merge: T8 pgvector 混合检索与缺口埋点（T-P2-03/04/06，TD-08）
This commit is contained in:
@@ -0,0 +1,297 @@
|
||||
"""混合检索集成：真实 PostgreSQL + pgvector，嵌入用 DeterministicEmbedder（不联网）.
|
||||
|
||||
证明三件单测证明不了的事（任务文档 §4）：
|
||||
1. 硬过滤零漏 —— 禁忌/draft 实体即使向量最近也绝不出现（T-P2-04 红线）；
|
||||
2. 关键词路径在 tsv 'simple' 分词下真实命中；
|
||||
3. 无结果 query 落 kb_gap 且 hit_count 递增（T-P2-06）。
|
||||
外加：向量化重跑整组替换且逐行带版本（T-P2-03）、kb_asset partial 索引可用。
|
||||
|
||||
Run: MEIDA_RUN_INTEGRATION=1 .venv/bin/pytest -m integration -q --no-cov
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
import uuid
|
||||
from collections.abc import AsyncIterator
|
||||
from typing import Any
|
||||
|
||||
import pytest
|
||||
from sqlalchemy import delete, select, text
|
||||
|
||||
from packages.domain.ids import next_id
|
||||
from packages.retrieval.embedding import DeterministicEmbedder
|
||||
from packages.retrieval.repo import ChunkInput, hybrid_search, upsert_chunks
|
||||
from packages.retrieval.service import RetrievalService, UserContext
|
||||
from packages.shared.db import dispose_engine, session_scope
|
||||
from packages.shared.models import KbAsset, KbChunk, KbEntity, KbGap
|
||||
|
||||
|
||||
def _infra_available() -> bool:
|
||||
return bool(os.getenv("MEIDA_RUN_INTEGRATION"))
|
||||
|
||||
|
||||
pytestmark = [
|
||||
pytest.mark.integration,
|
||||
pytest.mark.skipif(
|
||||
not _infra_available(),
|
||||
reason="set MEIDA_RUN_INTEGRATION=1 with PostgreSQL and Redis running",
|
||||
),
|
||||
]
|
||||
|
||||
|
||||
def _unique(name: str) -> str:
|
||||
return f"it-t8-{name}-{uuid.uuid4().hex[:10]}"
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
async def _hermetic_engine() -> AsyncIterator[None]:
|
||||
"""每个用例前后重建全局 engine：asyncpg 连接绑定创建它的事件循环，
|
||||
pytest-asyncio 给每个用例一个新 loop，复用池里旧连接会报
|
||||
"attached to a different loop"。app 系用例由 lifespan 的 dispose_engine
|
||||
兜底；直接用 session_scope 的本文件必须自己管。"""
|
||||
await dispose_engine()
|
||||
yield
|
||||
await dispose_engine()
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
async def clean_kb() -> AsyncIterator[dict[str, list[Any]]]:
|
||||
"""登记本用例造的脏数据并在结束时清掉：集成库长活，不能留假实体."""
|
||||
registry: dict[str, list[Any]] = {"entities": [], "gaps": []}
|
||||
yield registry
|
||||
async with session_scope() as session:
|
||||
if registry["entities"]:
|
||||
ids = registry["entities"]
|
||||
await session.execute(delete(KbChunk).where(KbChunk.entity_id.in_(ids)))
|
||||
await session.execute(delete(KbAsset).where(KbAsset.entity_id.in_(ids)))
|
||||
await session.execute(delete(KbEntity).where(KbEntity.entity_id.in_(ids)))
|
||||
for query_text in registry["gaps"]:
|
||||
await session.execute(delete(KbGap).where(KbGap.query_text == query_text))
|
||||
|
||||
|
||||
async def _add_entity(
|
||||
registry: dict[str, list[Any]],
|
||||
*,
|
||||
entity_type: str = "item",
|
||||
status: str = "published",
|
||||
attrs: dict[str, Any] | None = None,
|
||||
) -> int:
|
||||
entity_id = next_id()
|
||||
registry["entities"].append(entity_id)
|
||||
async with session_scope() as session:
|
||||
session.add(
|
||||
KbEntity(
|
||||
entity_id=entity_id,
|
||||
entity_type=entity_type,
|
||||
name=_unique(entity_type),
|
||||
attrs=attrs or {},
|
||||
status=status,
|
||||
)
|
||||
)
|
||||
return entity_id
|
||||
|
||||
|
||||
class RenamedEmbedder:
|
||||
"""换名复用确定性向量：验证 model/version 逐行落库与整组替换（T-P2-03）."""
