Extractor: Dossier (voice/dictation)
Deep relationship-dossier extraction for voice and dictation captures.
What it does
When you finish a capture in MyVault (a voice memo, a dictated note about someone you met, a quick reflection on a conversation), this capability runs MyVault’s dossier-grade LLM pass over the transcript and turns it into structured findings inside the Knowledge Extractor.
Concretely, for each person mentioned in the capture it produces:
- A person finding carrying their name, company, role, location, how you met, what they’re focused on, background, sentiment, key themes, what they want, what you want from them, and personal notes — anything the dictation surfaced.
- A follow-up finding if you mentioned next steps for that person, so reminders can pick it up.
These findings flow through the Extractor’s normal review and routing just like findings from any other source — once approved, person records land in your contacts / dossier and follow-ups land in reminders.
This capability is selective by design. It only runs on capture transcripts handed in by MyVault. Emails, pasted text, calendar events, and every other source continue through the generic extractor so you don’t get duplicate findings.
Settings
Capture confidence (default 0.85)
The confidence value assigned to person and follow-up findings
produced from your dictation. Captures are first-hand — you stated
the information directly — so the default is high. The Extractor
uses confidence to decide whether a finding can auto-route or needs
your review; raise this if you want capture findings to auto-route
more often, lower it if you want a closer look at each one before it
lands.
The value must be between 0 and 1. Any individual person in a
capture can carry their own confidence (when the LLM is uncertain
about a name or detail), and that per-person value is used instead.
Commands
This capability has no user commands. It runs automatically when MyVault hands a finished capture to the Extractor.
Integrations
- MyVault — required. The dossier LLM pass lives in MyVault and is invoked through it. Without MyVault installed and enabled, this capability stays idle and captures fall through to the generic extractor (which produces shallower, email-shaped findings).
- Knowledge Extractor — required. This is the host that calls into this capability. Without the Extractor installed there is nothing to plug into.
- An LLM provider — required. The dossier pass is an LLM call, routed through MyVault’s configuration. Whatever provider you’ve set up for the agent (Claude, Gemini, or another) is used.
Known limits
- Captures only. This capability deliberately ignores emails, pasted text, calendar items, and every non-capture source. The generic extractor handles those.
- No MyVault, no depth. When MyVault isn’t installed the capability returns no findings — it doesn’t fall back to a lighter extraction. The capture is still preserved as raw text in the extractor’s ledger and can be re-extracted later if you install MyVault.
- Names are required. Person findings need a name. People mentioned only obliquely (“my friend from the conference”) are skipped rather than guessed at.
- LLM failures degrade gracefully. If the LLM call errors or times out, the capability returns no findings rather than breaking the capture flow. The raw transcript is still preserved.
- No state of its own. This capability doesn’t accumulate any history between captures — each capture is extracted on its own. Repeated dictations about the same person create repeated findings; the Extractor and the contacts surface handle deduplication.