matchmaking-v2/app/engine/curate.py

"""The senior recruiter — Opus curates the sifted ~18 into the honest shortlist and writes the report
cards in the interview-service's Gemini video-analysis voice. **Opus's call is final.**

One gateway call: profile + the ~18 sifted jobs → strict JSON of the KEPT jobs (honest count ≤ the
pool — fewer when warranted, never padded), each carrying the full report-card contract. Returns `None`
on any failure (no key / bad JSON / API error) so the caller falls back to the sift + templated cards.
"""
from __future__ import annotations

import json
import logging

from app.config import get_settings
from app.engine.llm import curate_enabled, gateway_client
from app.engine.schema import Growth, MatchDim, MatchResult

logger = logging.getLogger(__name__)

_SYSTEM = """You are Scout's senior recruiter. You receive a candidate's PROFILE and a pre-sifted list of \
JOBS (already narrowed to the right city and title by cheaper tools). Two jobs:

1. CURATE — keep ONLY the jobs genuinely worth this candidate's time. Be honest: if only 8 of the jobs \
are real matches, return 8. NEVER pad to a number. Drop weak, duplicate, or off-target roles.

2. For each KEPT job, write a REPORT CARD in this voice (modeled on a supportive video-coaching analysis):
   - archetype: a short, characterful label — e.g. "The Stretch Worth Taking", "The Safe Powerhouse", \
"The Skill-Adjacent Pivot". Memorable, never generic.
   - one_line: ONE honest, balanced sentence — name the strength AND the gap. Never pure hype.
   - breakdown: 3-4 dimensions (e.g. Skill match, Experience, Location, Industry), each with an honest \
0-100 score, a level ("Strong" >=72 / "Solid" 52-71 / "Light" <52), and a specific one-line note grounded \
in THIS job's actual skills/title.
   - coach_note: one warm, actionable line — what to lead with, or what to shore up before applying.
   - growth (optional): {"text","from","to"} — the gap to close and the score lift closing it buys.
   - score: honest 0-100 overall fit. fit: "fit" (a current-state match) or "stretch" (a genuine reach).

Honesty rules: scores reflect reality — a partial match is in the 60s-70s, not the 90s. Ground every note \
in the job's real content; invent nothing. Frame as helpful guidance, not a hiring verdict.

NON-TECH roles (sales, marketing, finance, HR, operations, support, supply-chain, legal, admin): the SPINE \
of the match is role/function fit + responsibility overlap + industry + seniority — NOT a skill-tag checklist. \
These postings rarely list skills, and a missing skill list is NOT a negative signal — score on the \
responsibilities and role alignment, never penalize the candidate for the board's empty skill field. A strong \
role+industry+responsibility+seniority match with no listed tags is a genuine ~80 match, not a ~65. Make one \
breakdown dimension "Role fit" (same function / adjacent / different).

Respond with ONLY a JSON object (no prose, no markdown fences):
{"kept":[{"id","score","fit","archetype","one_line","breakdown":[{"name","score","level","note"}],\
"coach_note","growth":{"text","from","to"}}]}
Order "kept" best-first. Use each job's exact "id"."""


def _profile_brief(prefs: dict, ctx: dict | None) -> dict:
    ctx = ctx or {}
    return {
        "target_title": prefs.get("title"),
        "target_roles": prefs.get("role"),
        "target_location": prefs.get("location"),
        "seniority": prefs.get("experience"),
        "years": prefs.get("years"),
        "target_industry": prefs.get("industry"),
        "skills": ctx.get("skills"),
        "current_role": ctx.get("current_role"),
        "stretch_appetite": prefs.get("stretch", "balanced"),
    }


def _job_brief(j: dict) -> dict:
    d = j.get("details") or {}
    return {
        "id": j["id"],
        "title": j.get("title"),
        "company": j.get("organization"),
        "location": f"{j.get('location_city') or ''} {j.get('location_country') or ''}".strip(),
        "role_category": d.get("role_category"),
        "industry": d.get("industry"),
        "seniority": j.get("seniority_level"),
        "pay": j.get("payLabel"),
        "skills": (d.get("skills") or j.get("required_skills") or [])[:12],
        "responsibilities": (d.get("description") or "")[:1200],   # was 500 — Opus scored from a stub
        "prelim_score": j.get("matchScore"),
    }


def _parse_json(text: str) -> dict:
    """Tolerant parse — strip markdown fences / surrounding prose if the model added any."""
    t = (text or "").strip()
    if t.startswith("```"):
        t = t.split("```", 2)[1].removeprefix("json").strip() if "```" in t[3:] else t.strip("`")
    i, j = t.find("{"), t.rfind("}")
    if i != -1 and j != -1:
        t = t[i : j + 1]
    return json.loads(t)


def _to_match(k: dict) -> MatchResult:
    dims = [
        MatchDim(name=str(d["name"]), score=int(d["score"]), level=d["level"], note=d.get("note"))
        for d in (k.get("breakdown") or [])
    ]
    g = k.get("growth")
    growth = Growth(text=g["text"], **{"from": int(g["from"]), "to": int(g["to"])}) if g else None
    return MatchResult(
        score=int(k["score"]),
        fit=k["fit"],
        archetype=k.get("archetype"),
        one_line=k.get("one_line"),
        reason=k.get("one_line") or k.get("archetype") or "Curated by Scout",
        breakdown=dims,
        growth=growth,
        coach_note=k.get("coach_note"),
        proofReady=any(d.level == "Strong" for d in dims),
        factors={},
    )


def curate(prefs: dict, ctx: dict | None, sifted_jobs: list[dict]) -> list[dict] | None:
    """Opus reads the sifted pool → the curated jobs (kept, best-first, with real report cards).
    `None` on any failure so the caller degrades to the sift + templated cards."""
    client = gateway_client()
    if not (curate_enabled() and client and sifted_jobs):
        return None
    s = get_settings()
    payload = {"profile": _profile_brief(prefs, ctx), "jobs": [_job_brief(j) for j in sifted_jobs]}
    try:
        resp = client.chat.completions.create(
            model=s.CURATE_MODEL,
            messages=[
                {"role": "system", "content": _SYSTEM},
                {"role": "user", "content": json.dumps(payload, ensure_ascii=False)},
            ],
            max_tokens=4000,
        )
        data = _parse_json(resp.choices[0].message.content)
    except Exception as e:  # noqa: BLE001 — any failure → fall back to the sift
        logger.warning("curate (Opus) failed, falling back to sift: %s", e)
        return None

    by_id = {j["id"]: j for j in sifted_jobs}
    out: list[dict] = []
    for k in data.get("kept") or []:
        job = by_id.get(k.get("id"))
        if not job:
            continue
        try:
            m = _to_match(k)
        except Exception:  # noqa: BLE001 — skip a malformed card, keep the rest
            continue
        job["match"] = m.as_dict()
        job["matchScore"] = m.score
        out.append(job)
    return out or None