|
||||
|
||||
def __init__(self, model_name: str, version: str) -> None:
|
||||
self._inner = DeterministicEmbedder(dimensions=1024)
|
||||
self.model_name = model_name
|
||||
self.version = version
|
||||
self.dimensions = 1024
|
||||
|
||||
async def embed(self, texts: list[str]) -> list[list[float]]:
|
||||
return await self._inner.embed(texts)
|
||||
|
||||
|
||||
# --- 1. 硬过滤零漏（T-P2-04 红线）--------------------------------------------
|
||||
async def test_hard_filter_never_leaks_taboo_or_draft(clean_kb: dict[str, list[Any]]) -> None:
|
||||
embedder = DeterministicEmbedder(dimensions=1024)
|
||||
query_text = "pear shape body dress advice"
|
||||
# 防御性登记：若回归导致空结果被记进 kb_gap，用例结束也能清干净。
|
||||
clean_kb["gaps"].append(query_text)
|
||||
|
||||
taboo_id = await _add_entity(
|
||||
clean_kb, attrs={"occasions": ["wedding"], "avoid_body_shapes": ["apple"]}
|
||||
)
|
||||
draft_id = await _add_entity(clean_kb, status="draft")
|
||||
good_id = await _add_entity(clean_kb, attrs={"occasions": ["wedding"]})
|
||||
|
||||
async with session_scope() as session:
|
||||
# 禁忌与 draft 的 chunk 和 query 完全同文：不过滤的话它们必然霸占
|
||||
# 向量第一名（距离 0）—— 这样断言"零出现"才有说服力。
|
||||
await upsert_chunks(session, taboo_id, [ChunkInput(text=query_text)], embedder)
|
||||
await upsert_chunks(session, draft_id, [ChunkInput(text=query_text)], embedder)
|
||||
await upsert_chunks(
|
||||
session,
|
||||
good_id,
|
||||
[ChunkInput(text="soft knit sweater paired with a high waist skirt")],
|
||||
embedder,
|
||||
)
|
||||
|
||||
async with session_scope() as session:
|
||||
svc = RetrievalService(session, embedding=embedder)
|
||||
outcome = await svc.search(
|
||||
query_text,
|
||||
user_ctx=UserContext(body_shape="apple", occasion="wedding", entity_types=("item",)),
|
||||
)
|
||||
|
||||
assert outcome.degraded is False
|
||||
assert outcome.results, "正常 published 实体应被召回"
|
||||
assert {r.entity_id for r in outcome.results} == {good_id}
|
||||
|
||||
# 对照组：去掉禁忌/场景过滤后禁忌实体确实在库里且可被召回 ——
|
||||
# 证明上面拦截它的是硬过滤，而不是数据没写进去。draft 永远不出现。
|
||||
query_vec = (await embedder.embed([query_text]))[0]
|
||||
async with session_scope() as session:
|
||||
control = await hybrid_search(session, query_vec, query_text)
|
||||
control_entities = {r.entity_id for r in control}
|
||||
assert taboo_id in control_entities
|
||||
assert draft_id not in control_entities
|
||||
|
||||
|
||||
# --- 2. 关键词路径（tsv 'simple' 分词）----------------------------------------
|
||||
async def test_keyword_recall_ranks_matching_chunk_first(
|
||||
clean_kb: dict[str, list[Any]],
|
||||
) -> None:
|
||||
embedder = DeterministicEmbedder(dimensions=1024)
|
||||
entity_id = await _add_entity(clean_kb)
|
||||
matching = "breathable linen shirt keeps you cool in summer"
|
||||
unrelated = "heavy wool coat for freezing winter nights"
|
||||
|
||||
async with session_scope() as session:
|
||||
await upsert_chunks(
|
||||
session,
|
||||
entity_id,
|
||||
[ChunkInput(text=matching), ChunkInput(text=unrelated)],
|
||||
