Spaces:
Sleeping
Sleeping
Aissugur Kurganbekov
Claude Opus 4.8
Add hybrid retrieval (TF-IDF + embeddings), OCR, A/B eval, and answer trust/UX
f65a6b3 | """Focused unit tests for the answer metadata dict (the shared-contract 5th | |
| element of answer_question). These assert the metadata's shape, the confidence | |
| range/calibration, evidence_status, and chunk counts — independently of the | |
| answer text itself.""" | |
| from app import config | |
| from app.answering import answer_question | |
| from tests.conftest import make_chunk | |
| INVOICE = ( | |
| "Invoice Number INV-2025-0473. Invoice Date March 3 2025. Due Date April 2 " | |
| "2025. Subtotal $2,230.00. Tax $183.98. Total Amount Due $2,413.98. " | |
| "Payment terms net 30." | |
| ) | |
| # Every metadata dict must carry exactly these keys (shared contract). | |
| EXPECTED_KEYS = { | |
| "question_type", | |
| "evidence_status", | |
| "confidence", | |
| "top_score", | |
| "matched_terms", | |
| "required_terms", | |
| "retrieved_chunks", | |
| "cited_chunks", | |
| "requested_mode", | |
| "answer_mode", | |
| "latency_ms", | |
| } | |
| def test_metadata_supported_answer_shape_and_confidence(): | |
| chunks = [make_chunk(INVOICE)] | |
| _answer, abstained, _mode, citations, metadata = answer_question( | |
| "What is the total amount due?", chunks | |
| ) | |
| assert abstained is False | |
| # Contract: the full key set is always present. | |
| assert set(metadata) == EXPECTED_KEYS | |
| assert metadata["evidence_status"] == "supported" | |
| # Confidence is a 0..1 float; a solid invoice-total match should be high. | |
| assert 0.0 <= metadata["confidence"] <= 1.0 | |
| assert metadata["confidence"] >= 0.6 | |
| # cited_chunks mirrors the returned citations. | |
| assert metadata["cited_chunks"] == len(citations) | |
| assert metadata["cited_chunks"] >= 1 | |
| assert metadata["question_type"] == "invoice_total" | |
| assert metadata["latency_ms"] >= 0 | |
| assert isinstance(metadata["latency_ms"], int) | |
| def test_metadata_abstained_low_confidence_and_no_citations(): | |
| chunks = [make_chunk("The weather today is sunny with a light breeze.")] | |
| _answer, abstained, _mode, _citations, metadata = answer_question( | |
| "What is the total amount due?", chunks | |
| ) | |
| assert abstained is True | |
| assert set(metadata) == EXPECTED_KEYS | |
| assert metadata["evidence_status"] == "abstained" | |
| assert metadata["confidence"] == 0.0 | |
| assert metadata["cited_chunks"] == 0 | |
| def test_metadata_confidence_in_unit_range_for_empty_document(): | |
| _answer, abstained, _mode, _citations, metadata = answer_question("anything?", []) | |
| assert abstained is True | |
| assert 0.0 <= metadata["confidence"] <= 1.0 | |
| assert metadata["evidence_status"] == "abstained" | |
| assert metadata["retrieved_chunks"] == 0 | |
| def test_metadata_question_type_payment_terms(): | |
| chunks = [make_chunk(INVOICE)] | |
| _answer, _abstained, _mode, _citations, metadata = answer_question( | |
| "What are the payment terms?", chunks | |
| ) | |
| assert metadata["question_type"] == "payment_terms" | |
| def test_metadata_match_text_populated_on_citation(): | |
| chunks = [make_chunk(INVOICE)] | |
| _answer, _abstained, _mode, citations, _metadata = answer_question( | |
| "What is the total amount due?", chunks | |
| ) | |
| assert citations | |
| # The focused span the question matched is surfaced for UI highlighting and | |
| # is a verbatim substring of the chunk text. | |
| assert citations[0].match_text is not None | |
| assert citations[0].match_text in INVOICE | |