Which LLM models are recommended for HyDE generation?

It is best to use small, fast, and cost-effective models like gpt-4o-mini or Claude 3 Haiku, as the goal is to generate technical context rather than high-level reasoning.

Does using HyDE increase search latency?

Yes, because it requires an initial LLM call. To mitigate this, the skill uses fast models like gpt-4o-mini, implements aggressive caching, and includes a fallback mechanism that reverts to standard embedding if the LLM takes too long.

What is HyDE and how does it improve retrieval?

HyDE stands for Hypothetical Document Embeddings. It works by using an LLM to generate a 'fake' answer to a query, then embedding that answer instead of the query. This improves retrieval because the hypothetical answer is more likely to share vocabulary with the actual documents.

When should I avoid using HyDE?

HyDE is less effective for exact term searches (where keyword search is better) or code snippet searches. It is most powerful for conceptual, natural language, or abstract questions.

HyDE Semantic Retrieval

Name: HyDE Semantic Retrieval
Author: yonatangross

byyonatangross

•

数据科学与机器学习

Enhances RAG retrieval accuracy by generating hypothetical answer documents to bridge vocabulary gaps in semantic search.

The HyDE (Hypothetical Document Embeddings) skill implements an advanced Retrieval-Augmented Generation pattern designed to solve the common 'vocabulary mismatch' problem in vector search. By transforming abstract user queries into detailed hypothetical answers before embedding them, it ensures significantly higher similarity scores when matching against technical documentation that may use specialized terminology. This production-ready implementation includes essential features for real-world deployment, such as aggressive LRU caching, multi-concept batching, and configurable fallback mechanisms to ensure system reliability even when LLM generation times out.

主要功能

0129 GitHub stars

02Hypothetical Document Generation for context-rich embeddings

03Graceful timeout fallbacks to direct query embedding

04Vocabulary bridging between user intent and technical docs

05Production-grade LRU caching to minimize LLM costs

06Multi-concept batch processing for complex queries

使用场景

01Improving search relevance for abstract or conceptual user queries

02Building robust RAG pipelines for documentation with heavy technical jargon

03Optimizing semantic retrieval performance in AI-powered support bots

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add yonatangross/skillforge-claude-plugin hyde-retrieval

For use in Claude.ai and ChatGPT

Download Skill