knowledge-meeting-ingest

# Meeting Ingestion Skill — Unified Pipeline Follow gbrain's meeting-ingestion phases and verification gate. Read [shared conventions](references/conventions.md), [publication](references/publication.md), and `knowledge-entities`. Source-specific fetch/parse is the host agent's job; normalize once, then run this pipeline. Do not build a per-vendor pipeline or a Knowledge ingestion job. ## Contract - One page per REAL meeting; split multi-meeting recordings first. - Retain the full transcript and provenance; recorder summaries are claims, not facts. Speakers are resolved by evidence, never by guess. - Meeting page includes attendees, Summary, Key Decisions, Action Items and Notable Quotes. Empty sections carry an explicit source-grounded reason. - Every sufficiently identified actual attendee gets a substantive person page. Every relevant company/project/concept gets a Timeline entry and back-link. - Timeline context and bidirectional back-links cover all resolved mentioned entities with pages. Unverified claims do not fan out. - Supported decisions, commitments and events become independently retrievable claims. Explicit relationships carry evidence, not just co-mention. - Never report a meeting fully ingested until every output and verification item is accounted for. Verify sequence independently from substance. ## The normalized transcript record Reduce the input to this shape in task working context; it is not a new API: ```yaml source: "<recorder name, or 'manual'>" source_id: "<recording ID, if supplied>" title: "Meeting Title" date: YYYY-MM-DD # null if unknown time: "HH:MM TZ" # null if unknown duration: "45m" # null if unknown attendees: # source claims, not verified identities - name: "..." email: "..." # only if source provides it role: "..." # only if known transcript_segments: - speaker: "..." # resolved name OR explicitly unresolved speaker_raw: "..." # retain original label for traceability text: "..." raw_transcript_text: "..." # complete available transcript, never truncate source_summary: "..." # recorder summary is a CLAIM, not a FACT source_url: "..." # if supplied ``` Retain source segment order/timestamps when present. A summary is two lossy layers removed (transcription + summarization). Retain the original export and transcript with `knowledge_attach`, then attach returned evidence to the meeting page. Normalized text is a labeled derivative, not a replacement for the original. Inspect for captured credentials before retention; redact live secrets and label the redacted derivative honestly. Respect requested privacy limits; never call a redacted file byte-identical to the input. ## Phases ### Phase 1: Normalize the input Acquire the user-selected source with available host tools and normalize it. Empty transcripts, summary-only payloads and in-progress recordings do not pass this phase. Report the missing source; do not create a completed meeting from a summary alone. If the user requested saving only notes, use `knowledge-capture` for attributed notes and explicitly state that meeting ingestion is not complete. Transcription is a host capability, not a Knowledge tool. For API acquisition, read [source acquisition](references/source-acquisition.md): verify date filters, follow every transcript cursor, and record coverage before normalizing. It includes the Granola endpoint/parameter reference. ### Phase 2: Split detection — one recording is not always one meeting Investigate roster shifts, topic hard-cuts, fresh introductions/context resets, or an agenda naming multiple sessions. Two or more signals strongly suggest a split. Same people in one flowing conversation changing topics is ONE meeting. For a supported split, partition the transcript into contiguous nonoverlapping segment ranges: no dropped or duplicated segments, original order preserved. Run remaining phases once per meeting, with its own title/date/attendees. Identify shared source ID and exact boundaries on each page; cross-link siblings after IDs/paths are known so dedup does not merge them back. Flag uncertain boundaries. ### Phase 3: Dedup across recorders Search existing meetings by source ID, title/attendees and date. Gbrain's date ±1 day, attendee overlap ≥50%, similar title rule is a candidate screen, not proof. Read the candidates and coverage before deciding. If the same meeting exists, reconcile into it instead of creating a duplicate. Retain both source IDs/originals and coverage differences. Use the richer transcript for coverage; source completeness is not speaker-label accuracy. Record source identifiers in prose or allowed ontology fields, never arbitrary frontmatter. Inspect existing assertions/connections as well as the meeting page. ### Phase 4: Speaker resolution — before writing the page 1. Start with the source attendee list and explicit roster/user evidence. 2. **Never guess.** Use `[Room]` or `UNKNOWN` with an uncertainty flag when needed. 3. Resolve candidates through `knowledge-entities`; read existing evidence. 4. Record raw label → supported identity or unresolved, with basis and qualitative confidence (high/medium/low). Confidence is attribution confidence, not truth. A roster-confirmed attendee still needs evidence linking them to the track. 5. **Content-identity check:** compare a speaker's claimed role/company with the named person's evidence. A contradiction signals possible mislabeling; inspect before reattributing a whole track. Do not treat every new role as an error. 6. **Phantom-speaker check:** more tracks than confirmed attendees, mid-thought handoffs or labels never addressing each other suggest over-splitting. Merge labels only when evidence supports the identification, not from counts alone. 