"""Embedding + trainable head: the "few labelled examples" path that trains in seconds. Encode examples + label rubrics (as extra pseudo-examples) with a multilingual sentence encoder, fit a logistic-regression head, predict calibrated-ish probabilities. Requires the `embed` extra: uv sync --extra embed Default model is multilingual and small; swap via --embed-model. """ from __future__ import annotations import hashlib import json from ..spec import LabelSet, Post, Result DEFAULT_MODEL = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2" class EmbedBackend: name = "embed" def __init__(self, model: str = DEFAULT_MODEL): from sentence_transformers import SentenceTransformer # lazy: heavy import self.encoder = SentenceTransformer(model) self._heads: dict[str, object] = {} def _key(self, ls: LabelSet) -> str: return hashlib.sha1(json.dumps(ls.model_dump(), sort_keys=True).encode()).hexdigest() def _head(self, ls: LabelSet): k = self._key(ls) if k not in self._heads: from sklearn.linear_model import LogisticRegression texts = [e.text for e in ls.examples] + [f"{l.key}: {l.rubric}" for l in ls.labels] ys = [e.label for e in ls.examples] + [l.key for l in ls.labels] X = self.encoder.encode(texts, normalize_embeddings=True) self._heads[k] = LogisticRegression(max_iter=1000, C=4.0).fit(X, ys) return self._heads[k] async def classify(self, ls: LabelSet, posts: list[Post]) -> dict[str, Result]: head = self._head(ls) X = self.encoder.encode([p.text for p in posts], normalize_embeddings=True) out: dict[str, Result] = {} for p, row in zip(posts, head.predict_proba(X)): probs = {str(c): float(v) for c, v in zip(head.classes_, row)} best = max(probs, key=probs.get) out[p.id] = Result(label=best, probabilities=probs, confidence=probs[best], backend=self.name) return out