Class HydeRetriever

Whether HyDE is enabled.

Generate a hypothetical answer for a query.

Generate multiple hypothetical documents from different perspectives.

Each hypothesis approaches the query from a different angle, improving recall by covering more of the semantic space. Uses chain-of-thought prompting to ensure diverse, high-quality hypotheses.

The system prompt asks the LLM to generate N diverse hypotheses:

• Hypothesis 1: Technical/formal perspective
• Hypothesis 2: Practical/example perspective
• Hypothesis 3: Overview/summary perspective
• (Additional hypotheses explore further angles)

Throws

If the LLM call fails.

Example

const { hypotheses, latencyMs } = await retriever.generateMultipleHypotheses(
  'How does BM25 scoring work?',
  3,
);
// hypotheses[0]: Technical explanation with formulas
// hypotheses[1]: Practical example with code
// hypotheses[2]: High-level conceptual overview
Copy

Multi-hypothesis retrieval: generates N diverse hypotheses, searches with each, and merges results by deduplication (keeping the highest score per document).

This dramatically improves recall compared to single-hypothesis HyDE because one bad hypothesis doesn't ruin everything — other hypotheses can still find relevant documents from different angles.

Pipeline:

1. Generate N hypotheses via generateMultipleHypotheses
2. Embed each hypothesis
3. Search the vector store with each embedding
4. Union all results, deduplicate by document ID, keep highest score

Example

const result = await retriever.retrieveMulti({
  query: 'How does BM25 work?',
  vectorStore: myStore,
  collectionName: 'knowledge-base',
  hypothesisCount: 3,
});
console.log(`Found ${result.queryResult.documents.length} unique docs from ${result.hypothesisCount} hypotheses`);
Copy

Embed the hypothesis and search the vector store. Uses adaptive thresholding: starts at initialThreshold, steps down until results are found or minThreshold is reached.

Convenience: retrieve and format as augmented context string.