Advanced10 min

Query Transformation

Techniques to improve retrieval by transforming user queries before search: HyDE, multi-query expansion, step-back prompting, and query decomposition.

Quick Reference

→HyDE: generate a hypothetical answer, embed that instead of the question — bridges the query-document gap
→Multi-query: expand one question into 3-5 variations to improve recall across different phrasings
→Step-back prompting: ask a more general question first to retrieve broader context
→Query decomposition: break a complex question into sub-queries, retrieve for each independently
→Query transformation addresses the fundamental vocabulary mismatch between how users ask and how documents are written

The Query-Document Vocabulary Mismatch

Users ask questions in natural language. Documents are written in technical prose. The gap between these two styles is the root cause of most retrieval failures. A user asks 'Why is my app slow?' but the relevant document says 'Performance optimization: reducing API response latency through connection pooling.' The embedding model may not bridge this gap because the surface-level vocabulary is completely different. Query transformation techniques reshape the user's question to better match document vocabulary.

Asymmetric retrieval

Queries are short (5-15 words) and in question form. Documents are long (100-500 words) and in declarative form. Embedding models are trained on both, but the embedding space is inherently asymmetric. Query transformation makes the query look more like a document, or generates multiple query variants to cast a wider net.

▸Short query, long document: user asks 'auth error' but the doc explains OAuth2 token refresh flows in detail
▸Question vs statement: user asks 'How does X work?' but the doc states 'X operates by...'
▸Colloquial vs technical: user says 'make it faster' but docs discuss 'latency optimization' and 'throughput engineering'
▸Incomplete context: user asks 'What about pricing?' referencing a product discussed 3 messages ago

HyDE: Hypothetical Document Embeddings

HyDE (Hypothetical Document Embeddings) is one of the most powerful query transformation techniques. Instead of embedding the user's question, you ask an LLM to generate a hypothetical answer — even if it's wrong — and embed that instead. The hypothetical answer is written in document style and uses document vocabulary, making it much closer in the embedding space to the actual relevant documents.

Multi-Query Expansion

Multi-query expansion uses an LLM to generate 3-5 variations of the original question, then retrieves documents for each variation and deduplicates the results. This dramatically improves recall because different phrasings match different documents. A single question might miss relevant docs, but five variations cast a much wider net.

Sign in to read this article

This is a premium article. Sign in with your Google account to continue.