VelesDB + LangChain Integration

Difficulty: Intermediate | Showcases: Hybrid search (dense + sparse), RRF fusion, LangChain VectorStore interface

Example showing VelesDB as a hybrid dense+sparse vector store for LangChain applications, using the published langchain-velesdb connector — adopting VelesDB is a single dependency change.

Why VelesDB for LangChain?

Most RAG pipelines need both semantic search (dense vectors) and keyword search (sparse/BM25). This typically requires running two separate systems (e.g., Pinecone + Elasticsearch) and writing glue code to fuse results.

VelesDB handles dense + sparse + fusion in a single engine:

• Dense search via HNSW with AVX2/AVX-512 SIMD acceleration
• Sparse search via inverted index with MaxScore optimization
• Hybrid fusion via Reciprocal Rank Fusion (RRF) built in
• Sub-millisecond latency, local-first, no cloud dependency

Prerequisites

pip install langchain-velesdb

Usage

python hybrid_search.py

The example uses deterministic synthetic embeddings so it runs without an embedding model or API key. In production, replace DemoEmbeddings with a real embedding model (OpenAI, Sentence-Transformers, Cohere, etc.).

What the Example Shows

1. langchain_velesdb.VelesDBVectorStore — the published LangChain connector
2. add_texts with metadata and sparse vectors in one call
3. Dense-only search via similarity_search_with_score
4. Hybrid dense+sparse search by passing sparse_vector= (RRF fusion)
5. Hybrid vector+BM25 search via hybrid_search

Expected Output

Inserted 20 documents

=== Dense-Only Search ===
  [0.xxxx] LangChain provides abstractions for building LLM applications
  ...

=== Hybrid Search (Dense + Sparse) ===
  [0.xxxx] Sub-millisecond latency is critical for real-time AI applications
  ...

=== Hybrid Search (Vector + BM25) ===
  [0.xxxx] Hybrid search fuses dense and sparse signals for better recall
  ...

Adapting for Production

from langchain_openai import OpenAIEmbeddings
from langchain_velesdb import VelesDBVectorStore

store = VelesDBVectorStore(
    embedding=OpenAIEmbeddings(),
    path="./data",
    collection_name="my_docs",
)

# add_texts auto-generates embeddings via the embedding model
store.add_texts(["Document text here..."])