Krisp case — transcripts → segments → wiki, on the fleet rails
Status: internal / unreleased. A worked use case that runs the fleet agent runtime on real-shaped data: a raw call/voice transcript becomes a topic-segmented map and then a knowledge note with [[WikiLinks]] into the vault graph.
Pipeline (2 chained roles)
transcripts/<id>.md raw transcript (deterministic ingest — source of truth, never the LLM)
│ edit fires change_webhook (role: transcript-segment, trigger transcripts/**)
▼
segments/<id>.md step 1: topic-boundary map (Minto headlines, [MM:SS–MM:SS])
│ write fires change_webhook (role: transcript-wiki, trigger segments/**)
▼
wiki/<id>.md step 2: knowledge note + [[WikiLinks]] (resolves into the trip2g graph)
Chain works because both roles set max_depth: 3 (segment's depth-1 write is allowed to fire the wiki role); the chain terminates because nothing triggers on wiki/**. Both roles use for_each: changed_files (the bodies reference change_file).
Design decision
Raw-source-saving is deterministic ingest, not the LLM: the raw transcript is preserved verbatim (auditable, exact, re-processable). The LLM only processes (segment, extract). Re-run a role on the preserved raw note when the prompt/model improves — no re-fetch from Krisp.
Files
roles/transcript-segment.md— step 1 (ported fromtrip2g_agent_queue/krisp/prompts/segments.md).roles/transcript-wiki.md— step 2 (ported/adapted fromidea_note.md; reads the source transcript viaread_notefor detail).transcripts/sample-call.md— synthetic 3-min planning call (3 topic shifts) for cheap testing.segments/sample-call.md,wiki/sample-call.md— captured outputs from a real run.
Validated run (2026-06-29)
Both steps run end-to-end on the fleet's own engine (agentruntime) via the cmd/fleet --once offline harness, model qwen/qwen3-14b over OpenRouter:
- step 1:
completed, ~7.6K tokens, 2 steps — 3 correct topic segments, Minto headlines. - step 2:
completed, ~11.2K tokens, 2 steps — knowledge note + 7[[WikiLinks]](Блог, Контекстная реклама, Сеньор-разработчик, …).
Cost ≈ fractions of a cent total. (cmd/fleet --onceruns a single role-note offline — same engine as the daemon, no trip2g connection — seedocs/dev/fleet_run.mdfor the full fleet run.)
To run on the real fleet
Push roles/* and transcripts/sample-call.md into a trip2g instance (roles/ is the fleet's --agents-folder), start cmd/fleet (see docs/dev/fleet_run.md) with --default-model qwen/qwen3-14b --llm-base-url https://openrouter.ai/api/v1, then edit the transcript → the chain runs and writes segments/* + wiki/*.
Next
- Port
trip2g_agent_queue/krisp/extract_krisp_transcripts→ write real Krisp transcripts astranscripts/*(deterministic ingest). - Quality compare Qwen-14B vs gpt-5.4-mini on the same input.