embedder,
|
||||
)
|
||||
|
||||
# 查询文本与两个 chunk 都不同文 → 向量路分不出高下（近似正交），
|
||||
# 排序由关键词路（0.2 权重、归一后差距 1.0）决定。
|
||||
async with session_scope() as session:
|
||||
results = await hybrid_search(
|
||||
session, (await embedder.embed(["linen shirt summer"]))[0], "linen shirt summer"
|
||||
)
|
||||
|
||||
assert [r.text for r in results] == [matching, unrelated]
|
||||
assert results[0].kw_score == pytest.approx(1.0)
|
||||
assert results[1].kw_score == 0.0
|
||||
|
||||
|
||||
# --- 3. 向量化可重跑（T-P2-03）------------------------------------------------
|
||||
async def test_upsert_chunks_is_versioned_and_rerunnable(
|
||||
clean_kb: dict[str, list[Any]],
|
||||
) -> None:
|
||||
entity_id = await _add_entity(clean_kb)
|
||||
|
||||
async with session_scope() as session:
|
||||
first = await upsert_chunks(
|
||||
session,
|
||||
entity_id,
|
||||
[ChunkInput(text="rule one"), ChunkInput(text="rule two")],
|
||||
RenamedEmbedder("bge-m3", "v1"),
|
||||
)
|
||||
assert {c.embedding_model for c in first} == {"bge-m3"}
|
||||
assert {c.embedding_version for c in first} == {"v1"}
|
||||
|
||||
# 换模型名（= 换版本）重跑：整组替换，不残留旧版本向量
|
||||
async with session_scope() as session:
|
||||
await upsert_chunks(
|
||||
session,
|
||||
entity_id,
|
||||
[ChunkInput(text="rule one"), ChunkInput(text="rule two"), ChunkInput(text="rule new")],
|
||||
RenamedEmbedder("bge-m4", "v2"),
|
||||
)
|
||||
|
||||
async with session_scope() as session:
|
||||
rows = (
|
||||
(await session.execute(select(KbChunk).where(KbChunk.entity_id == entity_id)))
|
||||
.scalars()
|
||||
.all()
|
||||
)
|
||||
assert len(rows) == 3
|
||||
assert {r.embedding_model for r in rows} == {"bge-m4"}
|
||||
assert {r.embedding_version for r in rows} == {"v2"}
|
||||
|
||||
|
||||
# --- 4. 无结果 query 落 kb_gap（T-P2-06）--------------------------------------
|
||||
async def test_no_result_query_lands_in_kb_gap_with_counter(
|
||||
clean_kb: dict[str, list[Any]],
|
||||
) -> None:
|
||||
query_text = _unique("missing")
|
||||
clean_kb["gaps"].append(query_text)
|
||||
svc_embedder = DeterministicEmbedder(dimensions=1024)
|
||||
|
||||
async with session_scope() as session:
|
||||
svc = RetrievalService(session, embedding=svc_embedder)
|
||||
first = await svc.search(query_text, user_ctx=UserContext(entity_types=("nonexistent",)))
|
||||
second = await svc.search(query_text, user_ctx=UserContext(entity_types=("nonexistent",)))
|
||||
|
||||
assert first.degraded and first.results == []
|
||||
assert second.degraded and second.results == []
|
||||
|
||||
async with session_scope() as session:
|
||||
gaps = (
|
||||
(await session.execute(select(KbGap).where(KbGap.query_text == query_text)))
|
||||
.scalars()
|
||||
.all()
|
||||
)
|
||||
assert len(gaps) == 1 # 同 query + 同 filter_ctx 只有一行
|
||||
assert gaps[0].hit_count == 2
|
||||
assert gaps[0].filter_ctx["entity_types"] == ["nonexistent"]
|
||||
|
||||
|
||||
# --- 5. kb_asset 按 entity 查询 + partial 索引可用 -----------------------------
|
