Spaces:
Sleeping
Sleeping
Aissugur Kurganbekov
Claude Opus 4.8
Add hybrid retrieval (TF-IDF + embeddings), OCR, A/B eval, and answer trust/UX
f65a6b3 | """Hybrid (TF-IDF + dense) retrieval fusion — tested offline. | |
| We mock the query embedding so the test needs no model download. The point is to | |
| prove the Reciprocal-Rank-Fusion logic actually lets a SEMANTIC match surface | |
| above a chunk that scored higher on lexical TF-IDF, and that it degrades cleanly | |
| to the TF-IDF order when embeddings are unavailable. | |
| """ | |
| import numpy as np | |
| from app import embeddings | |
| from app.models import Chunk | |
| from app.retrieval import ScoredChunk, hybrid_rank | |
| def _vec(*xs) -> bytes: | |
| return np.asarray(xs, dtype=np.float32).tobytes() | |
| def test_rrf_surfaces_semantic_match_over_higher_tfidf(monkeypatch): | |
| # Query embeds to [1,0,0]. Chunk C is the semantic match; A/B are not. | |
| monkeypatch.setattr(embeddings, "embed_query", lambda q: np.asarray([1, 0, 0], dtype=np.float32)) | |
| A = Chunk(id=1, document_id=1, chunk_index=0, page=1, text="alpha", embedding=_vec(0.1, 1, 0)) | |
| B = Chunk(id=2, document_id=1, chunk_index=1, page=1, text="beta", embedding=_vec(0, 0, 1)) | |
| C = Chunk(id=3, document_id=1, chunk_index=2, page=1, text="gamma", embedding=_vec(1, 0, 0)) | |
| chunks = [A, B, C] | |
| # Lexical TF-IDF order: A best, B mid, C worst. | |
| tfidf = [ScoredChunk(A, 0.5), ScoredChunk(B, 0.3), ScoredChunk(C, 0.05)] | |
| fused = hybrid_rank("q", chunks, tfidf, top_k=3) | |
| order = [s.chunk.id for s in fused] | |
| # C (strong semantic, weak lexical) must now rank ABOVE B (weak on both), | |
| # even though B scored higher on TF-IDF — i.e. the dense signal moved it up. | |
| assert order.index(3) < order.index(2) | |
| # Display score is still the lexical TF-IDF score (citations stay interpretable). | |
| assert {s.chunk.id: s.score for s in fused}[3] == 0.05 | |
| def test_falls_back_to_tfidf_when_embeddings_unavailable(monkeypatch): | |
| monkeypatch.setattr(embeddings, "embed_query", lambda q: None) | |
| a = Chunk(id=1, document_id=1, chunk_index=0, text="a") | |
| b = Chunk(id=2, document_id=1, chunk_index=1, text="b") | |
| tfidf = [ScoredChunk(a, 0.5), ScoredChunk(b, 0.2)] | |
| fused = hybrid_rank("q", [a, b], tfidf, top_k=4) | |
| assert [s.chunk.id for s in fused] == [1, 2] | |