knowledge-platform/tests/unit/test_dense_retrieval.py
2026-09-01 15:27:53 +08:00

110 lines
3.4 KiB
Python

"""Unit tests for query embedding and provider-neutral Dense Top-N retrieval."""
from math import nan
from uuid import uuid4
import pytest
from ekp_core.dense_retrieval import (
DenseRetrievalContractError,
DenseRetrievalService,
InvalidRetrievalQueryError,
)
from ekp_domain import IndexProfile
from ekp_embedding import FakeEmbeddingProvider
from ekp_vector_store import DenseSearchRequest, FakeVectorStore, VectorRecord
def _record(
*,
vector: tuple[float, ...],
profile_id: str = "profile-v1",
filename: str = "policy.md",
) -> VectorRecord:
"""Build a minimal traceable vector row for exact-search assertions."""
return VectorRecord(
index_record_id=uuid4(),
chunk_id=uuid4(),
document_id=uuid4(),
filename=filename,
page=1,
page_end=1,
section="制度",
chunk_index=0,
content="测试内容",
embedding=vector,
embedding_model="fake",
index_profile_id=profile_id,
)
def test_fake_vector_store_filters_profile_before_cosine_top_n() -> None:
store = FakeVectorStore()
closest = _record(vector=(1.0, 0.0), filename="closest.md")
farther = _record(vector=(0.5, 0.5), filename="farther.md")
shadow = _record(vector=(1.0, 0.0), profile_id="shadow-v2", filename="shadow.md")
store.upsert((farther, shadow, closest))
candidates = store.dense_search(DenseSearchRequest((1.0, 0.0), "profile-v1", 2))
assert [candidate.filename for candidate in candidates] == ["closest.md", "farther.md"]
assert all(candidate.index_profile_id == "profile-v1" for candidate in candidates)
assert candidates[0].score == pytest.approx(1.0)
def test_dense_service_uses_default_top_twenty_and_trims_query() -> None:
embedding = FakeEmbeddingProvider(model_name="fake", dimensions=8)
store = FakeVectorStore()
service = DenseRetrievalService(
embedding,
store,
IndexProfile("profile-v1", "structure_v1", "fake", 8),
)
result = service.search(" 住宿标准 ")
assert result.query == "住宿标准"
assert result.index_profile_id == "profile-v1"
assert result.candidates == ()
assert store.dense_search_calls == 1
assert store.search_requests[0].top_n == 20
assert store.search_requests[0].index_profile_id == "profile-v1"
def test_dense_service_rejects_empty_query() -> None:
service = DenseRetrievalService(
FakeEmbeddingProvider(model_name="fake", dimensions=8),
FakeVectorStore(),
IndexProfile("profile-v1", "structure_v1", "fake", 8),
)
with pytest.raises(InvalidRetrievalQueryError):
service.search(" ")
class InvalidEmbeddingProvider:
"""Protocol-compatible provider returning a deliberately invalid vector."""
model_name = "fake"
dimensions = 2
max_batch_size = 1
def embed(self, texts: object) -> tuple[tuple[float, ...], ...]:
return ((nan, 0.0),)
def test_dense_service_rejects_non_finite_query_embedding() -> None:
service = DenseRetrievalService(
InvalidEmbeddingProvider(),
FakeVectorStore(),
IndexProfile("profile-v1", "structure_v1", "fake", 2),
)
with pytest.raises(DenseRetrievalContractError, match="non-finite"):
service.search("退款")
def test_dense_search_request_rejects_invalid_top_n() -> None:
with pytest.raises(ValueError, match="top_n"):
DenseSearchRequest((1.0, 0.0), "profile-v1", 0)