||||
async def test_tryon_ready_assets_query_uses_partial_index(
|
||||
clean_kb: dict[str, list[Any]],
|
||||
) -> None:
|
||||
entity_id = await _add_entity(clean_kb)
|
||||
async with session_scope() as session:
|
||||
session.add_all(
|
||||
[
|
||||
KbAsset(
|
||||
asset_id=next_id(),
|
||||
entity_id=entity_id,
|
||||
asset_type="front",
|
||||
cos_key=_unique("front"),
|
||||
tryon_ready=True,
|
||||
review_status="approved",
|
||||
),
|
||||
KbAsset(
|
||||
asset_id=next_id(),
|
||||
entity_id=entity_id,
|
||||
asset_type="side",
|
||||
cos_key=_unique("side"),
|
||||
tryon_ready=False,
|
||||
review_status="pending",
|
||||
),
|
||||
]
|
||||
)
|
||||
|
||||
async with session_scope() as session:
|
||||
ready = (
|
||||
(
|
||||
await session.execute(
|
||||
select(KbAsset).where(
|
||||
KbAsset.entity_id == entity_id, KbAsset.tryon_ready.is_(True)
|
||||
)
|
||||
)
|
||||
)
|
||||
.scalars()
|
||||
.all()
|
||||
)
|
||||
assert len(ready) == 1
|
||||
assert ready[0].tryon_ready is True
|
||||
|
||||
# 关闭顺序扫描逼 planner 走索引，证明 partial 索引真的可用
|
||||
# （SET LOCAL 只影响当前事务，随 session_scope 提交即失效）。
|
||||
await session.execute(text("SET LOCAL enable_seqscan = off"))
|
||||
plan_lines = (
|
||||
(await session.execute(text("EXPLAIN SELECT asset_id FROM kb_asset WHERE tryon_ready")))
|
||||
.scalars()
|
||||
.all()
|
||||
)
|
||||
plan_text = "\n".join(str(line) for line in plan_lines)
|
||||
assert "ix_kb_asset_tryon_ready" in plan_text
|
||||
@@ -0,0 +1,99 @@
|
||||
"""DeterministicEmbedder 与工厂：纯数学，不连库不联网（T-P2-03 接口位）."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import math
|
||||
|
||||
import pytest
|
||||
|
||||
from packages.retrieval.embedding import (
|
||||
DETERMINISTIC_MODEL_NAME,
|
||||
DeterministicEmbedder,
|
||||
build_embedding_provider,
|
||||
)
|
||||
from packages.shared.config import get_settings
|
||||
from packages.shared.errors import AppError, ErrorCode
|
||||
|
||||
|
||||
def _dot(a: list[float], b: list[float]) -> float:
|
||||
return sum(x * y for x, y in zip(a, b, strict=True))
|
||||
|
||||
|
||||
async def test_same_text_embeds_to_same_vector() -> None:
|
||||
embedder = DeterministicEmbedder(dimensions=1024)
|
||||
first, second = await embedder.embed(["白衬衫怎么搭配", "白衬衫怎么搭配"])
|
||||
assert first == second
|
||||
|
||||
|
||||
async def test_dimensions_match_declaration() -> None:
|
||||
vec = (await DeterministicEmbedder(dimensions=1024).embed(["hello"]))[0]
|
||||
assert len(vec) == 1024
|
||||
small = (await DeterministicEmbedder(dimensions=8).embed(["hello"]))[0]
|
||||
assert len(small) == 8
|
||||
|
||||
|
||||
async def test_vectors_are_l2_normalized() -> None:
|
||||
for text in ("hello", "中文知识片段", ""):
|
||||
vec = (await DeterministicEmbedder(dimensions=1024).embed([text]))[0]
|
||||
assert math.isclose(_dot(vec, vec), 1.0, rel_tol=1e-6, abs_tol=1e-9)
|
||||
|
||||
|
||||
async def test_different_texts_are_near_orthogonal() -> None:
|
||||
"""不同文本近似正交（|dot| ≪ 1）：向量召回管线在无真模型时仍可测."""
