Skip to main content

Class: HybridSearcher

Defined in: packages/agentos/src/rag/search/HybridSearcher.ts:113

Hybrid dense+sparse searcher combining vector embeddings with BM25.

Uses Reciprocal Rank Fusion (RRF) to merge results from both retrieval systems, capturing both semantic similarity and exact keyword matches.

Examples​

const bm25 = new BM25Index();
bm25.addDocuments(documents);

const hybrid = new HybridSearcher(vectorStore, embeddingManager, bm25, {
  denseWeight: 0.7,
  sparseWeight: 0.3,
  fusionMethod: 'rrf',
});

const results = await hybrid.search(
  'error TS2304 type declarations',
  'my-collection',
  10,
);
const hybrid = new HybridSearcher(vectorStore, embeddingManager, bm25, {
  fusionMethod: 'weighted-sum',
  denseWeight: 0.6,
  sparseWeight: 0.4,
});

Constructors​

Constructor​

new HybridSearcher(vectorStore, embeddingManager, bm25Index, config?): HybridSearcher

Defined in: packages/agentos/src/rag/search/HybridSearcher.ts:142

Creates a new HybridSearcher.

Parameters​

vectorStore​

IVectorStore

Dense vector store for semantic search.

embeddingManager​

IEmbeddingManager

Manager for generating query embeddings.

bm25Index​

BM25Index

BM25 sparse keyword index.

config?​

HybridSearcherConfig

Optional configuration overrides.

Returns​

HybridSearcher

Example​

const searcher = new HybridSearcher(store, embeddings, bm25, {
  denseWeight: 0.7,
  sparseWeight: 0.3,
});

Methods​

search(query, collectionName, topK?, queryOptions?): Promise<HybridResult[]>

Defined in: packages/agentos/src/rag/search/HybridSearcher.ts:184

Searches both dense and sparse indexes, then fuses results.

Pipeline:

  1. Generate query embedding via the embedding manager
  2. Query the dense vector store for semantically similar documents
  3. Query the BM25 sparse index for keyword-matching documents
  4. Fuse both result sets using the configured fusion method (RRF by default)
  5. Return the top K results sorted by fused score

Parameters​

query​

string

The search query text.

collectionName​

string

Vector store collection to search.

topK?​

number = 10

Maximum number of results to return.

queryOptions?​

Partial<QueryOptions>

Additional options for the vector store query.

Returns​

Promise<HybridResult[]>

Fused results sorted by relevance.

Throws​

If embedding generation fails.

Example​

const results = await hybrid.search('error TS2304', 'knowledge-base', 5);
for (const r of results) {
  console.log(`${r.id}: fused=${r.score.toFixed(4)} dense=${r.denseRank} sparse=${r.sparseRank}`);
}