Semih b635b529e2 feat(phase6e): eval framework + prompt editor + daily brief + retention + bundle mode ...

Schema:
- eval_sets (promptTag, cases JSON: [{id, timeline, expected: {...}}])
- eval_runs (per-execution scoring: passedSchema/Severity/Rubric, cost, tokens, results)

Prompts:
- 2 new seed templates: upgrade_hesitation (flash), pattern_bundle (pro)
- pickPromptTag() routes 'upgrade_hesitation' tag to upgrade_hesitation prompt
- Editor UI at /insights/settings/prompts/[id]: edit system/user/schema/tier/temp,
  publishPromptVersion() creates new version + deactivates old (CRUD with auto bump)
- setPromptActive() to toggle versions

Daily Brief (/insights/brief):
- Last 24h: sessions processed, insights produced (with severity breakdown),
  cost, cache hit rate; today's top 5 priorities; week stats (shipped/validated/regressed)

Retention cron (04:15 UTC daily):
- Delete sessions_meta + session_custom_events older than 90d (unless referenced by
  active insight)
- Delete compressed_sessions rows + MinIO timeline blobs older than 180d
- raw_metadata 30d (currently no-op; we don't persist raw metadata to MinIO)

Eval framework:
- apps/worker/src/jobs/eval-run.ts: runs cases against current/specified prompt version,
  scores schema_pass + severity_match + rubric_substring; stores EvalRun
- apps/worker BullMQ queue handler for 'eval-run' job name
- apps/web installed bullmq; /lib/queue.ts thin Queue accessor
- Web actions: createEvalSet, triggerEvalRun (enqueues job to insight-pipeline queue)
- UI: /insights/settings/eval-sets list, /new create form (paste JSON cases),
  /[id] detail with Run button + recent runs + per-case JSON

Bundle mode (analyze job):
- Pull 3x batch, group by fingerprint
- Groups ≥ INSIGHT_BUNDLE_THRESHOLD (default 3) → use 'pattern_bundle' prompt
- Timeline = primary rep + PATTERN BUNDLE summary block (occurrences, unique users, deltas)
- Insight stores ALL group session IDs as relatedSessionIds; all marked analyzed in one tx
- Cost amortized: 1 LLM call per group

Nav/Cmd+K:
- Inbox header links: Brief, Patterns, Eval sets
- Palette: Daily brief, Eval sets entries

Deferred to backlog: embedding-similarity cross-fingerprint clustering, Telegram brief delivery,
Sase.tr-side data-private audit (separate repo).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

2026-05-14 06:11:48 +00:00

192 KiB

Raw Blame History

View Raw