Unit 6 of 9 · Intermediate

Grounding an agent in your own information

In one read

An agent's model was trained on general information, not your filings, your policies, or today's data. Retrieval is how it works with material it never saw in training. The technique — often called RAG, retrieval-augmented generation — is straightforward: at the moment a question is asked, the system fetches the documents or records most relevant to it and places them in the context window, so the model answers from that material rather than its general memory. This is how an agent reads a specific filing, applies your written policy, or reasons over a live data set instead of guessing from what it half-remembers. It also explains a common failure: when an agent gives a confidently wrong answer about your own information, the cause is usually bad retrieval — the right document was never fetched — not a bad model. Fix what the agent is shown, and the answer often fixes itself.

The model doesn't know your information

The model at the center of an agent was trained on broad, general text. It never saw your filings, your internal policy, your desk's positions, or today's numbers. Ask it about those from memory and, at best, it reasons from general patterns; at worst, it invents.

Retrieval is how an agent works with material the model was never trained on. The common name is RAG — retrieval-augmented generation — and despite the acronym it is a simple idea.

How retrieval works

Recall the desk from the 101 unit: the model can only reason about what is on its context window right now. Retrieval is the act of putting the right papers on that desk at the right moment.

  1. A question comes in.
  2. The system fetches the documents or records most relevant to that question — from your document store, a database, a live feed.
  3. Those are placed into the context window, alongside the question.
  4. The model answers from that material, not from its general memory.

So retrieval is how an agent reads a specific 10-K, applies your written risk policy, or reasons over this morning's data — grounded in real source material instead of half-remembered generalities.

Why bad retrieval looks like a bad model

Here is the practical lesson executives most often miss. When an agent gives a confident, wrong answer about your own information, the culprit is usually not the model — it is retrieval. The right document was never fetched, so it never reached the desk, so the model answered from generalities and got it wrong.

The fix is rarely a smarter model. It is fixing what the agent is shown: better sources, better fetching, the relevant record actually retrieved. Improve the input and the answer often corrects itself.

See it in kestrel

Kestrel grounds its agent in real, recorded session data rather than a half-remembered market — the "documents" here are the tape itself. Watch a run on a generic instrument, no signup, no card:

npx kestrel.markets sim fomc-rate-decision-whipsaw

It prints a certified proof URL you can recompute yourself. Drop the kestrel.markets MCP server into your client and the grounding is set up again next session — no account in between.

Recompute it

Every claim in this unit recomputes from a certified proof — no account, no card.

/proof/art_d29415f0cf502f4a218a9cba
Keep readingHow an agent reaches out of the chat box

Agent twin: /education/agents-fundamentals/af-context-and-retrieval/content.md