Advanced Patterns: Agentic, Graph and Multimodal RAG

When one retrieval is not enough

Some questions need more than a single search. Agentic RAG lets the model decide to search several times, rewrite its query, choose between tools (documents, SQL, web) and stop when it has enough evidence; it is more capable but slower, costlier and harder to test. Graph RAG builds a knowledge graph of entities and relations to answer questions that span many documents ("which suppliers are linked to project X?"). Multimodal RAG retrieves images, tables and charts, for example through captions or multimodal embeddings. Structured data questions are often best answered by generating SQL rather than retrieving text. Adopt these only after the basic pipeline is measured and its limits are clear.

Quick check: When is agentic RAG most useful?

• To avoid evaluation
• For one-line FAQ lookups
• When a question needs several searches or tools to answer
• To remove the LLM