rag-and-agentslisted

# RAG & AI Agent Development — Production Reference ## The Decision Framework **Start naive → add complexity only when evaluation shows a quality ceiling.** Progression: 1. Naive RAG (embed-retrieve-stuff) 2. Hybrid search + semantic reranking 3. Parent-child chunking + better parsing 4. Contextual retrieval (Anthropic) 5. Advanced RAG (query transforms, multi-query, decomposition) 6. GraphRAG or agents — only when steps above have hit their ceiling **Each step adds cost. Advance only when a 50–200 QA golden set proves it.** --- ## RAG Fundamentals **Four problems RAG solves:** 1. Hallucination (grounds answers in retrieved documents) 2. Knowledge cutoff (retrieves current private data) 3. Private-data access (indexes your corpus) 4. Verifiable sourcing (enables citations) **The full pipeline:** ingestion → chunking → embedding → indexing → query processing → retrieval → reranking → context assembly → generation **RAG vs Fine-tuning vs Long-context:** - **RAG**: dynamic/proprietary knowledge needing citations; audit trail - **Fine-tuning**: changing behavior, format, tone, domain style - **Long-context stuffing**: single-document deep reasoning where the whole doc fits; no extra infra They combine — fine-tune for domain language, RAG for facts. **"Lost in the middle" (Liu et al., TACL 2024):** performance degrades significantly when relevant information is in the middle of long contexts, even for explicitly long-context models. Critical info should be first or last