Why narrow beats general

An assistant presented as able to answer anything about your company sets the visitor's expectation accordingly. They will ask about anything, including the things you never documented, the things that require account access, and the things that are somebody's job rather than a page.

Most of those questions will fail. Each failure is measured by the visitor against the promise, so a general assistant with good coverage of most of your business still reads as unreliable, while a narrow one with complete coverage of one area reads as competent.

Nothing about the underlying quality changed. What changed is what the visitor was told to expect, which is entirely within your control and usually set by accident.

By product line. One assistant for one product, with its documentation, its pricing and its support material. Useful when different products have different customers, different vocabularies or different support teams.

By audience. Separate handling for prospects and existing customers is the most common version, and often the most valuable, because the two want opposite things. A prospect wants to know whether it fits and what it costs. A customer wants to know how to do something specific and does not want to be sold to.

By site section. An assistant that appears only in the documentation area, or only on the pricing pages, or only at checkout, scoped to the material for that section. This is the easiest to implement because it maps to pages you already have, and it is the one that most reliably improves answer quality, because the visitor's likely question is already constrained by where they are standing.

The mechanism is not mysterious. Retrieval selects passages by similarity, and every additional document is another opportunity to select the wrong one. Documents that are similar in wording but different in scope are the dangerous case: the enterprise pricing page and the small business pricing page, the current terms and the terms for legacy accounts, the same policy for two different countries.

In a narrow corpus those near collisions largely disappear, because the competing document is not there. The question about pricing has one pricing page to match against rather than four, and the answer is right by construction rather than by luck.

This is also why adding material does not monotonically improve an assistant. Past a point, each additional document adds more collision risk than coverage, and the useful move becomes removing things rather than adding them.

A narrow assistant has fewer documents, which means fewer review dates, fewer owners, and a realistic chance that somebody actually reads all of it once a quarter.

It also localises the blast radius of a mistake. If the documentation assistant is wrong, it is wrong about documentation, and the person who owns documentation can fix it without consulting anybody. A general assistant that is wrong could be wrong about anything, which means every incident needs triage before it needs a fix.

And it makes the ownership question tractable. It is much easier to name an owner for the assistant that covers one product than for the one that speaks for the entire company, which is exactly the ownership problem that leaves general assistants unmaintained.

A narrow assistant can state its scope in one sentence, and that sentence does real work. I can answer questions about the documentation for this product. I answer questions about ordering, delivery and returns.

Told that, a visitor can immediately judge whether to ask. If it covers their question they proceed with confidence. If it does not they go elsewhere without wasting a turn, which is a better outcome than a failed attempt.

It also makes the refusal coherent. A narrow assistant refusing something outside its stated scope is behaving exactly as advertised, and reads as a correctly designed tool rather than as a broken one. The same refusal from something that claimed to answer anything reads as a failure.

You do not have to rebuild. Look at a month of conversations and sort the questions by area. There will usually be one or two areas that account for most of the volume and answer well, and a scattering of others that answer badly.

Cut the corpus down to the areas that work, change the opening message to state that scope, and route everything else to a person. The refusal rate will rise and the quality of what remains will rise faster.

Then, if there is a second area with real demand, set it up separately with its own material rather than merging it back in. Two narrow assistants on different parts of a site are more accurate and more maintainable than one that covers both, and the only thing you lose is the tidiness of having one of something.