|
||||
|
||||
left, right = await DeterministicEmbedder(dimensions=1024).embed(
|
||||
["interview outfit advice", "winter skincare routine"]
|
||||
)
|
||||
assert left != right
|
||||
assert abs(_dot(left, right)) < 0.1
|
||||
|
||||
|
||||
async def test_batch_preserves_order_and_length() -> None:
|
||||
texts = ["alpha", "beta", "gamma"]
|
||||
vectors = await DeterministicEmbedder(dimensions=16).embed(texts)
|
||||
assert len(vectors) == len(texts)
|
||||
singles = [(await DeterministicEmbedder(dimensions=16).embed([t]))[0] for t in texts]
|
||||
assert vectors == singles
|
||||
|
||||
|
||||
def test_metadata_is_stable() -> None:
|
||||
embedder = DeterministicEmbedder(dimensions=8)
|
||||
assert embedder.model_name == DETERMINISTIC_MODEL_NAME
|
||||
assert embedder.version
|
||||
assert embedder.dimensions == 8
|
||||
|
||||
|
||||
def test_invalid_dimensions_rejected() -> None:
|
||||
with pytest.raises(ValueError, match="dimensions"):
|
||||
DeterministicEmbedder(dimensions=0)
|
||||
|
||||
|
||||
def test_factory_builds_deterministic_embedder(monkeypatch: pytest.MonkeyPatch) -> None:
|
||||
# 带 alias 的 Settings 字段按字段名构造会被静默忽略 —— 统一走
|
||||
# setenv + cache_clear（上批经验，见任务文档 §2）。
|
||||
monkeypatch.setenv("KB_EMBEDDING_MODEL", "deterministic-v1")
|
||||
monkeypatch.setenv("KB_VECTOR_DIMENSIONS", "1024")
|
||||
get_settings.cache_clear()
|
||||
try:
|
||||
settings = get_settings()
|
||||
provider = build_embedding_provider(settings)
|
||||
assert isinstance(provider, DeterministicEmbedder)
|
||||
assert provider.model_name == "deterministic-v1"
|
||||
assert provider.dimensions == 1024
|
||||
finally:
|
||||
monkeypatch.undo()
|
||||
get_settings.cache_clear()
|
||||
|
||||
|
||||
def test_factory_rejects_unknown_model(monkeypatch: pytest.MonkeyPatch) -> None:
|
||||
monkeypatch.setenv("KB_EMBEDDING_MODEL", "bge-m3-onnx")
|
||||
get_settings.cache_clear()
|
||||
try:
|
||||
settings = get_settings()
|
||||
assert settings.kb_embedding_model == "bge-m3-onnx"
|
||||
with pytest.raises(AppError) as excinfo:
|
||||
build_embedding_provider(settings)
|
||||
assert excinfo.value.code == ErrorCode.CONFIG_ERROR
|
||||
finally:
|
||||
monkeypatch.undo()
|
||||
get_settings.cache_clear()
|
||||
@@ -0,0 +1,158 @@
|
||||
"""RRF 融合 + 权重 + 新鲜度的数学正确性（T-P2-04 验收 3）：构造已知名次断言最终序."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import itertools
|
||||
from datetime import UTC, datetime, timedelta
|
||||
|
||||
import pytest
|
||||
|
||||
from packages.retrieval.rerank import (
|
||||
DEFAULT_RERANK_WEIGHTS,
|
||||
FRESHNESS_HALF_LIFE_DAYS,
|
||||
RRF_K,
|
||||
RecallRow,
|
||||
freshness_score,
|
||||
rerank,
|
||||
rrf_component,
|
||||
rrf_score,
|
||||
)
|
||||
|
||||
NOW = datetime(2026, 8, 14, 12, 0, 0, tzinfo=UTC)
|
||||
|
||||
|
||||
def _row(
|
||||
chunk_id: int,
|
||||
*,
|
||||
vec_rank: int | None = None,
|
||||
kw_rank: int | None = None,
|
||||
updated_at: datetime = NOW,
|
||||
entity_id: int = 1,
|
||||
) -> RecallRow:
|
||||
return RecallRow(
|
||||
chunk_id=chunk_id,
|
||||
entity_id=entity_id,
|
||||
chunk_type="text",
|
||||
text=f"chunk-{chunk_id}",
|
||||
updated_at=updated_at,
|
||||
vec_rank=vec_rank,
|
||||
kw_rank=kw_rank,
|
||||
)
|
||||
|
||||
|
||||
# --- 单路 RRF 分量 ------------------------------------------------------------
|
||||
def test_rrf_component_is_max_normalized() -> None:
|
||||
assert rrf_component(1) == pytest.approx(1.0) # (60+1)/(60+1)
|
||||
assert rrf_component(2) == pytest.approx(61 / 62)
|
||||
assert rrf_component(RRF_K + 2) == pytest.approx(0.5) # 61/122：名次再排后一半
|
||||
assert rrf_component(None) == 0.0
|
||||
assert rrf_component(0) == 0.0
|
||||
assert rrf_component(-3) == 0.0
|
||||
|
||||
|
||||
def test_rrf_component_monotonic_non_increasing() -> None:
|
||||
values = [rrf_component(r) for r in range(1, 20)]
|
||||
assert all(a >= b for a, b in itertools.pairwise(values))
|
||||
|
||||
|
||||
def test_rrf_score_is_sum_of_raw_reciprocals() -> None:
|
||||
assert rrf_score(1, 2) == pytest.approx(1 / 61 + 1 / 62)
|
||||
assert rrf_score(None, 1) == pytest.approx(1 / 61)
|
||||
assert rrf_score(1, None) == pytest.approx(1 / 61)
|
||||
assert rrf_score(None, None) == 0.0
|
||||
|
||||
|
||||
# --- 新鲜度 ------------------------------------------------------------------
|
||||
def test_freshness_half_life_decay() -> None:
|
||||
assert freshness_score(NOW, now=NOW) == pytest.approx(1.0)
|
||||
old = NOW - timedelta(days=FRESHNESS_HALF_LIFE_DAYS)
|
||||
assert freshness_score(old, now=NOW) == pytest.approx(0.5)
|
||||
older = NOW - timedelta(days=2 * FRESHNESS_HALF_LIFE_DAYS)
|
||||
assert freshness_score(older, now=NOW) == pytest.approx(0.25)
|
||||
# 时钟漂移产生未来时间戳 → 按 0 天处理，不放大分数
|
||||
future = NOW + timedelta(hours=1)
|
||||
assert freshness_score(future, now=NOW) == pytest.approx(1.0)
|
||||
|
||||
|
||||
# --- 加权公式 ----------------------------------------------------------------
|
||||
def test_final_score_follows_architecture_formula() -> None:
|
||||
"""final = 0.45×vec + 0.2×kw + 0.25×rule + 0.1×fresh，逐项核对."""
