Your dense retriever performs poorly on terse user queries against prose heavy docs. How does HyDE (Hypothetical Document Embeddings) help, and via what mechanism?

Attempt it

Quick reference

TL;DR

HyDE asks an LLM to draft a fake answer, embeds that draft instead of the query, and retrieves against it, moving the search anchor into the document cluster.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Imagine searching a library by handing the librarian your question scribbled on a sticky note: 'best espresso machine?' The shelves are full of long, formal product reviews, and your terse note doesn't look anything like them, so the match is weak. Now imagine you first write a short fake review yourself ('This machine pulls rich shots with a stable temperature') even if the details are wrong. You hand THAT to the librarian instead. It looks like the real reviews on the shelf, so the librarian finds genuine matches fast. HyDE does exactly this: an LLM writes a hypothetical answer document, you embed that document instead of the bare query, and you search with it. The fake answer can contain errors; what matters is that it shares the shape and style of the real documents.

Key concepts

Retrieval-Augmented Generation Embeddings Vector databases Prompt engineering

Quick scan · 2–4 min

Everything important, quickly.

4 min: state the query-doc mismatch, explain HyDE's embed-the-fake-answer mechanism, note accuracy irrelevance, then cover cost/latency and when fine-tuning or hybrid wins instead.

Code example

python

# HyDE: embed a hypothetical answer, not the raw query
query = "best espresso machine for small kitchens?"

hyp_doc = llm.generate(
    f"Write a short passage that answers: {query}"
)  # may hallucinate; that's fine

hyp_vec = embed(hyp_doc)        # embed the synthetic DOC, not the query
chunks = index.search(hyp_vec, top_k=5)

# optional: average with the raw query vector to reduce variance
# anchor = 0.5 * hyp_vec + 0.5 * embed(query)

Side-by-side

Technique	What gets embedded	Retrains?	Main cost
HyDE	LLM-generated hypothetical answer doc	No	Extra LLM call per query
Raw dense retrieval	The user's query text	No	Cheapest, weakest on style gap
Multi-query expansion	Several LLM rewrites of the query	No	Multiple searches per query
Fine-tuned embeddings	Query + docs in aligned space	Yes	Training data + GPU time

Real-world examples

Real products, models, and research that use this idea.

LangChain ships a built in HyDE retriever that wraps any LLM and embedding model for drop in hypothetical-document retrieval.
LlamaIndex exposes HyDEQueryTransform as a query transformation you can chain before any vector index retrieval step.
Enterprise RAG teams use HyDE on zero-shot domains where fine-tuning a domain embedding model is too costly to justify.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy does the hypothetical document help even when it contains factual errors?

Sign in or go Pro to see the approach for this follow-up.

QHow would you reduce variance when a single hypothetical draft is unreliable?

Sign in or go Pro to see the approach for this follow-up.

QOn what kind of corpus would HyDE actively hurt retrieval quality?

Sign in or go Pro to see the approach for this follow-up.

QHow does HyDE compare to fine-tuning a domain embedding model for the same gap?

Sign in or go Pro to see the approach for this follow-up.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.