Part of Build with Claude Using Code and Guides
Engineering Blueprint
Safety checkedBuild Retrieval Augmented Generation Systems
Techniques for enhancing Claude with domain-specific knowledge using RAG, including contextual embeddings and re-ranking.
What problem does this solve?
Demonstrates how to augment Claude's responses with external knowledge, removing hallucinations for fact-dependent tasks. The guide covers RAG system architecture, performance optimization techniques like summary indexing and re-ranking, and building evaluation suites to measure precision, recall, and accuracy.
How does it work?
- Prepare your knowledge base and chunk documents. 2. Choose a retrieval strategy (semantic search, BM25, or both). 3. Optionally apply contextual embeddings to add context to chunks before retrieval. 4. Integrate retrieval into Claude prompts and measure accuracy with evaluation datasets. Output: a production-grade RAG pipeline with measured performance metrics.
Included Skills
No skills in this group
What's required to run this?
Includes sample data from Anthropic documentation. Evaluation uses Promptfoo config files for automated testing.
Key Features
Summary indexing
Improve retrieval precision by indexing summaries instead of raw chunks
Contextual embeddings
Solve context-loss in chunked retrieval by embedding chunks with surrounding context
Re-ranking strategies
Post-filter retrieved documents to boost relevant results to the top
Evaluation suite
Measure RAG accuracy with Promptfoo datasets covering single-hop and multi-hop queries
About This Blueprint
- Industry
- Technology
More Blueprints to explore
ICP-Filtered Outreach Lists from G2 Buyer Intent
Stop wasting rep time on weak intent signals and mismatched accounts. The workflow ranks ICP-fit buyers by readiness, finds the right contact, and supplies tailored outreach angles for the strongest opportunities.
Check Point Next Generation Firewalls (NGFWs)Ishani D.
Account Manager at G2