peptiderepo
c79fbb570e
feat: wave 2 eval harness with lm-as-judge and cli command
...
New eval framework:
- eval/dataset.json: 20 frozen peptide topic eval items
- class-opik-eval-scorer.php: LLM-as-judge using Gemini Flash 1.5
Scores on: scientific_signal, readability, style_adherence (0-10)
- class-opik-eval-runner.php: orchestrator for full eval pass
Loads dataset, runs generation (no publish/images), scores, pushes to Opik
run(int $limit = 0, bool $dry_run = false): array
- class-wp-cli-commands.php: registers wp prautoblogger opik:eval
Flags: --limit=<n> (default 0 = all), --dry-run (score only)
Features:
- Rate limiting: 500ms between items
- Cost tracking: eval judge calls logged to opik_eval_judge stage
- Dry-run mode: test locally without Opik API push
- Experiment naming: prautoblogger-eval-v<VERSION>-<YYYY-MM-DD>
All files <= 300 lines. Version 0.13.0. Ready for PR.
2026-04-24 04:25:41 +00:00
peptiderepo
fbbb387311
feat: Opik LLM observability tracing skeleton (Wave 1) — v0.12.0
...
Wires PRAutoblogger into Comet Opik cloud for per-call tracing.
- Thin PHP REST client (class-opik-client.php, 276 LOC)
- Async fire-and-forget dispatch via wp_schedule_single_event
- Feature flag PRAUTOBLOGGER_OPIK_ENABLED defaults OFF
- One trace per article, 5 spans across content-generator + chief-editor
- Exponential backoff 2s/4s/8s, max 3 retries, drops on failure
- Admin settings page (api-key configured indicator only, never rendered)
- ARCHITECTURE.md + CONVENTIONS.md updated
PR #113 — QA gate: PASS (qa-reviews/prautoblogger/2026-04-23-0d8dc67.md)
2026-04-23 22:36:24 +07:00
