RAG Architect guides Claude through building retrieval-augmented generation in five steps: requirements, vector store design, chunking, the retrieval pipeline and evaluation. Each step has a checkpoint to pass before the next.
The skill includes Python examples for chunking with metadata, idempotent indexing into Qdrant with OpenAI embeddings, hybrid vector and BM25 search with tenant filters, Cohere reranking and RAGAS evaluation.
What You Get
- A five-step workflow with a validation checkpoint for each step
- Five implementation examples: chunking, embedding and indexing, hybrid search, reranking and retrieval evaluation
- Reference files for vector databases, embedding models, chunking strategies, retrieval optimization and RAG evaluation
- Rules for hybrid search, metadata filters, reranking, deterministic IDs, embedding versioning and measured retrieval metrics
- An output template with an architecture diagram, database trade-offs, chunking rationale, pipeline design and an evaluation plan
Common Use Cases
- Choosing between Pinecone, Weaviate, Chroma, pgvector and Qdrant
- Setting chunk size and overlap by testing on domain data
- Adding BM25 and reranking on top of vector search
- Building a multi-tenant retrieval index with metadata filters
- Measuring retrieval quality with precision, recall and MRR
What Makes It Unique
The skill treats retrieval quality as something to measure apart from answer quality. Checkpoints assert metadata and deduplication, evaluation targets are numeric, and it warns against defaults such as an unevaluated chunk size of 512. Embedding choices are versioned so a model migration can be planned.
Under The Hood
The skill is instructions and reference markdown that Claude loads by topic; it ships no scripts. Its examples call the OpenAI embeddings API, the Cohere rerank API with a key read from an environment variable, and a Qdrant server on localhost, and the indexing example recreates a collection, which deletes the existing one. Claude writes Python files in your project and may run an evaluation script if you have one.