An assistant is only as good as what it is allowed to read

The model is the last mile

A support assistant is a retrieval system with a language model on the end of it. Almost everything that makes one useful, or dangerous, happens before the model is called at all — in the step that decides which of your own records it is looking at.

Swapping one model for a better one changes the phrasing. It does not change whether the right paragraph was in front of it.

Retrieval is the product

Retrieval here means an index over the material the company actually answers from: the help centre, the policy documents, past resolved tickets, and the fields on the order record. Postgres with pgvector is usually enough, and the index is not the hard part.

Deciding what belongs in it is. A document that is out of date, internal-only, or written for a different market is not neutral once it is retrievable. It is a wrong answer waiting for the question that matches it.

Chunking is an editorial job

Splitting documents for retrieval gets treated as a parameter. It is closer to editing. A policy cut in the middle loses the condition that made it true, and the half that survives into the index reads as unconditional.

Citations are not decoration

An answer a person cannot check is not an answer. It is a claim with a confident tone.

Every response should carry the records it was built from, linked, so the person reading it can go and look. That single requirement changes the system: it forces retrieval to be honest about what it found, and it gives the support team somewhere to send a correction other than the output.

Permissions travel with the query

The index does not get its own view of the world. A retrieval step that ignores who is asking will eventually put one customer record in front of another customer, and it will do it for the first time in production.

Filtering at query time, from the identity of the requester rather than from a flag on the document, is the only version of this that holds as the corpus grows.

Teach it to stop

The most valuable behaviour is refusal. Below a confidence threshold, or with no supporting record above a similarity floor, the correct output is that the assistant does not know and a person is being fetched. That is a design decision, and it belongs in the retrieval layer rather than in the prompt.

hits = retrieve(question, viewer)
if best(hits) < floor: escalate()

Freshness beats cleverness

An index rebuilt on a schedule is behind the business by exactly that schedule. Where the source is a database, read it live and cache the result. Where it is a document, make publishing update the index, so the answer changes when the policy changes rather than at the weekend.

What good looks like here

A useful assistant answers narrowly, shows its sources, refuses more often than feels comfortable at first, and gets braver as the material behind it improves. The work is in the material and in the boundary. It was never in the wording.