Hypothetical Document Embeddings (HyDE)

Standard dense retrieval embeds a short user query and searches for similar document chunks. This creates a vocabulary mismatch problem: queries use terse, keyword-heavy language, while documents use expansive, descriptive prose. Embedding both in the same space often fails to capture their semantic equivalence. HyDE solves this by first transforming the query into the document space.

• Problem: Query 'What is a cat?' is short; relevant documents discuss 'feline domestication history'.
• Result: Raw query embeddings often miss nuanced, long-form answers.

HyDE instructs a generative language model (like GPT-3.5 or Llama 2) to create a hypothetical document that answers the user's query. The prompt is typically: 'Write a passage that answers the question: [QUERY]'. The model generates a synthetic text that captures the expected style, tone, and factual content of a real answer, even if the generated facts are hallucinated.

• Key Insight: The synthetic document does not need to be factually correct.
• Purpose: It only needs to 'look like' a real document to produce a useful embedding vector.

The generated hypothetical document is then passed through a dense embedding model (e.g., Instructor-XL, Contriever) to produce a vector. This vector is used to query the vector database instead of the raw query embedding. Because the synthetic document mimics the structure of real indexed documents, its embedding lands in a dense region of the vector space populated by relevant, real documents.

• Process: Query → LLM → Hypothetical Doc → Embedding Model → Vector → Vector DB Search
• Contrast: Standard retrieval skips the LLM and Hypothetical Doc steps.

HyDE is fundamentally a zero-shot technique. It requires no fine-tuning, no labeled training data, and no model retraining. It works by chaining two pre-existing, off-the-shelf models: a generative LLM and a contrastively trained embedding model. Furthermore, because the generative model can produce text in any language it supports, HyDE is inherently language agnostic.

• No Training: Works immediately with any generative model and any embedding model.
• Multilingual: Generate the hypothetical document in the same language as the target corpus for cross-lingual retrieval.

HyDE consistently outperforms standard dense retrieval (using the raw query embedding) on benchmarks like TREC DL and BEIR. It is particularly effective for complex, fact-seeking queries. The primary trade-off is latency and cost: a full LLM generation step is added before the vector search.

• Latency: Adds 1-3 seconds for the LLM generation call.
• Cost: Incurs token generation costs for every query.
• Mitigation: Use smaller, faster instruction-tuned models (e.g., Llama-3-8B) to minimize overhead.

HyDE is closely related to Query2Doc, a contemporaneous technique. Both generate synthetic text to augment retrieval. The distinction is subtle: Query2Doc appends the generated pseudo-document to the original query before embedding, while classic HyDE replaces the query entirely with the hypothetical document's embedding.

• HyDE: Embed only the hypothetical document.
• Query2Doc: Concatenate query + generated doc, then embed.
• Modern Practice: Many implementations now use a hybrid, embedding both and fusing results via Reciprocal Rank Fusion (RRF).

The broader category of techniques that HyDE belongs to, where a user's original query is transformed or augmented to improve retrieval recall. HyDE is a specific, generative form of this.

• Paraphrasing: Generating alternative phrasings of the same query.
• Synonym Injection: Adding related terms to capture lexical variations.
• HyDE: Generating a synthetic answer document to bridge the semantic gap.

A foundational bi-encoder architecture that uses separate models to embed queries and documents into a shared vector space. HyDE often operates on top of a DPR index.

• Bi-Encoder: Independent encoding of query and document.
• Dot-Product Similarity: Measures relevance between the two vectors.
• Limitation: Struggles with asymmetric queries where the user asks a question but the document contains a statement.

The retrieval paradigm that HyDE enhances by moving beyond keyword matching to understand conceptual intent. HyDE converts a short, abstract query into a dense, concrete representation.

• Vector Similarity: Finds documents based on meaning, not just terms.
• Embedding Models: Transform text into high-dimensional vectors.
• Challenge: Short queries provide limited semantic signal, which HyDE directly addresses.

A two-stage retrieval refinement often paired with HyDE. HyDE performs the initial, high-recall retrieval, while a cross-encoder re-ranks the top candidates with high precision.

• Stage 1 (HyDE): Fast, approximate retrieval of ~100 candidates.
• Stage 2 (Cross-Encoder): Jointly processes the query and each candidate for an exact relevance score.
• Trade-off: Maximizes both recall and precision at the cost of added latency.

Embedding models that accept task-specific instructions, allowing a single model to handle asymmetric tasks like query-document matching. This is an alternative to HyDE that bakes the transformation into the model itself.

• Task Prompting: Prefix the query with an instruction like 'Represent this question for retrieval.'
• Unified Model: One model for both query and document encoding.
• Comparison: HyDE uses an external LLM, while this uses an internal model mechanism.

A strategy where a single document is indexed with multiple vectors representing different aspects, summaries, or hypothetical questions. This is the inverse of HyDE, where the document side is expanded.

• Query Generation: Generate likely questions a document answers and index them.
• Summary Indexing: Index a summary alongside the full text.
• Synergy: Can be combined with HyDE for a dual-expansion approach.