alvis
7d4c29e137
Centralizes user-level features (completion_rate_30d, dismiss_rate_30d, mean_dwell_ms_30d, preferred_hour, tip_volume_30d) in a TS registry that owns both definition and SQL aggregation, since the data lives in the TS-owned SQLite tables (tip_views/tip_feedback). Lazy TTL refresh keeps recommend latency bounded; values persist in user_profile_features (KV). ml/serving accepts profile_features on /score + /generate but does not yet consume them — extending the bandit feature vector changes D and resets every user's learned state, so that's a deliberate phase-B step. Includes ml/features/profile_schema.py as a contract mirror with a sync test that diffs name sets against registry.ts. ADR-0011 records the data-locality reasoning (registry in TS, not Python as the issue originally suggested). Phase B (deferred): event-driven incremental updates, bandit consumption with state migration, admin per-user profile page, staleness alerts. Refs #81. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
33 lines
2.0 KiB
Markdown
33 lines
2.0 KiB
Markdown
# ml/
|
||
|
||
Python. Owns models, features, training, online scoring.
|
||
|
||
| Dir | Role | Phase |
|
||
|---|---|---|
|
||
| `serving/` | FastAPI online scorer (`/score`, `/generate`) + LiteLLM gateway + prompt registry (`prompts.py`), called by `recommender` | 1–2 |
|
||
| `features/` | context assembler (`context.py`): signals → `PromptContext`; Feast adapter later | 2 |
|
||
| `pipelines/` | batch feature + training DAGs (Prefect/Airflow) | 4 |
|
||
| `registry/` | MLflow-backed model registry integration | 4 |
|
||
| `experiments/` | A/B assignment + multi-armed bandit policies | 4 |
|
||
| `notebooks/` | research; never imported by production code | — |
|
||
|
||
## Principles
|
||
|
||
- Every model has a **model card** in `registry/` describing inputs, offline metrics, fairness checks, and rollout history.
|
||
- Online inference must be stateless and < 50ms p99.
|
||
- Training reads from the offline feature store; serving reads from the online feature store; definitions are shared (no train/serve skew).
|
||
- Shadow deploys before any policy change that affects real users.
|
||
|
||
## Profile-feature contract
|
||
|
||
User-level features (completion rate, preferred hour, tip volume…) are computed
|
||
by the TypeScript recommender and shipped to ml/serving on every `/score` and
|
||
`/generate` call as `profile_features: dict | None`. The Python mirror in
|
||
`features/profile_schema.py` documents the available names + dtypes — keep it
|
||
in sync with `services/api/src/profile/registry.ts` (a CI-style test asserts
|
||
the name sets match). See ADR-0011.
|
||
|
||
## Prompt registry
|
||
|
||
`serving/prompts.py` keys tip-generation prompts by stable version string. Adding a new variant means adding an entry — no caller changes. Selection precedence: `POST /generate` body's `prompt_version` field → env `DEFAULT_PROMPT_VERSION` → `"v1"`. The TypeScript recommender drives selection via `TIP_PROMPT_VERSION` (single value or comma-separated rotation); the version actually used flows back in the response and is persisted to `tip_scores.prompt_version` so the admin reward-analytics dashboard can bucket reactions per variant.
|