level_classifier_tool.py

from __future__ import annotations

from dataclasses import dataclass
from typing import Dict, List, Tuple, Iterable, Optional, Literal, Callable, Any
import math
import os

# Optional heavy deps are imported lazily when needed
_TOK = None
_MODEL = None
_TORCH = None

Agg = Literal["mean", "max", "topk_mean"]


# --------------------------- Embedding backend ---------------------------

@dataclass
class HFEmbeddingBackend:
    """
    Minimal huggingface transformers encoder for sentence-level embeddings.
    Uses mean pooling over last_hidden_state and L2 normalizes the result.
    """
    model_name: str = "sentence-transformers/all-MiniLM-L6-v2"
    # "cuda" | "cpu" | None -> (env or "cpu")
    # We default to CPU on Spaces to avoid ZeroGPU device mixups.
    device: Optional[str] = None

    def _lazy_import(self) -> None:
        global _TOK, _MODEL, _TORCH
        if _TORCH is None:
            import torch as _torch
            _TORCH = _torch
        if _TOK is None or _MODEL is None:
            from transformers import AutoTokenizer, AutoModel  # type: ignore
            _TOK = AutoTokenizer.from_pretrained(self.model_name)
            _MODEL = AutoModel.from_pretrained(self.model_name)
        # Prefer explicit device -> env override -> default to CPU
        dev = self.device or os.getenv("EMBEDDING_DEVICE") or "cpu"
        _MODEL.to(dev).eval()
        self.device = dev

    def encode(self, texts: Iterable[str], batch_size: int = 32) -> "tuple[_TORCH.Tensor, list[str]]":
        """
        Returns (embeddings, texts_list). Embeddings are a CPU torch.Tensor [N, D], unit-normalized.
        """
        self._lazy_import()
        torch = _TORCH  # local alias
        texts_list = list(texts)
        if not texts_list:
            # Hidden size available after _lazy_import
            return torch.empty((0, _MODEL.config.hidden_size)), []  # type: ignore

        all_out = []
        with torch.inference_mode():
            for i in range(0, len(texts_list), batch_size):
                batch = texts_list[i:i + batch_size]
                # Tokenize and move to model device
                enc = _TOK(batch, padding=True, truncation=True, return_tensors="pt").to(self.device)  # type: ignore
                out = _MODEL(**enc)
                last = out.last_hidden_state  # [B, T, H]
                mask = enc["attention_mask"].unsqueeze(-1)  # [B, T, 1]
                # mean pool
                summed = (last * mask).sum(dim=1)
                counts = mask.sum(dim=1).clamp(min=1)
                pooled = summed / counts
                # L2 normalize
                pooled = pooled / pooled.norm(dim=1, keepdim=True).clamp(min=1e-12)
                # Collect on CPU for downstream ops
                all_out.append(pooled.cpu())

        embs = torch.cat(all_out, dim=0) if all_out else torch.empty((0, _MODEL.config.hidden_size))  # type: ignore
        return embs, texts_list


# --------------------------- Utilities ---------------------------

def _normalize_whitespace(s: str) -> str:
    return " ".join(s.strip().split())


def _default_preprocess(s: str) -> str:
    # Keep simple, deterministic preprocessing. Users can override with a custom callable.
    return _normalize_whitespace(s)


@dataclass
class PhraseIndex:
    phrases_by_level: Dict[str, List[str]]
    embeddings_by_level: Dict[str, "Any"]
    model_name: str


def build_phrase_index(
    backend: HFEmbeddingBackend,
    phrases_by_level: Dict[str, Iterable[str]],
) -> PhraseIndex:
    """
    Pre-encode all anchor phrases per level into a searchable index.
    """
    # Flatten texts while preserving level boundaries
    cleaned: Dict[str, List[str]] = {lvl: [_default_preprocess(p) for p in phrases] for lvl, phrases in phrases_by_level.items()}
    all_texts: List[str] = []
    spans: List[Tuple[str, int, int]] = []  # (level, start, end) in the flat list
    cur = 0
    for lvl, plist in cleaned.items():
        start = cur
        all_texts.extend(plist)
        cur += len(plist)
        spans.append((lvl, start, cur))

    embs, _ = backend.encode(all_texts)  # embs is a CPU torch.Tensor [N, D]

