Retrieval and Search Tools

Retrieval as a tool

Models do not know your private documents or yesterday's data. A search tool (keyword, vector or SQL) lets the agent look things up when it needs them, rather than stuffing everything into the prompt. This is the "agentic" form of RAG: the model decides what to search for, reads the results and can search again with a better query.

What agents can reach

Most tools fall into a few families: knowledge, computation, systems and the screen.

A search tool that cites sources

Return short snippets with ids so answers can cite them and the model can fetch a full document only when needed.

@tool
def search_docs(query: str, limit: int = 5) -> str:
    """Search the help-centre articles. Returns up to `limit` snippets as
    [doc_id] snippet. Call get_doc(doc_id) for the full text."""
    hits = index.search(query, limit)
    return "\n".join(f"[{h.id}] {h.snippet[:200]}" for h in hits)

Retrieved text is untrusted

A document in your index might contain instructions aimed at the model. Treat retrieved text as data, never as commands, and see the safety section for defences.

Quick check: What does "agentic RAG" add to plain RAG?

• It removes the need for documents
• The model decides what to search and can search again
• It guarantees correct answers
• It works without a search index