rag-architect_skill

Overview

This skill codifies expertise for designing production-grade Retrieval-Augmented Generation (RAG) systems, vector databases, and knowledge-grounded AI applications. It focuses on pragmatic decisions: vector store selection, chunking and ingestion pipelines, hybrid retrieval, reranking, and continuous evaluation. Use it to make RAG systems accurate, efficient, and maintainable at scale.

How this skill works

I inspect system requirements (latency, accuracy, scale, multi-tenancy) and produce a concrete design: vector store choice, schema and indexing strategy, chunking rules, embedding selection, and retrieval pipeline including hybrid search and reranking. I enforce operational constraints such as idempotent ingestion, metadata enrichment, monitoring, and versioning of embeddings. Deliverables include architecture diagrams, trade-off analyses, chunking examples, and an evaluation plan with metrics and benchmarks.

When to use it

• Building chatbots, agent assistants, or Q&A systems that must ground answers in documents
• Choosing or migrating a vector database and defining indexing/sharding strategy
• Designing document ingestion, chunking, deduplication, and metadata enrichment pipelines
• Implementing semantic search, hybrid (vector + keyword) retrieval, and reranking
• Evaluating retrieval quality, debugging failures, and optimizing relevance/latency

Best practices

• Evaluate multiple embedding models on representative domain data and track embedding drift
• Always implement hybrid search (vector + keyword) and reranking for production relevance
• Enrich chunks with metadata (source, timestamp, section) and support filters for multi-tenant data
• Make ingestion idempotent with deduplication and preprocessing to avoid storing raw noisy documents
• Measure retrieval metrics (precision@k, recall@k, MRR, NDCG) and monitor latency and quality over time
• Version embeddings and plan model migrations; avoid coupling embedding model tightly to application code

Example use cases

• Designing an enterprise knowledge base with per-customer filters and low-latency search
• Selecting between Pinecone, Qdrant, Chroma, or pgvector with trade-offs for cost, scale, and operational control
• Building a content ingestion pipeline that chunks legal texts with semantic boundaries and overlap rules
• Implementing a hybrid search flow: query expansion -> vector + BM25 -> reranker -> LLM context assembly
• Creating an evaluation plan and dashboards to track precision@k, MRR, and retrieval regressions during embeddings updates