Skip to content

Vol. VI · Context & Memory · VI.4

Retrieval and Document Agents

LlamaIndex's claim: the hard problem is often the corpus, not the orchestrator. Agentic RAG is a router over query engines, not a smarter chatbot.

6 min read

Doctrine

  • If the knowledge lives in documents, start with indexes and evals on retrieval, not with a crew.
  • Agentic RAG: the model chooses which retriever, rewrite, or subquery — with traces.
  • Firecrawl and similar tools feed the index. They are not a substitute for one.

When LlamaIndex, when LangGraph

Corpus-heavy Q&A: LlamaIndex (or Dify for visual). Multi-step operations with tools and state: LangGraph. Many products need both: retrieve, then act. Keep the retrieval eval separate from the agent eval or you will not know which half broke.

Anti-patterns

  • Dumping crawl results into the prompt and calling it RAG.
  • No citation path from answer to chunk.

Related