Answering

No single model provider is available all the time.

Routing is not about chasing the best model. It is about cost, latency and the fact that every provider has incidents, and yours will happen on a weekday afternoon.

One question, several possible modelsA router sends a question to a primary model and falls back when it is unavailable.QuestionRouterPrimaryFallbackCost, latency and availability
  • Private launchOnboarding selected teams now
  • Answers from your pages onlyIt refuses when your content does not cover it

What it does

Routing decides which model answers a given question. The decision is made on difficulty, cost and availability, and it includes a fallback path for when the first choice is unavailable, rate limited or too slow.

Why it matters

Most website questions are not hard. Reading three retrieved passages and phrasing an answer from them is comprehension, not reasoning, and a mid tier model does it as well as an expensive one for a fraction of the cost per conversation.

The reliability argument is the stronger one. When a provider has an incident, an assistant without a fallback becomes a visible outage on your marketing site. With one, it becomes a slightly different tone of answer that nobody notices.

The thing routing does not fix is retrieval. If the wrong passage was retrieved, the best available model will answer confidently and wrongly, because from its position it was handed the relevant context and asked to use it.

How it works

Classify the question

Short factual lookups, comparison questions and multi part questions have different requirements. Classification is cheap and happens before any expensive call.

Route on cost and difficulty

Simple lookups go to a fast, inexpensive model. Questions needing synthesis across several passages go higher.

Watch the clock

If the primary has not begun responding within the latency budget, the request moves. Time to first token is what a visitor feels, not total time.

Fall back on failure

Provider error, rate limit or timeout moves the request to the next model rather than showing an error.

Keep the same grounding

Whichever model answers, it receives the same retrieved passages and the same instruction to refuse when they do not cover the question. Routing changes who phrases the answer, never what it is allowed to say.

One question, several possible modelsA router sends a question to a primary model and falls back when it is unavailable.QuestionRouterPrimaryFallbackCost, latency and availability

What routing cannot fix

Routing changes who phrases the answer. It does not change what the answer is made of, and almost every accuracy complaint is about the material rather than the phrasing.

If retrieval returned a passage from an outdated policy page, every model in the chain will answer from it, because each one was handed that passage and told to use it. The expensive model will produce a more fluent version of the same error, which is worse, because fluency is what makes people believe it.

This is why upgrading the model is usually the wrong first response to a bad answer. Read the citation. If the cited page is wrong or stale, the model was working correctly and the index was not.

The reverse also holds. If the cited page is right and the answer misread it, that is a phrasing problem and routing is the correct lever. Distinguishing the two takes ten seconds with a citation and is impossible without one.

What is built and what is planned

Routing across providers with automatic fallback is core to how Creobot is built. Per plan control over which models are used is not built and is not being promised.

Where the data goes

The question and the retrieved passages are sent to whichever provider handles the request. The subprocessor list names every provider in that path. Security and data boundaries.

Questions

Several, and the specific set will be named on the subprocessor list before anyone is charged. Naming a lineup now would be a claim we would have to revise.

Not today, and it is not on the near list. The honest reason is that most questions do not benefit from a more expensive model, and the ones that do are better handled by routing than by a global setting.

Slightly, in phrasing. It receives the same retrieved passages and the same grounding instruction, so what it is permitted to say is identical.

The assistant says it cannot answer right now rather than showing a broken widget, and the conversation can escalate. A chat window that silently fails is worse than one that admits it.

In wording, marginally. In substance, no, because substance comes from your passages rather than from the model's own knowledge. This is a benefit of grounding that is easy to miss.

Usually not. If retrieval returned the wrong passage, a better model produces a more fluent wrong answer. Check the citation before changing the model.

Signal

Is this the capability you need?

Access is by waitlist, demo request or public-content pilot. Tell us what your site has to answer and we will say honestly whether this covers it.