|
||||
|
||||
row = _row(1, vec_rank=1, kw_rank=3)
|
||||
boost = {1: 0.5}
|
||||
scored = rerank([row], rule_boost=boost, now=NOW)
|
||||
expected = 0.45 * 1.0 + 0.2 * (61 / 63) + 0.25 * 0.5 + 0.1 * 1.0
|
||||
assert scored[0].final_score == pytest.approx(expected)
|
||||
assert scored[0].vec_score == pytest.approx(1.0)
|
||||
assert scored[0].kw_score == pytest.approx(61 / 63)
|
||||
assert scored[0].rule_score == pytest.approx(0.5)
|
||||
assert scored[0].freshness_score == pytest.approx(1.0)
|
||||
|
||||
|
||||
def test_default_weights_match_architecture_section_8_3() -> None:
|
||||
assert DEFAULT_RERANK_WEIGHTS == {
|
||||
"vector": 0.45,
|
||||
"keyword": 0.2,
|
||||
"rule": 0.25,
|
||||
"freshness": 0.1,
|
||||
}
|
||||
|
||||
|
||||
def test_dual_recall_beats_single_recall() -> None:
|
||||
"""两路都命中的 chunk 应压过只有向量路命中、名次更高的 chunk."""
|
||||
|
||||
both = _row(1, vec_rank=5, kw_rank=2)
|
||||
vec_only = _row(2, vec_rank=1)
|
||||
scored = rerank([both, vec_only], now=NOW)
|
||||
by_id = {c.chunk_id: c for c in scored}
|
||||
assert by_id[1].final_score > by_id[2].final_score
|
||||
|
||||
|
||||
def test_rule_boost_lifts_lower_recall_chunk() -> None:
|
||||
"""规则权重 0.25×boost 可以把召回名次落后的 chunk 抬到最前（自建知识库的意义）."""
|
||||
|
||||
rank5 = _row(1, vec_rank=5)
|
||||
rank1 = _row(2, vec_rank=1)
|
||||
scored = rerank([rank5, rank1], rule_boost={1: 1.0}, now=NOW)
|
||||
assert scored[0].chunk_id == 1
|
||||
assert scored[0].rule_score == pytest.approx(1.0)
|
||||
|
||||
|
||||
def test_keyword_only_chunk_gets_zero_vector_score() -> None:
|
||||
scored = rerank([_row(7, kw_rank=1)], now=NOW)
|
||||
assert scored[0].vec_score == 0.0
|
||||
assert scored[0].kw_score == pytest.approx(1.0)
|
||||
|
||||
|
||||
def test_freshness_breaks_ties_toward_newer_chunk() -> None:
|
||||
"""名次相同（现实中少见，但公式上可能）时，更新鲜者胜."""
