Files
matchmaking-v2/app/engine/sift.py
raulgupta b1e9cdd182 Engine v2: embeddings sift + Opus curator + Postgres feed persistence
- A1: fire the hard filter (Apify is precise); embedding vibe-ranker (embed.py)
  blends with the floor-free white-box (rank.py) to sift ~90 -> top 18
- A2: Opus curator (curate.py) reads the 18 -> honest <=count + Gemini-voiced
  report cards; graceful fallback to the white-box + templated cards
- LLM via opencode.ai/zen gateway (llm.py): Opus chat + direct-OpenAI embeddings
- Run LLM work off the event loop (asyncio.to_thread) so the Redis response publishes
- C1: dedicated Postgres (app/db/) persists the per-user feed; get_scout_feed
  replays it so matches survive navigation/refresh
- match contract + report-card fields (schema.py); skills.py; tests/
2026-06-19 15:51:16 +05:30

44 lines
1.9 KiB
Python

"""The sift — combine the two cheap rankers (white-box ‖ embedding vibe) by weighted-average
RANK fusion and return the **top-K candidate pool** for the curator (Opus).
This is NOT the final cut — it only decides which ~18 jobs the expensive brain reads (retrieve cheap,
rerank expensive). Each returned job already carries its white-box `match` block (the templated
fallback the curator's output later replaces).
"""
from __future__ import annotations
from app.config import get_settings
from app.engine import embed as _embed
from app.engine import rank as _rank
W_WHITEBOX, W_VIBE = 0.55, 0.45 # white-box (checklist) leads slightly; vibe (semantics) sharpens
def _ranks(score_by_id: dict[str, float]) -> dict[str, float]:
"""id → normalized rank in [0,1] (1 = best, 0 = worst). Position-based, so scale-free."""
order = sorted(score_by_id, key=lambda i: -score_by_id[i])
n = len(order)
return {i: (1.0 - pos / (n - 1)) if n > 1 else 1.0 for pos, i in enumerate(order)}
def sift(prefs: dict, ctx: dict | None, jobs: list[dict], k: int | None = None):
"""Returns (top_k_jobs, debug). top_k_jobs are ranked best-first; each keeps its white-box match."""
s = get_settings()
k = k or s.SIFT_TOP_K
ranked = _rank.rank(prefs, ctx, jobs)["opportunities"] # white-box scores ALL (no early cut)
if not ranked:
return [], {"mode": "empty", "scored": 0, "k": 0}
wb = {j["id"]: j["matchScore"] / 100.0 for j in ranked}
vibe = _embed.vibe_scores(prefs, ctx, ranked) # id → cosine, or None
if vibe:
wr, vr = _ranks(wb), _ranks(vibe)
fused = {i: W_WHITEBOX * wr[i] + W_VIBE * vr.get(i, 0.0) for i in wb}
mode = "whitebox+vibe"
else:
fused = _ranks(wb)
mode = "whitebox-only"
top = sorted(ranked, key=lambda j: -fused[j["id"]])[:k]
return top, {"mode": mode, "scored": len(ranked), "k": len(top)}