7. **User ground truth:** an explicit user identification corrects the mapping. Retain it as attributed evidence; preserve the original recorder transcript. Route repairs of existing assertions through `knowledge-correct`. 8. Search plausible spelling variants before creating a transcript-only person. Once identity is confirmed, reuse the page and optionally add the alias while preserving existing aliases. A near-match alone is not a confirmed identity. ### Phase 5: Create meeting page Use a clean source title; store the occurrence date/time in the meeting ontology's `fields.occurred_at` at source precision. Include source timezone context in the body. Do not append an ingestion date to a title or guess a timezone to resolve a relative deadline. A date-only value is not a midnight instant. Use the current ontology's meeting definition and actual resolved page links: ```markdown # {Meeting Title} **Attendees:** {resolved people with links; uncertain identities flagged} **Date:** {known date, otherwise unknown} **Duration:** {if available} ## Summary {3–5 key outcomes, each citing the lines that support it} ## Key Decisions {Decisions with context and citations. If none: _No decisions — discussion only._} ## Action Items {Supported commitments, owners/deadlines when known, citations. If none: _None — exploratory conversation._} ## Notable Quotes {Verbatim transcript spans, attributed, in blockquotes with citations. If none: _No notable quotes — operational/logistics meeting._} ## Discussion Notes {Structured notes by topic, with citations} ## Source Context {Only relevant provenance: source ID/URL, raw speaker label → supported identity or unresolved with basis, coverage/split boundaries or transcription limitations. Use readable inline citations, not artifact paths or hashes.} ``` Summary, Key Decisions, Action Items and Notable Quotes are required. Additional sections are additive. Before writing “none”, check the transcript. A bare `None` is not a reason. Quotes are VERBATIM; paraphrase in a quote block is fabrication. The transcript remains an attachment, not a truncated page-body replacement. Source Context is optional when this information is already clear in allowed fields or nearby prose; speaker-resolution evidence must still be retained. Cite the meeting's own observed path for transcript passages, with verified line labels, and resolved entity paths for attendees/mentions. Keep hashes and exact locators in attached evidence, originals in File details, and no Sources footer. Follow publication's create-then-link procedure rather than predicting filenames. For a complete paired example, read [meeting and entity example](references/meeting-entity-example.md). ### Phase 6: Claim verification + consistency check (gate for every entity write) Reconcile each proposed assertion against earlier meetings about the same entity, not only reruns of this source. Use shared publication guidance to distinguish reuse, additional support, dated reaffirmation, updated plan, contradiction and explicit correction. Similar titles alone never authorize supersession. | Claim type | Verification bar | |---|---| | Role/relationship change | Explicit transcript passage, not enthusiasm or implication | | Ownership/attribution | Explicit ownership language or corroborating evidence; saying a name is not owning it | | New proper noun | Search Knowledge for canonical spelling, then public sources if appropriate; otherwise label unverified spelling | | Major life/deal event | Verbatim support; inspect existing evidence and chronology | Also check existing Knowledge for contradictions and logical plausibility. Surprising claims deserve extra scrutiny, not a story that rationalizes them. Preserve conflicting accounts with both citations; neither older synthesis nor newer transcript automatically wins. An explicit correction is distinct from a changed plan or independently conflicting source. **Downgrade:** an inference or uncertain identity stays labeled on the meeting page; it must not enter entity State, metadata, timelines or relationships as fact. Failed verification must not fan out. After this gate, create/reuse separately useful claims with `knowledge_create_claim`: decisions, commitments, material facts/beliefs and dated events. Use `assertion_origin: source` for transcript statements and name the speaker in attribution, including unresolved labels. Preserve conditions such as “subject to review”. Use resolved entity IDs and a meeting dependency, plus exact transcript evidence. Known owner and due date belong on a commitment; omit unknown ones. Relative dates require sufficient source date/timezone context. No task or reminder is created. ### Phase 7: Attendee enrichment (MANDATORY) For EACH actual attendee, run `knowledge-entities`: 1. Search and read the existing person page, or create a sourced one if identity is sufficient. Participation and actual contribution are meaningful context; do not create a content-free stub or fabricate a profile for UNKNOWN. 2. Add a one-line dated “Attended {meeting}” Timeline entry, newest first, with a link back and citation, in the `knowledge-entities` Timeline format. 