feat: M2 AI tips — LiteLLM gateway, context assembler, end-to-end generation pipeline

Issues closed: #86, #87, #88, #89, #90, #91, #79, #80, #82

infra:
- docker-compose `ai` profile: Ollama + LiteLLM services
- infra/litellm/litellm_config.yaml: tip-generator / embedder / judge aliases
- .env.example: LITELLM_URL, LITELLM_MASTER_KEY, OLLAMA_URL

ml/serving:
- POST /generate: calls LiteLLM tip-generator alias, returns TipCandidate[]
- JSON retry loop (2 retries with correction prompt on malformed response)
- _parse_llm_json strips markdown fences

ml/features:
- context.py: build_context() assembles user signals → PromptContext
  (sorts overdue/high-priority tasks first for LLM prompt quality)

shared-types:
- TipKind, TipSource, TipCandidate types
- Tip gains kind + rationale fields

services/api:
- recommender: 3-stage pipeline (assemble → score → serve)
  Stage 1: Todoist tasks + LLM candidates fetched in parallel
  Stage 2: egreedy bandit scores merged candidate pool
  Stage 3: serve + log with prompt_version, llm_model, tip_kind
- tip_scores: prompt_version, llm_model, tip_kind columns + migrations
- config: LITELLM_URL added
- integrations: surface token_status in /integrations response

tests:
- ml/serving/tests/test_generate.py: 13 tests (retry, 502/503, fence variants)
- ml/features/test_context.py: 9 tests (sorting, edge cases)
- services/api recommender.unit.test.ts: 16 pure-function tests (inferReward, dueAgeDays)
- services/api recommender.test.ts: 4 integration tests (tip_scores columns, LLM fallback)
- shared-types: TipCandidate, rationale, full TipFeedback action set

docs:
- ADR-0008: LiteLLM AI gateway decision
- overview.md: M2 pipeline description updated
- ml/README.md: serving + features roles updated

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-04-17 14:09:02 +00:00
parent 85367aeaa0
commit ffdf70733f
22 changed files with 1017 additions and 45 deletions

