What is Contextual Retrieval?

It is a technique that prepends a short, 1-2 sentence summary of a full document to each individual chunk before embedding, ensuring the chunk retains its situational meaning during the retrieval process.

Why does this skill recommend hybrid search?

Hybrid search combines BM25 (keyword matching) and vector similarity (semantic matching). Anthropic's research shows this combination, when paired with contextualized chunks, provides the lowest retrieval failure rate.

When should I skip using contextual retrieval?

You may skip it if your chunks are already self-contained (like Q&A pairs), if you require extremely low-latency indexing for streaming data, or if you have a massive volume of small, unrelated documents where indexing cost is a primary concern.

How does this skill help reduce API costs?

The skill implements Anthropic's prompt caching features, which allow the main document to be cached once and reused across hundreds of chunks, reducing context generation costs by approximately 90%.

Contextual Retrieval for RAG

Name: Contextual Retrieval for RAG
Author: yonatangross

byyonatangross

•

데이터 과학 및 ML

Implements Anthropic's contextual retrieval technique to improve RAG performance by prepending situational metadata to document chunks.

This skill provides standardized implementation patterns for Anthropic's Contextual Retrieval method, a technique designed to significantly reduce retrieval failures in RAG systems. By prepending brief, LLM-generated context to each document chunk before embedding, it ensures that individual pieces of information retain their document-level meaning—such as dates, entities, or section headers—that are often lost during traditional chunking. The skill includes production-ready code for prompt-cached context generation, hybrid search combining BM25 with vector similarity, and optimized indexing pipelines, enabling developers to build AI applications with up to 67% fewer retrieval errors.

주요 기능

0169 GitHub stars

02Standardized retrieval pipeline architectures for high-precision AI applications

03Automated context generation for document chunks using Claude Sonnet

04Hybrid search integration combining BM25 keyword matching with vector embeddings

05Efficient prompt caching implementation to reduce indexing costs by up to 90%

06Scalable parallel processing patterns for high-volume document indexing

사용 사례

01Optimizing technical knowledge bases where information is buried in dense, multi-page manuals

02Improving accuracy in financial or legal RAG applications where document metadata is critical

03Building customer support bots that must distinguish between specific product versions or dates

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add yonatangross/orchestkit contextual-retrieval

For use in Claude.ai and ChatGPT

Download Skill