168 lines
5.6 KiB
Python

"""Unit tests for reranked search and canonical source enrichment."""
from collections.abc import Sequence
from uuid import UUID, uuid4
import pytest
from sqlalchemy import create_engine
from sqlalchemy.orm import Session, sessionmaker
from ekp_core.database import create_session_factory
from ekp_core.dense_retrieval import DenseRetrievalResult
from ekp_core.search import CanonicalSourceError, RerankerContractError, SearchService
from ekp_domain import Base, Chunk, Document, DocumentStatus
from ekp_reranker import FakeRerankerProvider, RerankResult
from ekp_vector_store import DenseSearchCandidate
class StubDenseRetriever:
"""Return prebuilt candidates while recording the query."""
def __init__(self, candidates: tuple[DenseSearchCandidate, ...]) -> None:
self.candidates = candidates
self.queries: list[str] = []
def search(self, query: str, top_n: int | None = None) -> DenseRetrievalResult:
self.queries.append(query)
return DenseRetrievalResult(query.strip(), "profile-v1", self.candidates)
def _candidate(
chunk_id: UUID,
document_id: UUID,
content: str,
*,
dense_score: float,
) -> DenseSearchCandidate:
return DenseSearchCandidate(
index_record_id=uuid4(),
chunk_id=chunk_id,
document_id=document_id,
filename="stale-index-name.md",
page=None,
page_end=None,
section="stale-index-section",
chunk_index=0,
content=content,
score=dense_score,
index_profile_id="profile-v1",
)
def _canonical_source_session() -> tuple[sessionmaker[Session], UUID, UUID, UUID]:
engine = create_engine("sqlite+pysqlite:///:memory:")
Base.metadata.create_all(engine)
session_factory = create_session_factory(engine)
document_id = uuid4()
first_chunk_id = uuid4()
second_chunk_id = uuid4()
with session_factory() as session:
session.add(
Document(
id=document_id,
filename="policy.md",
media_type="text/markdown",
size_bytes=100,
storage_key=f"documents/{document_id}/policy.md",
status=DocumentStatus.READY,
)
)
session.add_all(
[
Chunk(
chunk_id=first_chunk_id,
document_id=document_id,
filename="policy.md",
page=1,
page_end=1,
section="报销材料",
section_path=["报销材料"],
chunk_index=0,
content="提交报销材料",
token_count=6,
strategy_version="structure_v1",
),
Chunk(
chunk_id=second_chunk_id,
document_id=document_id,
filename="policy.md",
page=2,
page_end=2,
section="住宿标准",
section_path=["住宿标准"],
chunk_index=1,
content="员工出差住宿标准说明",
token_count=9,
strategy_version="structure_v1",
),
]
)
session.commit()
return session_factory, document_id, first_chunk_id, second_chunk_id
def test_search_reranks_and_rehydrates_source_from_canonical_chunk() -> None:
session_factory, document_id, first_chunk_id, second_chunk_id = _canonical_source_session()
dense = StubDenseRetriever(
(
_candidate(first_chunk_id, document_id, "提交报销材料", dense_score=0.91),
_candidate(second_chunk_id, document_id, "员工出差住宿标准说明", dense_score=0.72),
)
)
reranker = FakeRerankerProvider()
service = SearchService(session_factory, dense, reranker, default_top_k=1)
result = service.search("员工出差住宿标准")
assert len(result.results) == 1
source = result.results[0]
assert source.chunk_id == second_chunk_id
assert source.filename == "policy.md"
assert source.page == 2
assert source.section == "住宿标准"
assert source.content == "员工出差住宿标准说明"
assert source.score == 0.72
assert source.rerank_score == 1.0
def test_search_rejects_stale_vector_content() -> None:
session_factory, document_id, first_chunk_id, _ = _canonical_source_session()
dense = StubDenseRetriever(
(_candidate(first_chunk_id, document_id, "索引中的旧内容", dense_score=0.8),)
)
service = SearchService(
session_factory,
dense,
FakeRerankerProvider(),
)
with pytest.raises(CanonicalSourceError, match="stale"):
service.search("旧内容", top_k=1)
class DuplicateReranker:
model_name = "bad"
max_documents = 64
def rerank(
self,
query: str,
documents: Sequence[str],
top_n: int,
) -> tuple[RerankResult, ...]:
return (RerankResult(0, 0.9), RerankResult(0, 0.8))
def test_search_rejects_duplicate_reranker_indexes() -> None:
session_factory, document_id, first_chunk_id, second_chunk_id = _canonical_source_session()
dense = StubDenseRetriever(
(
_candidate(first_chunk_id, document_id, "提交报销材料", dense_score=0.9),
_candidate(second_chunk_id, document_id, "员工出差住宿标准说明", dense_score=0.8),
)
)
service = SearchService(session_factory, dense, DuplicateReranker(), default_top_k=2)
with pytest.raises(RerankerContractError, match="duplicate"):
service.search("住宿标准")