Call → Knowledge Base pipeline — design (2026-07-02)

Consolidated design for turning call transcripts (Krisp) into a linked, navigable trip2g knowledge base. Captures the decisions from the 2026-07-02 session so nothing is lost. Sample vault: ~/krisp-run/vault-out/. Article: PR #102. Pipeline evidence: ~/krisp-run/2026-07-02_boundary_and_extraction.md.

Goal

Extraction INTO a knowledge base (linked concept graph + navigable notes), NOT summary + chat-with-a-bot. The KB is durable, auditable, and compounds over time.

Note taxonomy

  • transcript (transcripts/YYYY-MM-DD_slug.md) — raw Krisp transcript, verbatim, never edited. The source. type: transcript, call_id (ulid), created_at from the UUIDv7 id.
  • call / разбор (calls/YYYY-MM-DD_slug.md) — the analysis. First paragraph = summary (magazine card). Links to its transcript (frontmatter transcript: + inline). title_source: inferred (Krisp's calendar title is unreliable — see below), speakers, needs_review.
  • daily (daily/YYYY-MM-DD.md) — entry point of the pipeline. Action-item checkboxes at the TOP (today; "сделаю завтра" → next day's daily), log below, no day summary. Prev/next-day links.
  • log (log/topic.md) — append-only journal of one recurring thought; each addition under a ### [[YYYY-MM-DD]] heading; never rewrite, only append.
  • concept / term (concepts/slug.md) — the glossary graph. [[wikilinks]], aliases, mentions (unquoted int).

See project_secondbrain_note_taxonomy for the user's methodology this models.

Provenance (invariant)

Every fact traces to the source. We have topic boundaries, so: quote the raw segment verbatim as a blockquote with a link + timerange to the transcript. The quote IS the evidence (zero hallucination in the evidence); the LLM разбор/summary sits above it. Applies to разбор topic sections and concept "Упоминания". Approximate segment-start time is fine; never fabricate exact times.

Pipeline = event-driven fleet cascade

Standalone ~/krisp-run/build_vault.py is the "run it yourself without fleet" path. The trip2g-native path is a chain of fleet role-notes wired by change-webhooks:

  1. Ingest — python executor: code role pulls raw calls (Krisp API) → writes transcripts/*.md. The ONLY source-specific stage.
  2. change-webhook on transcripts/** → topic segmentation (mini-V2 coarse boundaries) writes segments.
  3. segments written → extraction/distribution (растаскивание) writes concept notes, appends logs, writes daily action-checkboxes, and the разбор note linking back to the transcript.

Drop a raw transcript → the KB assembles itself. Dogfoods change-webhooks + exec:code. See project_kb_construction_engine, project_agent_runtime.

Authoritative metadata (Krisp is unreliable)

  • Time: decode the UUIDv7/ULID id (int(id[:8],16)<<16 ms), NOT the local clock (user's clock is wrong) and NOT Krisp's calendar match (it picks the wrong event by time — documented-fragile, purely time-based). Drives created_at (RFC3339, the key the magazine sorts on) + daily bucketing.
  • Title: prefer LLM-inferred topic title (title_source: inferred); Krisp's calendar title is actively wrong for this user.
  • Speakers: Krisp API participant list; fall back to inference; needs_review: true.

Entity resolution (the hard part)

  • Specificity gate: concept notes only for proper nouns or terms that recur with a stable meaning. Lowercase generic descriptions ("мероприятие про агентов") stay as prose or resolve to a real named entity — they do NOT become notes.
  • Referent-aware reconcile (not name-only): on a surface-name match, compare the new mention's context to the existing note's definition. Same referent → append. Different referent → mint a NEW note with a disambiguating title (... (HR, 2026-06-19) vs ... (маркетинг, 2026-06-27)). Never silently merge. This is the fix for the observed over-merge (Claude/Codex → ChatGPT).
  • Stable slug + alias table so links survive re-titling/disambiguation.
  • needs_review on low-confidence merges. Identity = note id/content, not the surface title (Zettelkasten).
  • Drift within one note is fine (evergreen append) as long as it's the same referent; the bug is false merges, fixed by the above.

Calls magazine (home), terms magazine, by-day (daily notes), _header/_footer. Magazine mode = content: [magazine] + magazine_include_files glob; date key is created_at (RFC3339), not date. See docs/dev/2026-07-02_call_magazine_layout.md.

Maintenance fleet roles (future)

  • link-healer — reads note warnings (GraphQL AdminNoteWarnings / trip2g lint), fixes broken [[wikilinks]] (create missing stub or repoint) via patch_note; needs_review when ambiguous.
  • concept-gardener — cron role that finds notes conflating two referents (contradictory definition / diverging mentions) and splits them.

Cost

~$0.13–0.18 per 50-min call, all gpt-5.4-mini. Boundaries mini-V2 coarse. Re-runs reuse cached extractions (work-vault/) at ~$0.

Status

Sample vault at ~/krisp-run/vault-out/ (8 calls); served locally via memcli up. Article PR #102. Reconcile hardening (specificity gate + referent-aware + disambiguation) and the maintenance roles are designed here but not yet implemented.