve recall latency, is PGVector-compatible, and has BM25 all in one PostgreSQL DB.

Would you please consider reopening this ticket?

[qdrddr]

I just tested replacing the DB image in docker-compose.yaml

image: tensorchord/vchord-suite:pg${HINDSIGHT_DB_VERSION:-18-latest}

Instead of pgvector - It works for me with text-embedding-3-small (1536 dimensions)

Though if you want to allow using models with higher than 2k dimensionality and be able to use HybridSearch, you'll need to update the code.

When I try text-embedding-3-large with 3072 dimensions
I see this message: column cannot have more than 2000 dimensions for hnsw index.

2026-02-11 18:50:45,111 - INFO - hindsight_api.migrations - Embedding dimension mismatch: database has 384, model requires 3072
2026-02-11 18:50:45,112 - INFO - hindsight_api.migrations - Altering embedding column dimension from 384 to 3072
2026-02-11 18:50:45,121 - WARNING - hindsight_api.engine.memory_engine - Failed to ensure embedding dimension for schema public: (psycopg2.errors.ProgramLimitExceeded) column cannot have more than 2000 dimensions for hnsw index

[SQL: 
                CREATE INDEX IF NOT EXISTS idx_memory_units_embedding_hnsw
                ON public.memory_units
                USING hnsw (embedding vector_cosine_ops)
                WITH (m = 16, ef_construction = 64)
            ]
(Background on this error at: https://sqlalche.me/e/20/e3q8)

To support dimensionality more than 2000 we need to replace the hnsw index with vchordrq
https://github.com/vectorize-io/hindsight/blob/v0.4.10/hindsight-api/hindsight_api/migrations.py#L422-L442
https://github.com/vectorize-io/hindsight/blob/v0.4.10/hindsight-api/hindsight_api/models.py#L166
https://github.com/tensorchord/VectorChord/issues/273?utm_source=copilot.com

https://docs.vectorchord.ai/faqs/comparison-pgvector.html#vector-dimensions

[qdrddr]

#350