    # Slice embeddings back into level buckets
    torch = _TORCH
    embeddings_by_level: Dict[str, "Any"] = {}
    for lvl, start, end in spans:
        if end > start:
            embeddings_by_level[lvl] = embs[start:end]  # torch.Tensor slice [n_i, D]
        else:
            embeddings_by_level[lvl] = torch.empty((0, embs.shape[1]))  # type: ignore

    return PhraseIndex(
        phrases_by_level={lvl: list(pl) for lvl, pl in cleaned.items()},
        embeddings_by_level=embeddings_by_level,
        model_name=backend.model_name
    )


def _aggregate_sims(
    sims: "Any", agg: Agg, topk: int
) -> float:
    """
    Aggregate a 1D tensor of similarities into a single score.
    """
    torch = _TORCH
    if sims.numel() == 0:
        return float("nan")
    if agg == "mean":
        return float(sims.mean().item())
    if agg == "max":
        return float(sims.max().item())
    if agg == "topk_mean":
        k = min(topk, sims.numel())
        topk_vals, _ = torch.topk(sims, k)
        return float(topk_vals.mean().item())
    raise ValueError(f"Unknown agg: {agg}")


# --------------------------- Public API ---------------------------

def classify_levels_phrases(
    question: str,
    blooms_phrases: Dict[str, Iterable[str]],
    dok_phrases: Dict[str, Iterable[str]],
    *,
    model_name: str = "sentence-transformers/all-MiniLM-L6-v2",
    agg: Agg = "max",
    topk: int = 5,
    preprocess: Optional[Callable[[str], str]] = None,
    backend: Optional[HFEmbeddingBackend] = None,
    prebuilt_bloom_index: Optional[PhraseIndex] = None,
    prebuilt_dok_index: Optional[PhraseIndex] = None,
    return_phrase_matches: bool = True,
) -> Dict[str, Any]:
    """
    Score a question against Bloom's taxonomy and DOK (Depth of Knowledge)
    using cosine similarity to level-specific anchor phrases.
    """
    preprocess = preprocess or _default_preprocess
    question_clean = preprocess(question)

    # Prepare backend (defaults to CPU)
    be = backend or HFEmbeddingBackend(model_name=model_name)

    # Build / reuse indices
    bloom_index = prebuilt_bloom_index or build_phrase_index(be, blooms_phrases)
    dok_index = prebuilt_dok_index or build_phrase_index(be, dok_phrases)

    # Encode question -> CPU torch.Tensor [1, D]
    q_emb, _ = be.encode([question_clean])
    q_emb = q_emb[0:1]
    torch = _TORCH

    def _score_block(index: PhraseIndex) -> Tuple[Dict[str, float], Dict[str, List[Tuple[str, float]]]]:
        scores: Dict[str, float] = {}
        top_contribs: Dict[str, List[Tuple[str, float]]] = {}

        for lvl, phrases in index.phrases_by_level.items():
            embs = index.embeddings_by_level[lvl]  # torch.Tensor [N, D]
            if embs.numel() == 0:
                scores[lvl] = float("nan")
                top_contribs[lvl] = []
                continue
            # cosine similarity since embs and q_emb are unit-normalized
            sims = (q_emb @ embs.T).squeeze(0)
            scores[lvl] = _aggregate_sims(sims, agg, topk)
            if return_phrase_matches:
                k = min(5, sims.numel())
                vals, idxs = torch.topk(sims, k)
                top_contribs[lvl] = [(phrases[int(i)], float(v.item())) for v, i in zip(vals, idxs)]
        return scores, top_contribs

    bloom_scores, bloom_top = _score_block(bloom_index)
    dok_scores, dok_top = _score_block(dok_index)

    def _best(scores: Dict[str, float]) -> Tuple[str, float]:
        # max with NaN-safe handling
        best_lvl, best_val = None, -float("inf")
        for lvl, val in scores.items():
            if isinstance(val, float) and (not math.isnan(val)) and val > best_val:
                best_lvl, best_val = lvl, val
        return best_lvl or "", best_val

    best_bloom, best_bloom_val = _best(bloom_scores)
    best_dok, best_dok_val = _best(dok_scores)

    return {
        "question": question_clean,
        "model_name": be.model_name,
        "blooms": {
            "scores": bloom_scores,
            "best_level": best_bloom,
            "best_score": best_bloom_val,
            "top_phrases": bloom_top if return_phrase_matches else None,
        },
        "dok": {
            "scores": dok_scores,
            "best_level": best_dok,
            "best_score": best_dok_val,
            "top_phrases": dok_top if return_phrase_matches else None,
        },
        "config": {
            "agg": agg,
            "topk": topk if agg == "topk_mean" else None,
        },
    }