Skip to main content
When an agent answers a question using Pureframe AI, the answer isn’t a summary generated from a general impression of your library — it’s grounded in specific, retrievable evidence.

​
What “source-backed” means here

Every search_videos result carries the pieces needed to verify the claim, not just cite it:
  • video_id, start_secs / end_secs — exactly where the moment is
  • clip_url — a playable link to that exact segment
  • thumbnail_base64 — the actual frame the model saw, so you (or another model) can check the claim independently
  • text_snippet — the transcribed words, if the match came from speech
  • relevance_score — how confident the match is, so low-confidence results can be treated with appropriate skepticism
If an agent tells you “the presenter mentions a 20% discount at 4:32 in the Q3 demo,” that’s not an inference from a cached summary — it’s read directly from a transcript segment at timestamp_start: 272.0, and the URL to verify it is right there in the same tool response.

​
Why this matters for agent reliability

Video is opaque to most models by default — they can’t watch a long video before answering. Pureframe AI’s search step is what gives the model something concrete to point to, rather than having it guess or hallucinate content it never actually saw. If a model tries to answer without calling a tool first, it has no video_id or timestamp to offer — a useful tell that the answer isn’t grounded.

​
Building this into your own UI

If you’re building a product on top of Pureframe AI rather than using a chat agent directly, carry the same fields through to your UI: show the timestamp and a link to the clip next to any AI-generated summary of video content, not just the summary text alone. See Search Results for the full response shape these fields come from.