Try it now
A working sandbox. No sign-up, no project. Sample data only.
- Retrieves soundbites by idea, not by exact phrasing. Describe the gist and the chat locates the moment.
- Surfaces causality: ask why something happened and the answer draws connections across what multiple speakers said at different points in the shoot.
- Finds references and contradictions that would otherwise be cut because no one could place the exact line.
- Every answer is grounded in indexed transcripts; citations show speaker, file, and timecode for every claim.
Click a question to see a grounded answer with citations.
How it works
Three steps from raw material to result.
Where do Marcus and Elena contradict each other?
Ask in plain language. Try: 'Find the moment Marcus first admits doubt about the numbers', 'What does Elena say that contradicts the board's account?', or 'Any moment where someone mentions the inspection memo?'
The indexed transcripts are ranked by relevance to the question. The strongest passages are assembled into context, regardless of whether your phrasing matches the speaker's.
Their clearest contradiction is on who initiated the override. Elena says she made the call. Marcus says that's not how he remembers it.
Each claim in the answer is sourced to a specific file, timecode, and speaker. Jump to any citation directly from the answer.
Frequently asked questions
Can it make things up?
The model is constrained to answer only from the indexed transcripts. If the archive does not contain an answer, it says so. It does not draw on outside knowledge.
Does it search all files in the project?
Yes. The chat searches across every indexed transcript in the project. The index grows automatically as new files are transcribed.
What kinds of questions work best?
Soundbite retrieval ('find where Elena mentions the memo'), causality ('why does Marcus say the numbers stopped making sense, and what happened before that in the shoot?'), contradiction spotting ('where does anyone contradict the timeline Elena gives in Day 04?'), and thematic search ('find moments about silence or withholding').
Is the chat history saved?
The conversation persists within the session. Clearing the chat starts fresh; the indexed transcripts are unchanged.
How is this different from semantic search?
Semantic search returns ranked passages for a query. Chat synthesises an answer from those passages and writes it in plain language, with citations preserved so you can verify every claim and jump straight to the tape.
Related capabilities
Further reading
Background guides and comparisons.
Retrieval-augmented generation answers questions about indexed interview transcripts by retrieving relevant chunks and grounding the model's output in those chunks. Here is how the pipeline works and where it still fails.
Embedding-based semantic search finds passages by meaning rather than by matching words. Here is what an embedding is in plain terms, how the search works, and where it still misses things.