Skip to content
vectorstores Documentation

Vector Stores

Available Vector stores save embedding vectors of your ingested document chunks.

Search Modes

Section titled “Search Modes”

Most vector stores in this library support the following search modes:

  • Default: Standard vector similarity search using embeddings.
  • BM25: Keyword search using the BM25 algorithm.
  • Hybrid: A combination of vector similarity search and BM25 keyword search.

Scoring Formats

Section titled “Scoring Formats”

The score returned for each node depends on the search mode used:

  • Default (Vector): Scores are based on cosine similarity and range from 0.0 to 1.0, where 1.0 indicates a perfect semantic match.
  • BM25: Scores are calculated using the BM25 algorithm. These scores are unbounded (can be greater than 1.0) and represent keyword relevance. Higher scores indicate better matches.
  • Hybrid: Scores are calculated using Reciprocal Rank Fusion (RRF). RRF scores are typically very small (e.g., a rank 1 result has a score of $\approx 0.0164$ or $1.64%$). This method is robust as it only considers the rank position from each sub-search, not the raw similarity or BM25 scores.

You can specify the search mode in the query method of the vector store:

const result = await vectorStore.query({
  queryStr: "your keyword query",
  queryEmbedding: [0.1, 0.2, ...],
  mode: "hybrid",
  similarityTopK: 5,
  alpha: 0.5, // Importance of vector search vs keyword search (0.0 to 1.0)
});

Native vs Fallback Support

Section titled “Native vs Fallback Support”

Where possible, this library uses the native hybrid and BM25 implementations of the underlying vector store (e.g., Weaviate, ElasticSearch, MongoDB Atlas, PostgreSQL). If a vector store does not support these modes natively, a fallback implementation is used where possible.

Available Vector Stores

Section titled “Available Vector Stores”

Available Vector Stores are shown on the sidebar to the left. Additionally the following integrations exist without separate documentation:

Check the vectorstores Github for the most up to date overview of integrations.