"""matchmaking-v2 settings — lean, on-demand. No corpus DB, no nightly scrape.""" from __future__ import annotations from functools import lru_cache import os from pydantic_settings import BaseSettings class Settings(BaseSettings): model_config = {"env_file": ".env", "env_file_encoding": "utf-8", "extra": "ignore"} ENV: str = "development" LOG_LEVEL: str = "INFO" SERVICE_NAME: str = "matchmaking-service" # MUST stay "matchmaking-service" — orchestrator routes by card name API_PREFIX: str = "/api/v1" API_HOST: str = "0.0.0.0" API_PORT: int = 8000 CORS_ORIGINS: str = "*" # Public URL the orchestrator discovers this agent at (one of AGENT_URLS). AGENT_URL: str = "http://localhost:8006" # A2A (agent-to-agent) bearer auth — the orchestrator presents one of these. A2A_ALLOWED_KEYS: str = "dev-a2a-key" # Orchestrator comms (Redis Streams). Optional in dev — the worker no-ops if absent. ORCHESTRATOR_REDIS_URL: str | None = None REDIS_URL: str = "redis://localhost:6379/0" # Inputs / engine APIFY_TOKEN: str | None = None OPENAI_API_KEY: str | None = None # LLM gateway (opencode.ai/zen — OpenAI-compatible; reuse interview-service's DSPY key). # One `openai` client (base_url override) serves the Opus curator + embeddings. DSPY_API_BASE: str | None = None DSPY_API_KEY: str | None = None CURATE_MODEL: str = "claude-opus-4-8" # the senior recruiter (Opus, latest) — Opus's call is final SUGGEST_MODEL: str = "claude-haiku-4-5" # the bubbling engine — fast + cheap (latency-sensitive UI) EMBED_MODEL: str = "text-embedding-3-small" # the vibe-engineer (direct OpenAI — gateway has no embeddings route) ENGINE_LLM_ENABLED: bool = True # off / key absent → white-box sift + templated cards (safety net) SIFT_TOP_K: int = 22 # candidate pool the curator reads — widened from 18 for # recall, trimmed from 28 to keep the Opus payload lean MATCH_FLOOR: int = 50 # HARD filter — never surface a match scored below this # Presentation calibration of the HEADLINE score (breakdown stays raw evidence). Transparent + # monotonic + floor-anchored: a genuinely-strong match presents in the 90s; weak NEVER becomes strong. CALIBRATION_ENABLED: bool = True CALIBRATION_GAMMA: float = 1.3 # 50→50, 70→74, 90→94, 95→97 (see rubric.calibrate) # Result cache (dev cost-saver). Keyed by (actor, exact input) → saved job set on disk. # off = always live, never save (PRODUCTION default) # readwrite = replay if saved else live + save (DEV: one live sweep, then $0 replays) # read = replay only, error if missing (offline / deterministic) # refresh = always live + overwrite the saved set APIFY_CACHE_MODE: str = "off" APIFY_CACHE_DIR: str = ".apify_cache" # Active boards — "Balanced" India stack: all three filter to the exact city at the board # (hyper-relevant) and surface offsite apply links. Naukri=blackfalcondata feed, # Foundit=Monster India, LinkedIn=geoId/city. Others (ats/indeed/naukri_v1) stay registered, off. BOARDS_ENABLED: str = "naukri,foundit,linkedin,timesjobs,workindia" # signal-weighted; indeed dropped (0% skills, redundant) BOARD_TIMEOUT_S: float = 55.0 # per-board cap — a slow/hung board is skipped, the search stays fast # Warm pool ("the shelf") — bank surplus fetched jobs per (user, query) so we don't re-pay Apify for # a full sweep on every search. Serve from the pool; refill only when it dips. Reorder EARLY (top up # in the background) and hard-floor a blocking fetch so the pool never runs thin. POOL_ENABLED: bool = True POOL_REORDER_AT: int = 100 # fresh-unseen < this → background top-up (while still serving) POOL_SAFETY_FLOOR: int = 90 # fresh-unseen < this → BLOCK + fetch before serving (hard floor) # 100/90 keeps a DEEP pool (richer decks) at the cost of more # frequent refills — we pay for volume, deliberately. POOL_TTL_HOURS: int = 72 # banked jobs older than this are stale (likely filled) → dropped POOL_MAX_PER_QUERY: int = 400 # cap the bank per (user, query) — STORAGE cap (kills DB swell) + load cap # Exhaustion guard — a refill that adds fewer than this many NEW jobs means the query is tapped out; # we then RELAX the floor (serve what's pooled, stop block-fetching duplicates) until it expires. POOL_MIN_NEW_PER_REFILL: int = 8 POOL_EXHAUSTION_TTL_HOURS: int = 12 # re-try a tapped query after this (boards post new jobs ~daily) # Per-board fetch budget (the #1 Apify cost lever). Demo-sized to conserve credits. # Signal-weighted budgets → ~150 jobs/sweep, tilted to the high-signal boards. NAUKRI_MAX_JOBS: int = 30 # blackfalcondata maxResults; fetchDetails=True for offsite apply links FOUNDIT_MAX_JOBS: int = 40 LINKEDIN_MAX_JOBS: int = 30 TIMESJOBS_MAX_JOBS: int = 35 # shahidirfan — the richest board (skills+salary+experience) # ATS lane (off by default) — company-seeded global/remote "dream companies". ATS_MAX_PER_COMPANY: int = 4 ATS_COMPANIES: str = "stripe,databricks,gitlab,figma,ramp,notion,razorpay,zerodha" INDEED_MAX_JOBS: int = 30 WELLFOUND_MAX_JOBS: int = 15 # blackfalcondata/wellfound-scraper — registered, OFF (US-heavy) WORKINDIA_MAX_JOBS: int = 15 # shahidirfan/workindia-jobs-scraper — India blue/grey-collar (niche, low yield) # QScore (competence proxy) — consumed, not computed. Dashboard Momentum/QX reads it via REST. QSCORE_BASE_URL: str = "http://localhost:8004" QSCORE_AUTH_TOKEN: str | None = None QSCORE_ORG_ID: str = "growqr" # user-service — dashboard Day-streak reads metadata.current_streak via GET /api/state/{clerk_id} USER_SERVICE_BASE_URL: str = "http://localhost:8003" # dashboard-service — Q-Score trend series via GET /api/qscore-history/{clerk_id} DASHBOARD_SERVICE_BASE_URL: str = "http://localhost:8005" A2A_OUTBOUND_KEY: str = "dev-a2a-key" # Bearer token presented to sibling services # Storage for feedback/labels (NOT the old corpus). Optional until the learning loop lands. DATABASE_URL: str | None = None @property def cors_origins_list(self) -> list[str]: if self.CORS_ORIGINS.strip() == "*": return ["*"] return [o.strip() for o in self.CORS_ORIGINS.split(",") if o.strip()] @lru_cache def get_settings() -> Settings: if os.environ.get("PYTEST_CURRENT_TEST"): return Settings(_env_file=None) return Settings()