View File

@@ -26,12 +26,16 @@ from collections import deque
from pathlib import Path
from typing import Optional, Deque
import httpx
import numpy as np
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
app = FastAPI(title="oO ML Serving", version="1.0.0")
LITELLM_URL = os.getenv("LITELLM_URL", "http://localhost:4000")
LITELLM_MASTER_KEY = os.getenv("LITELLM_MASTER_KEY", "sk-oo-dev")
STATE_DIR = Path(os.getenv("STATE_DIR", "/tmp/oo-bandit-state"))
STATE_DIR.mkdir(parents=True, exist_ok=True)
@@ -166,6 +170,56 @@ class RewardResponse(BaseModel):
ok: bool
class PromptContext(BaseModel):
tasks: list[dict] = []
hour_of_day: int = 12
day_of_week: int = 0
extra: dict = {}
class GenerateRequest(BaseModel):
user_id: str
context: PromptContext = PromptContext()
n: int = 3
class TipCandidate(BaseModel):
id: str
content: str
source: str = "llm"
rationale: Optional[str] = None
class GenerateResponse(BaseModel):
candidates: list[TipCandidate]
model: str
prompt_tokens: int = 0
completion_tokens: int = 0
_GENERATE_SYSTEM = (
"You are a personal productivity coach. "
"Given the user's current context, generate actionable, specific tips. "
"Respond ONLY with a JSON array of objects, each with keys: "
'"id" (short slug), "content" (the tip, ≤2 sentences), "rationale" (why now, ≤1 sentence). '
"No markdown, no prose outside the JSON array."
)
def _build_prompt(ctx: PromptContext, n: int) -> str:
lines = [f"Time: {ctx.hour_of_day:02d}:00, day_of_week={ctx.day_of_week}"]
if ctx.tasks:
overdue = [t for t in ctx.tasks if t.get("is_overdue")]
lines.append(f"Tasks: {len(ctx.tasks)} total, {len(overdue)} overdue")
for t in ctx.tasks[:5]:
due = t.get("due_date", "no due date")
lines.append(f" - [{t.get('priority','?')}] {t.get('content','?')} (due: {due})")
for k, v in ctx.extra.items():
lines.append(f"{k}: {v}")
lines.append(f"\nGenerate {n} tips as a JSON array.")
return "\n".join(lines)
# ── Endpoints ──────────────────────────────────────────────────────────────
@app.get("/health")
@@ -173,6 +227,97 @@ def health():
return {"ok": True}
_RETRY_SUFFIX = (
"\n\nYour previous response was not valid JSON. "
"Reply ONLY with the JSON array — no prose, no markdown fences."
)
_MAX_GENERATE_RETRIES = 2
def _parse_llm_json(raw: str) -> list[dict]:
"""Strip markdown fences and parse JSON array. Raises ValueError on failure."""
text = raw.strip()
if text.startswith("```"):
parts = text.split("```")
text = parts[1] if len(parts) > 1 else text
if text.startswith("json"):
text = text[4:]
return json.loads(text)
@app.post("/generate", response_model=GenerateResponse)
async def generate(req: GenerateRequest) -> GenerateResponse:
"""Generate tip candidates via LiteLLM → tip-generator alias.
Retries up to _MAX_GENERATE_RETRIES times on malformed JSON, appending
a correction hint to the conversation so the model can self-correct.
"""
prompt = _build_prompt(req.context, req.n)
messages: list[dict] = [
{"role": "system", "content": _GENERATE_SYSTEM},
{"role": "user", "content": prompt},
]
headers = {"Authorization": f"Bearer {LITELLM_MASTER_KEY}"}
last_parse_error: str = ""
last_raw: str = ""
total_usage: dict = {"prompt_tokens": 0, "completion_tokens": 0}
model_used = "tip-generator"
async with httpx.AsyncClient(timeout=30.0) as client:
for attempt in range(1 + _MAX_GENERATE_RETRIES):
payload = {"model": "tip-generator", "messages": messages, "temperature": 0.7}
try:
resp = await client.post(
f"{LITELLM_URL}/chat/completions",
json=payload,
headers=headers,
)
resp.raise_for_status()
except httpx.HTTPStatusError as e:
raise HTTPException(status_code=502, detail=f"LiteLLM error: {e.response.text}")
except httpx.RequestError as e:
raise HTTPException(status_code=503, detail=f"LiteLLM unreachable: {e}")
data = resp.json()
usage = data.get("usage", {})
total_usage["prompt_tokens"] += usage.get("prompt_tokens", 0)
total_usage["completion_tokens"] += usage.get("completion_tokens", 0)
model_used = data.get("model", "tip-generator")
last_raw = data["choices"][0]["message"]["content"]
try:
items = _parse_llm_json(last_raw)
break
except (json.JSONDecodeError, ValueError) as e:
last_parse_error = str(e)
# Feed the bad reply back so the model can self-correct
messages.append({"role": "assistant", "content": last_raw})
messages.append({"role": "user", "content": _RETRY_SUFFIX})
else:
raise HTTPException(
status_code=502,
detail=f"LLM returned invalid JSON after {_MAX_GENERATE_RETRIES} retries: "
f"{last_parse_error}\n{last_raw[:200]}",
)
candidates = [
TipCandidate(
id=item.get("id", f"tip-{i}"),
content=item.get("content", ""),
rationale=item.get("rationale"),
)
for i, item in enumerate(items)
]
return GenerateResponse(
candidates=candidates,
model=model_used,
prompt_tokens=total_usage["prompt_tokens"],
completion_tokens=total_usage["completion_tokens"],
)
@app.post("/score", response_model=ScoreResponse)
def score(req: ScoreRequest) -> ScoreResponse:
if not req.candidates: