Reference
Glossary
This category uses words like grounded, cited, and evidence loosely enough that they have stopped meaning anything. These are the definitions we work to, including the ones that describe how we meter and delete your data.
- Terms
- 22
- Updated
- 8 October 2026
Index
Jump to a term.
Definitions
The vocabulary, defined precisely.
- Evidence layer
- The index that sits between raw sources and the answers people act on. It always keeps core transcript content intact; verified optional visual generations and document pages join as available, so each claim can point back at its real source.
- Frame citation
- A citation that points at a specific video frame rather than a line of transcript. It matters when the evidence was visible rather than spoken: an error toast, a form validation state, a URL, a value on a dashboard.
- Page citation
- A citation that names the page and section of a document an answer was read from. For scanned documents the page still has to be read visually first, which is why scanned pages are metered separately.
- Timestamp citation
- A citation fixed to a second in a recording, so playback can jump straight to the moment. Combined with the speaker and the verbatim quote, it is what makes a spoken claim checkable.
- Evidence pack
- A deliverable made of numbered claims plus an appendix of citations that resolve to moments, frames, or document pages. Used for UAT sign-off, client findings, and audit responses, where a summary would not be accepted. In Citesvue the building blocks are cited findings exports (PDF, Word, CSV); a one-click packaged brief is on the roadmap.
- Retrieval grounding
- Retrieving source passages before composing an answer, then constraining the answer to what those passages support. The opposite of generating first and looking for support afterwards, which is where confident fabrication comes from.
- Refusal behaviour
- What a grounded system does when the evidence is thin: decline, or answer with a stated hedge and show the weakest support, instead of producing a fluent claim nothing backs up.
- Typed artifact
- A structured finding extracted from a session. Citesvue creates seven transcript-derived types (bug, requirement, decision, action item, risk, question, insight) plus a recap summary. Core findings carry quote, speaker, timestamp, and (for bugs and risks) severity; optional visual evidence can later add screen corroboration without replacing the finding. Typed because it can be routed: a bug belongs in an issue tracker, a decision belongs in a record.
- Recap
- The first view of a processed recording: a short summary with key outcomes, where each outcome jumps the player to the moment behind it. The recap is generated from the cited record, so improving the prose can never quietly change what the meeting decided.
- Deep Dive
- The optional visual record of a video, added when you choose Analyze visual evidence: what was on screen at each moment, the scenes, on-screen text, visual findings, and frame citations. Audio-only recordings do not offer it, and a video shows no Deep Dive tab until visual evidence is requested.
- Meeting assistant
- The notetaker that joins a live meeting as a visible, named participant with a recording notice, records the call, and feeds the capture into the same transcript-first flow as an upload. Optional visual evidence remains a separate user action.
- Auto-join
- The calendar-driven mode of the meeting assistant: connect a calendar in Settings, under Calendars, and the assistant joins scheduled meetings on its own. In closed beta. The path available to everyone is pasting the meeting link on Meetings and choosing Send assistant.
- Project
- A grouping of recordings and documents (a client, a product, a sprint) that is also a question scope: asking a project answers from everything inside it, citing spoken evidence by timestamp and written evidence by page.
- Visual-evidence hours
- The monthly allowance for optional scene, OCR, visual-finding, and frame analysis: explicitly analyzing a 30-minute video uses 30 minutes. It pools across seats. Transcription, transcript artifacts, and transcript Q&A do not consume it.
- Voice profile
- A person in the workspace voice directory: a name, and optionally an email and role. Tagging a speaker links them to a profile, and the voiceprint kept from confirmed samples lets Citesvue recognise the same person in later recordings in that workspace. Confident matches are applied automatically; uncertain ones wait for someone to verify them. Removing a profile unlinks it from every speaker it named.
- UAT sign-off
- Formal acceptance by the business owner that a release meets its acceptance criteria. It transfers risk only when it names the build tested, the scope covered, and the defects knowingly accepted, each with evidence attached.
- Document pages
- The unit Citesvue meters document processing in: every page processed, in any supported format. It is the outer allowance that all document work is counted against.
- OCR pages
- A sub-allowance inside document pages, counting only the pages that had to be read visually because they were scanned. A scanned PDF page consumes one document page and one OCR page, so the smaller number is the effective ceiling.
- Hybrid retrieval
- Searching by exact wording and by meaning at the same time, then merging and re-ranking the results. Keyword search alone misses paraphrase; meaning-based search alone misses exact strings such as error codes and clause numbers.
- Deletion cascade
- Deleting a source removes everything derived from it as well: transcript, artifacts, any requested scenes, OCR and frames, and both core and evidence index entries. Without a cascade, deletion is only a hidden copy.
- Citesvue Guide
- The help assistant, opened from Ask the Guide at the bottom of the app's sidebar or the button on these documentation pages. It answers how-to questions from the Citesvue documentation and links to the pages it used; signed in, it can also answer about your own plan, usage, and connected tools. It is read-only and keeps your chats in your browser.
- Evidence-grounded AI
- A system designed so every output can be traced to a specific source location, and so unsupported outputs are refused rather than produced. Grounding is a property of the pipeline, not a claim about the model.
Put it to work