|
||||
|
||||
fresh = _row(1, vec_rank=2, updated_at=NOW)
|
||||
stale = _row(2, vec_rank=2, updated_at=NOW - timedelta(days=90))
|
||||
scored = rerank([fresh, stale], now=NOW)
|
||||
assert scored[0].chunk_id == 1
|
||||
delta = scored[0].final_score - scored[1].final_score
|
||||
assert delta == pytest.approx(0.1 * (1.0 - 0.5**3))
|
||||
|
||||
|
||||
def test_zero_weight_dimension_ignores_that_signal() -> None:
|
||||
old = _row(1, vec_rank=1, updated_at=NOW - timedelta(days=365))
|
||||
new = _row(2, vec_rank=2, updated_at=NOW)
|
||||
scored = rerank([old, new], weights={"vector": 1.0}, now=NOW)
|
||||
assert scored[0].chunk_id == 1 # 新鲜度权重为 0，旧的反而名次靠前
|
||||
|
||||
|
||||
def test_stable_ordering_and_top_k_truncation() -> None:
|
||||
rows = [_row(i, vec_rank=i) for i in (30, 10, 20, 40)]
|
||||
scored = rerank(rows, now=NOW, top_k=2)
|
||||
assert [c.chunk_id for c in scored] == [10, 20]
|
||||
|
||||
|
||||
def test_equal_final_scores_break_tie_by_chunk_id() -> None:
|
||||
rows = [_row(90, kw_rank=1), _row(10, kw_rank=1)]
|
||||
scored = rerank(rows, now=NOW)
|
||||
assert [c.chunk_id for c in scored] == [10, 90]
|
||||
|
||||
|
||||
def test_empty_input_returns_empty() -> None:
|
||||
assert rerank([], now=NOW) == []
|
||||
@@ -0,0 +1,179 @@
|
||||
"""RetrievalService 编排逻辑：repo 层打桩，不连库（编排正确性与 DB 无关）."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import Any
|
||||
|
||||
import pytest
|
||||
|
||||
from packages.retrieval import repo
|
||||
from packages.retrieval import service as service_mod
|
||||
from packages.retrieval.rerank import ScoredChunk
|
||||
from packages.retrieval.service import RetrievalService, UserContext
|
||||
|
||||
|
||||
class FakeEmbedder:
|
||||
"""满足 EmbeddingProvider Protocol 的最小桩：记录调用，返回固定向量."""
|
||||
|
||||
model_name = "fake"
|
||||
version = "v1"
|
||||
dimensions = 4
|
||||
|
||||
def __init__(self) -> None:
|
||||
self.calls: list[list[str]] = []
|
||||
|
||||
async def embed(self, texts: list[str]) -> list[list[float]]:
|
||||
self.calls.append(list(texts))
|
||||
return [[0.25, 0.25, 0.25, 0.25] for _ in texts]
|
||||
|
||||
|
||||
class FakeSession:
|
||||
def __init__(self) -> None:
|
||||
self.commits = 0
|
||||
self.rollbacks = 0
|
||||
|
||||
async def commit(self) -> None:
|
||||
self.commits += 1
|
||||
|
||||
async def rollback(self) -> None:
|
||||
self.rollbacks += 1
|
||||
|
||||
|
||||
def _scored(chunk_id: int = 1) -> ScoredChunk:
|
||||
return ScoredChunk(
|
||||
chunk_id=chunk_id,
|
||||
entity_id=100,
|
||||
chunk_type="text",
|
||||
text="chunk",
|
||||
vec_score=1.0,
|
||||
kw_score=0.0,
|
||||
rule_score=0.0,
|
||||
freshness_score=1.0,
|
||||
final_score=0.55,
|
||||
)
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def stubbed_repo(monkeypatch: pytest.MonkeyPatch) -> dict[str, Any]:
|
||||
"""把 repo 的两个 IO 函数换成记录桩，service 逻辑照常执行."""
