test_retrieval_confidence.py

"""
Unit tests for retrieval/confidence.py — pure functions, no network required.
"""
import pytest
from langchain_core.documents import Document

from retrieval.confidence import (
    build_vector_score_map,
    token_overlap_ratio,
    compute_text_confidence,
)
from retrieval.fusion import doc_identity_key


def _doc(content: str, source: str = "test.pdf", page=None) -> Document:
    meta = {"source": source}
    if page is not None:
        meta["page"] = page
    return Document(page_content=content, metadata=meta)


class TestBuildVectorScoreMap:
    def test_normalizes_score(self):
        doc = _doc("content")
        score_map = build_vector_score_map([(doc, 0.85)])
        key = doc_identity_key(doc)
        assert score_map[key] == pytest.approx(0.85)

    def test_clamps_above_one(self):
        doc = _doc("content")
        score_map = build_vector_score_map([(doc, 1.5)])
        key = doc_identity_key(doc)
        assert score_map[key] == 1.0

    def test_clamps_below_zero(self):
        doc = _doc("content")
        score_map = build_vector_score_map([(doc, -0.1)])
        key = doc_identity_key(doc)
        assert score_map[key] == 0.0

    def test_keeps_best_score_for_duplicates(self):
        doc = _doc("content")
        score_map = build_vector_score_map([(doc, 0.5), (doc, 0.9)])
        key = doc_identity_key(doc)
        assert score_map[key] == pytest.approx(0.9)


class TestTokenOverlapRatio:
    def test_full_overlap(self):
        doc = _doc("coolfire vibration watch alarm", source="coolfire watch.pdf")
        ratio = token_overlap_ratio("coolfire watch alarm", doc)
        assert ratio > 0.8

    def test_no_overlap(self):
        doc = _doc("unrelated content here", source="other.pdf")
        ratio = token_overlap_ratio("coolfire watch alarm", doc)
        assert ratio == 0.0

    def test_empty_question(self):
        doc = _doc("some content")
        ratio = token_overlap_ratio("", doc)
        assert ratio == 0.0

    def test_result_bounded(self):
        doc = _doc("word word word word word")
        ratio = token_overlap_ratio("word word word", doc)
        assert 0.0 <= ratio <= 1.0


class TestComputeTextConfidence:
    def test_empty_docs_returns_zero(self):
        assert compute_text_confidence("question", [], {}) == 0.0

    def test_high_score_raises_confidence(self):
        doc = _doc("coolfire alarm instructions", source="coolfire watch.pdf")
        key = doc_identity_key(doc)
        score_map = {key: 0.95}
        conf = compute_text_confidence("coolfire alarm", [doc], score_map)
        assert conf > 0.5

    def test_zero_score_lowers_confidence(self):
        doc = _doc("some content", source="unrelated.pdf")
        key = doc_identity_key(doc)
        score_map = {key: 0.0}
        conf = compute_text_confidence("coolfire alarm", [doc], score_map)
        assert conf < 0.7

    def test_result_bounded(self):
        doc = _doc("coolfire watch", source="coolfire.pdf")
        key = doc_identity_key(doc)
        score_map = {key: 1.0}
        conf = compute_text_confidence("coolfire watch", [doc], score_map)
        assert 0.0 <= conf <= 1.0