RAG Patterns

Use this skill to design, review, or optimize Retrieval-Augmented Generation systems, document Q&A, knowledge retrieval, agent memory, multimodal retrieval, and LLM context usage.

Invocation Triggers

• User asks for RAG architecture, chunking, embeddings, retrieval, reranking, vector databases, evaluation, or memory design.
• User is building document Q&A, agent memory, multimodal document processing, or retrieval tuning.
• User wants to reduce context usage through retrieval, caching, or memory patterns.

Core Workflow

1. Clarify data shape, query types, freshness needs, latency/cost limits, and quality bar.
2. Choose architecture tier: naive, advanced, agentic, or multimodal.
3. Select chunking strategy and metadata model.
4. Select embeddings and storage.
5. Design retrieval: vector, keyword, hybrid, query expansion, reranking, or routing.
6. Add memory, cache, and multimodal handling only when the use case needs them.
7. Define evaluation metrics and feedback loop.

Architecture Tiers

Tier	Shape	Use When
Naive RAG	Query -> embed -> vector search -> top-k -> LLM	Prototype or small corpus
Advanced RAG	Query rewrite -> hybrid search -> rerank -> LLM	Production text retrieval
Agentic RAG	Router chooses search, SQL, APIs, tools	Complex questions and multi-source tasks
Multimodal RAG	Parse text, tables, images, diagrams	PDFs, slides, charts, mixed media

Design Rules

• Use recursive or document-aware chunking for structured docs.
• Use parent-child retrieval when small chunks need larger context at generation time.
• Use hybrid search when keyword precision and semantic recall both matter.
• Use reranking when top-k quality is more important than latency.
• Match embedding model to language, context length, cost, and storage limits.
• Prefer evaluation with faithfulness, context relevance, answer relevance, recall, and precision.
• Avoid adding agents, multimodal parsing, or cache layers unless the failure mode justifies the complexity.

Safety and Quality Constraints

• Do not claim benchmark results without measured evidence.
• Do not store secrets, credentials, personal data, private files, or sensitive documents in vector stores or memory without explicit approval and a retention plan.
• Call out data freshness, access control, source attribution, and hallucination risk.
• Separate public examples from private corpus details.
• When using external APIs or managed vector databases, keep credentials in environment variables and avoid logging query payloads with sensitive content.

Output Expectations

• Architecture recommendation with tier, retrieval strategy, storage choice, and tradeoffs.
• Chunking and metadata plan.
• Embedding and vector database recommendation.
• Retrieval pipeline or pseudocode when useful.
• Evaluation plan and metrics.
• Known risks, follow-up tests, and rollout steps.

Reference Map

Need	Read
Architecture tiers, chunking strategies, parent-child pattern	architecture-and-chunking.md
Embedding selection, hybrid search, query expansion, reranking	embeddings-and-retrieval.md
Multimodal RAG, agent memory, memory MCP, LLM caching	multimodal-memory-and-caching.md
Vector database selection, evaluation metrics, cross-references	vector-db-and-evaluation.md