What happens when no transcript supports the question

When no transcript passage supports a question, SageTube tells you so instead of guessing. The chat returns a fixed message — “I do not have enough source support to answer reliably from the current corpus.” (app/Services/AnswerValidator.php:29) — the answer is labeled unsupported, and no citations are attached, because there is nothing to cite. Crucially, this is not a politeness the AI model may or may not observe: a server-side validator inspects every answer’s evidence references and forcibly downgrades any claim of support it cannot verify. Refusal is enforced in code, after the model has answered.

This post is a close-up on the refusal path of the grounding model described in Citations, Timestamps, and Trust.

Why refuse instead of answering anyway?#

A general-purpose chatbot answers from whatever its training absorbed. There is no fixed corpus behind the answer, so there is nothing to check the answer against — and when the training data runs thin, the fluent-sounding completion and the fabricated one look identical to the reader.

A SageTube Expert is the opposite arrangement: a closed corpus of specific transcripts, where every answer is supposed to be traceable back into that corpus. The model is instructed to answer using only the provided source passages (app/Client/StructuredAnswerClient.php:230), and — as we argued in Why every AI answer needs a citation — an answer you can’t trace is an answer you can’t verify. In that model, “the corpus doesn’t cover this” is a legitimate, useful answer. Papering over it with a plausible guess would break the one property the product exists to provide.

How does SageTube decide an answer is unsupported?#

Every answer must declare what kind of answer it is. The response schema forces the model to label each answer as one of directly_supported, synthesized, weakly_supported, unsupported, or corpus_meta (app/Client/StructuredAnswerClient.php:324), and to attach evidence references — IDs of the specific retrieved passages — to each point it makes.

The model’s self-assigned label is only the starting point. After the model responds, AnswerValidator re-checks the label against the evidence:

• Citations must point at real evidence. Any reference to a passage that was not actually retrieved for this question is stripped before anything else is evaluated (app/Services/AnswerValidator.php:119). The model cannot invent an “E7” that was never on the table.
• “Directly supported” requires surviving references. If an answer claims direct support but no point retains a verifiable reference after stripping, it is relabeled unsupported (app/Services/AnswerValidator.php:369).
• Title matches don’t count as evidence. An answer whose every reference resolves to a source card — a video’s title and metadata rather than its transcript content — is relabeled unsupported too (app/Services/AnswerValidator.php:369). Restating a title that merely echoes your question is not an answer from the video.
• “Synthesized” requires at least two distinct sources. Synthesis is a claim about drawing across sources; with one real source it is downgraded to a direct or weak answer, and with none it becomes unsupported (app/Services/AnswerValidator.php:420).
• An empty answer is a refusal. If the model returns empty answer text, the validator substitutes the stock refusal message and the unsupported label (app/Services/AnswerValidator.php:134).

The same stock message is also the hard fallback when retrieval itself fails with no evidence available (app/Services/ExpertAnswerService.php:482): the system degrades to an honest “I can’t answer this reliably,” never to an uncited guess.

What about videos that aren’t in the Expert at all?#

Two edge cases get explicit handling rather than a shrug.

If an Expert has sources that are still processing — or none at all — SageTube short-circuits before retrieval and explains the actual state of the knowledge base instead of answering from nothing (app/Services/ExpertAnswerService.php:88).

And when you use the Chrome extension on a YouTube video that hasn’t been added to your Expert, the assistant is explicitly instructed that it has no access to that video’s content and must say so: “Do NOT fabricate or guess the video’s content” (app/Services/ExpertAnswerService.php:117). It can still answer general questions from the sources it does have — it just won’t pretend to have watched a video it never indexed.

There is also a middle state between full support and refusal. When relevant passages had to be excluded because the evidence exceeded the context budget, the model is told to hedge and prefer the weakly_supported label (app/Client/StructuredAnswerClient.php:218), and answers built on degraded evidence are downgraded from stronger labels server-side (app/Services/ExpertAnswerService.php:412). Uncertainty is surfaced, not smoothed over.

How often does this actually happen?#

Refusals are not a theoretical guarantee — we count them. Every question a visitor asks on a public Ask page is recorded with its final answer type (app/Jobs/RecordAskPageQuestion.php:45). Over the 90 days ending July 31, 2026, visitors asked 202 questions on public Experts: 142 were answered as synthesized across multiple sources, 25 as directly_supported, and 35 — 17.3% — came back unsupported. Roughly one visitor question in six hits a topic the creator’s transcripts genuinely don’t cover, and each of those got an honest refusal instead of a fabricated answer.

We also monitor the refusal rate itself. An hourly production check computes each public Expert’s 7-day unsupported rate and pages the operator when any Expert with real traffic crosses 25% (app/Services/Qa/Tests/Infrastructure/AskPageUnsupportedRateTest.php:39) — because a spike in refusals can mean an off-topic audience, but it can also mean a retrieval bug making the corpus look emptier than it is. Refusing when the corpus is silent is correct; refusing when the corpus has the answer is a defect we alarm on.

What should you do when you get a refusal?#

Treat it as information about the corpus, not a dead end. If you’re the visitor, the creator likely never covered that topic — try rephrasing toward what the channel actually discusses, or explore other public Experts. If you’re the Expert’s owner, a refusal on a topic you know you’ve covered usually means the relevant videos aren’t indexed yet — add them, and the same question starts returning cited answers.

Either way, the refusal is the trust model working as designed: when SageTube gives you an answer with citations, it’s because the evidence is really there.