rag_skill

Overview

This skill provides practical patterns and a developer-ready playbook for building Retrieval-Augmented Generation (RAG) systems that ground LLM responses with external knowledge sources. It covers vector store selection, embedding model choices, document processing, retrieval strategies, and end-to-end RAG pipelines for production and local development. Use it to reduce hallucinations and deliver factual, source-cited answers from proprietary or multi-source knowledge bases.

How this skill works

The skill guides you to ingest documents, split them into chunks, generate embeddings, and store those vectors with metadata in a vector database. It describes retrieval strategies (dense, sparse, hybrid), reranking and filtering techniques, and how to wire a retriever into an LLM prompt template with chat memory and grounding checks. It also includes evaluation metrics and optimization steps for precision, latency, and faithfulness.

When to use it

• Building Q&A systems over proprietary documents or manuals
• Creating chatbots that must provide current, factual information
• Implementing semantic search with natural language queries
• Reducing hallucinations by grounding responses in retrieved sources
• Combining multiple knowledge sources (web, DB, docs) for unified answers

Best practices

• Choose a vector store that matches scale and deployment needs (Pinecone/Milvus for production, Chroma/FAISS for local)
• Preprocess and clean documents, add useful metadata for filtering and context
• Use 500–1000 token chunks with 10–20% overlap, then test variations for your corpus
• Start retrieval with a higher k (10–20) and apply reranking or filtering to improve precision
• Cache embeddings, batch ingestions, and monitor query latency and resource usage

Example use cases

• Document Q&A assistant that answers policy or support questions with citations
• Conversational product assistant that keeps context across multi-turn sessions
• Research tool that merges results from document, database, and web retrievers and reranks them
• Knowledge management system that exposes domain-specific content via semantic search
• Compliance auditor that filters and retrieves documents by metadata and date ranges