The RAG run: search → ground → answer

Background

The v0.2 agent retrieves first, then answers only from what it retrieved:

query → search(query) → format context → LLM(system + "Context:…\n\nQuestion:…") → grounded answer

Problem statement

Implement rag_run(query, search, llm) that runs the full retrieve-then-answer pipeline.

Input

• query — the user question string.
• search — callable query -> list[str] (the search tool).
• llm — callable messages -> str (the chat model).

Output

Returns the LLM's answer string. Internally:

1. docs = search(query).
2. context = "\n".join(f"- {d}" for d in docs).
3. messages = system "Answer based ONLY on the provided context. If the context doesn't cover the question, say so." + user f"Context:\n{context}\n\nQuestion: {query}".
4. return llm(messages).

Examples

Example 1 — the answer is grounded in the retrieved context

search = (q -> ["AlphaCode used CodeContests."])
llm    = (messages -> messages[-1]["content"])   # echoes the user message
rag_run("What benchmark?", search, llm)
# -> "Context:\n- AlphaCode used CodeContests.\n\nQuestion: What benchmark?"

Constraints

• Call search(query), format docs as bullets, build the two messages, return llm(messages).
• The user message must be exactly f"Context:\n{context}\n\nQuestion: {query}".
• The system message must instruct the model to answer only from the context.

Notes

• The agent never sees the search backend or the model — only their interfaces. Swap either without touching this loop.