What is a confidence score in AI answers?
A confidence score is a numeric estimate of how well-supported an answer is, most usefully derived from how closely retrieved passages matched the question.
Two very different things get called confidence. Retrieval confidence measures how close the best-matching passages were to the question — a property of the search step, computed before any answer is written. Model confidence is the model's own assessment of its answer, which it will happily report as high while inventing a price.
Why it matters
Only retrieval confidence is a usable control signal, because it is measured against something external. It answers a question with a real answer: did we find anything relevant? When the score is low, the honest response is to refuse and offer a person, and to log the question as a gap. A system that asks the model how sure it is has asked the wrong entity — a model that hallucinates does so with the same fluency it uses when correct, and its self-report inherits that fluency.
Examples
- A strong match against the shipping page, answered directly with a citation.
- A weak best match, refused and handed off — which is the correct outcome, not a failure.
- A model reporting high confidence in an invented delivery estimate, which is why it is not consulted.
Not the same as
Accuracy
Confidence is measured per answer before anyone knows if it is right. Accuracy is measured afterwards, across many answers.
Related
term
What is semantic search?
Keyword search asks whether two pieces of text share words. Semantic search asks whether they mean the same thing. The difference shows up immediately in customer questions, because customers do not use a company's vocabulary: someone asks 'can I send it back?' about a page headed 'Returns and exchanges', and the two share not one significant word.
term
What is cosine similarity?
It is the standard way to compare embeddings, and its defining property is that it ignores magnitude entirely. Two vectors pointing the same way score identically whether one is twice as long as the other. In practice, scores for text embeddings tend to fall in a narrow band — often 0.6 to 0.9 for anything vaguely related — which is why a raw score is meaningless without knowing the distribution for a given model.
term
What is a handoff in customer support?
Every automated support system needs a way out, and the quality of that exit matters as much as the answers. A good handoff carries the transcript, so the customer is not asked to repeat what they already typed — which is the single most common complaint about support automation, and one that is entirely avoidable.
term
What is grounding in AI?
A grounded system retrieves passages before it writes anything, puts them in front of the model, and instructs it to answer only from what it was given. The test of whether a system is grounded is simple and does not require access to its internals: ask it something, then ask it where the answer came from. A grounded system names a page. An ungrounded one describes its reasoning.
use case
An AI receptionist for your website
Most of what a receptionist repeats all day is already published: opening hours, where to park, what a first visit costs, whether you take a particular insurer. Those are answerable from the website. Booking, rescheduling and anything about a specific person are not.
term
What is an AI hallucination?
Language models predict plausible text, not true text. Asked a question with no supporting material, a model does not fall silent; it produces the most likely-sounding answer, which is often a wrong price, a policy that does not exist, or a citation to a page that was never written. The behaviour is not a bug to be patched out of the model. It is what the model does when nothing constrains it.
RubyRep is this, applied to your website.
Give it your URL and it reads your pages, then answers your customers from them.
Train My RubyRep