3. Create/reuse an evidenced attendance relationship via `knowledge_create_relationship`; inspect the current label/direction first. In the shipped ontology, `attended` runs **participant → meeting**, unlike gbrain's meeting auto-link direction. Knowledge has no auto-link post-hook. Back-link known people mentioned or speaking in the transcript too, not just attendees, but only supported identifications. Unknown speaker attribution can remain flagged without a fabricated entity. ### Phase 8: Timeline merge (MANDATORY) For each relevant company, project or concept discussed, find its page (or create one under `knowledge-entities`), add a Timeline entry and back-link to the meeting. Link back from the meeting to each entity. The same event belongs on ALL resolved relevant entities' timelines. Represent it as one event claim with `occurred_at`, all relevant `entities`, a meeting `dependencies` reference and evidence. That makes it discoverable from every entity's `knowledge_connections({id, direction: "claims"})`. Chronicle projects these structured dates into a queryable timeline. Set source_page and source_event=true on the single claim representing the meeting itself; its occurred_at must match the meeting field at source precision. Other events discussed retain their own dates and are not source_event. Retain dated human-readable timeline links on profiles. Different companies' feedback gets its own context, not a session-wide blob. Create additional relationships only when evidenced and supported by ontology labels/endpoint types. Preserve original thinking worth capturing in the user's exact words; label analysis separately. No unavailable signal-detector or background enrichment workflow is implied. ### Phase 9: Sync / retrieval No sync command is needed: published saves are immediately readable and search reconciles its own index. Check receipts and fresh retrieval. A pending index/Git checkpoint is not permission to create another copy. ## Verify before declaring done (HARD GATE) Run every check for every meeting, including quick logistics meetings. - **V1 — Required sections have substance.** Summary has concrete outcomes; Decisions, Actions and Quotes have content or a verified reason for absence. Independently useful assertions were created/reused, not left only in prose. - **V2 — Every entity link has a page AND timeline backlink.** Read each linked page; verify the source backlink, dated context, connected event claims and evidenced attendance edges where applicable. Inspect all changed records, compact metadata and relevant evidence (`includeEvidence: true`), following pagination. No guessed paths or empty stubs. Use `knowledge_verify` with independently planned `expectedLinks` for both authored directions and expected event associations/attendance endpoints. Discover graph context or investigate failures with compact connections and record links; do not repeat successful structural checks merely to obtain larger responses. Automatic reverse navigation does not prove an authored contextual link. - **V2b — State reads as current.** No State topic or paragraph on an updated page is named for a date, meeting or source; a changed fact was edited into its existing topic. - **V3 — Speaker map resolved or flagged.** No raw label silently masquerades as a known person. Every named speaker has an evidence basis/confidence. - **Structured readback.** Known commitment owners are resolved page IDs, not just names in prose. Supported deadlines and event dates retain source precision. Attendance edges have `source_page`; qualitative speaker-attribution confidence carries its basis and retained evidence in `attribution_confidence`. Unknown values stay absent. Meeting-event associations use `source_event` only for the meeting itself. Use aggregate expectations for exact owners/dates/source pages, associations and endpoints. Read compact metadata and explicitly requested evidence to judge attribution confidence and its basis; structural success does not replace that semantic review. - **V4 — Every quote grounded VERBATIM.** Locate each quoted span in retained transcript text. Read the exact lines, not just `evidence.verified`. Restore spoken phrasing or choose a supported contiguous span if it fails; never keep paraphrases in blockquotes. A summary cannot verify a transcript quote. - **V5 — Fabricated-attendee sanity checks.** Names absent from transcript and roster evidence, or email domains conflicting with known identity, require inspection. Do not auto-rename or fill addresses. Repair affected outputs within scope, then recheck. Uncertain identities stay explicitly uncertain. - **V6 — Sequence verify (order, not substance).** Extract narrated event atoms (before/during/after, place, people). Check claimed sequence against timestamps and source context independently of thematic heading order. A phase inversion or incompatible locations/times is a hard contradiction; never invent transit to connect them. Inspect dated connected claims for corroboration where useful. Use knowledge_chronicle_day with the meeting date and known timezone, and knowledge_timeline with relevant entity IDs. Verify source-event grouping and world dates independently from recorded_at. A missing event is a signal to investigate, not proof someone was absent. Fix contradictions and rerun, or record an explicit user waiver as acknowledged, NOT resolved. Genuinely ambiguous order stays flagged. Check the work log against receipts and records; verify exact lookup and a characteristic keyword. On repeat/interruption reconcile sources, existing claims and relationships before new writes, using the publication retry rules. Fix → re-check until checks pass or clearly report partial/not fully ingested. A user sequence waiver applies only to V6, not other failed checks. ## Sensitive meetings For legal or deeply personal content, keep the page factual and restrain biographical propagation/back-links. Respect the user's requested scope; clarify when needed before spreading sensitive content into other profiles. ## Output Format Report per real meeting: page ID/path, retained sources, attendees enriched, entities updated/reused, claims/actions/relationships captured, verification passed or specific failures, sequence PASS or explicit WAIVED with the remaining contradiction. List withheld claims, unresolved identities and missing work. Never say fully ingested before the checklist passes; a partial report is honest progress, not completion. No multi-record job or rollback guarantee is claimed.