"""
TF-IDF + Truncated SVD embedding for failure records.

No network downloads; no GPU required. Runs entirely with scikit-learn.

Feature construction combines:
- TF-IDF on a failure-context string (input text, expected label, prediction, reasoning type)
- TruncatedSVD to produce a dense, low-dimensional representation

Adapted from failure-induced-benchmarks/src/failure_geometry/embedding.py
(engineered sparse features) but extended to dense SVD projections for
KMeans and scatter visualisation.
"""

from __future__ import annotations

import numpy as np
from sklearn.decomposition import TruncatedSVD
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.preprocessing import normalize


EMBED_DIMS = 32


def _failure_text(failure: dict) -> str:
    return (
        f"input: {failure['x']} "
        f"expected: {failure['y']} "
        f"prediction: {failure['prediction']} "
        f"type: {failure['reasoning_type']} "
        f"model: {failure['model_id']}"
    )


def embed_failures(
    failures: list[dict],
    n_components: int = EMBED_DIMS,
) -> np.ndarray:
    if not failures:
        return np.empty((0, n_components))

    texts = [_failure_text(f) for f in failures]
    n = len(texts)

    vectorizer = TfidfVectorizer(
        max_features=800,
        ngram_range=(1, 2),
        sublinear_tf=True,
    )
    tfidf = vectorizer.fit_transform(texts)

    effective_dims = min(n_components, tfidf.shape[1] - 1, n - 1)
    if effective_dims < 2:
        arr = tfidf.toarray()
        return normalize(arr[:, : max(effective_dims, 1)])

    svd = TruncatedSVD(n_components=effective_dims, random_state=42)
    dense = svd.fit_transform(tfidf)
    return normalize(dense)


def embed_for_scatter(failures: list[dict]) -> np.ndarray:
    """Return exactly 2 components for PCA-style scatter plots."""
    if len(failures) < 3:
        return np.zeros((len(failures), 2))
    return embed_failures(failures, n_components=2)