|
||||
|
||||
calls: dict[str, Any] = {"hybrid_search": None, "gaps": [], "results": []}
|
||||
|
||||
async def fake_hybrid_search(
|
||||
session: Any, query_vec: Any, query_text: str, **kwargs: Any
|
||||
) -> list[ScoredChunk]:
|
||||
calls["hybrid_search"] = {"query_vec": query_vec, "query_text": query_text, **kwargs}
|
||||
return list(calls["results"])
|
||||
|
||||
async def fake_record_gap(session: Any, query_text: str, filter_ctx: Any) -> None:
|
||||
calls["gaps"].append((query_text, filter_ctx))
|
||||
|
||||
monkeypatch.setattr(repo, "hybrid_search", fake_hybrid_search)
|
||||
monkeypatch.setattr(repo, "record_gap", fake_record_gap)
|
||||
return calls
|
||||
|
||||
|
||||
async def test_search_with_hits_returns_results_without_gap(stubbed_repo: dict[str, Any]) -> None:
|
||||
stubbed_repo["results"] = [_scored(1), _scored(2)]
|
||||
embedder = FakeEmbedder()
|
||||
session = FakeSession()
|
||||
svc = RetrievalService(session, embedding=embedder)
|
||||
|
||||
outcome = await svc.search("面试穿什么", user_ctx=UserContext(body_shape="apple"))
|
||||
|
||||
assert outcome.degraded is False
|
||||
assert [c.chunk_id for c in outcome.results] == [1, 2]
|
||||
assert stubbed_repo["gaps"] == []
|
||||
assert session.commits == 0 # 有结果的路径不写库，事务留给调用方
|
||||
assert embedder.calls == [["面试穿什么"]] # embed 的输入就是 query 原文
|
||||
|
||||
|
||||
async def test_search_forwards_user_context_to_hybrid(stubbed_repo: dict[str, Any]) -> None:
|
||||
stubbed_repo["results"] = [_scored()]
|
||||
svc = RetrievalService(FakeSession(), embedding=FakeEmbedder())
|
||||
|
||||
await svc.search(
|
||||
"婚礼穿什么",
|
||||
user_ctx=UserContext(
|
||||
body_shape="pear",
|
||||
occasion="wedding",
|
||||
entity_types=("item", "style"),
|
||||
rule_boost={42: 0.8},
|
||||
),
|
||||
)
|
||||
|
||||
seen = stubbed_repo["hybrid_search"]
|
||||
assert seen is not None
|
||||
assert seen["query_text"] == "婚礼穿什么"
|
||||
assert seen["body_shape"] == "pear"
|
||||
assert seen["occasion"] == "wedding"
|
||||
assert seen["entity_types"] == ("item", "style")
|
||||
assert seen["rule_boost"] == {42: 0.8}
|
||||
|
||||
|
||||
async def test_empty_entity_types_is_forwarded_as_none(stubbed_repo: dict[str, Any]) -> None:
|
||||
stubbed_repo["results"] = []
|
||||
svc = RetrievalService(FakeSession(), embedding=FakeEmbedder())
|
||||
|
||||
await svc.search("随便看看", user_ctx=UserContext())
|
||||
|
||||
seen = stubbed_repo["hybrid_search"]
|
||||
assert seen is not None
|
||||
assert seen["entity_types"] is None # 空元组 = 不限类型，不是匹配空集合
|
||||
|
||||
|
||||
async def test_no_results_records_gap_and_degrades(stubbed_repo: dict[str, Any]) -> None:
|
||||
stubbed_repo["results"] = []
|
||||
session = FakeSession()
|
||||
svc = RetrievalService(session, embedding=FakeEmbedder())
|
||||
|
||||
outcome = await svc.search(
|
||||
"不存在的知识点", user_ctx=UserContext(body_shape="apple", entity_types=("item",))
|
||||
)
|
||||
|
||||
assert outcome.degraded is True
|
||||
assert outcome.results == []
|
||||
assert session.commits == 1 # gap 埋点立即提交：检索本身只读，别拖住调用方事务
|
||||
assert stubbed_repo["gaps"] == [
|
||||
(
|
||||
"不存在的知识点",
|
||||
{"body_shape": "apple", "occasion": None, "entity_types": ["item"]},
|
||||
)
|
||||
]
|
||||
|
||||
|
||||
async def test_gap_failure_degrades_but_does_not_raise(
|
||||
stubbed_repo: dict[str, Any], monkeypatch: pytest.MonkeyPatch
|
||||
) -> None:
|
||||
"""埋点失败不该把"检索空结果"变成异常 —— 空结果本身是合法结局."""
|
||||
|
||||
async def broken_gap(session: Any, query_text: str, filter_ctx: Any) -> None:
|
||||
raise RuntimeError("db down")
|
||||
|
||||
monkeypatch.setattr(repo, "record_gap", broken_gap)
|
||||
stubbed_repo["results"] = []
|
||||
session = FakeSession()
|
||||
svc = RetrievalService(session, embedding=FakeEmbedder())
|
||||
|
||||
outcome = await svc.search("anything")
|
||||
|
||||
assert outcome.degraded is True
|
||||
assert session.rollbacks == 1
|
||||
|
||||
|
||||
async def test_default_context_when_user_ctx_omitted(stubbed_repo: dict[str, Any]) -> None:
|
||||
stubbed_repo["results"] = []
|
||||
svc = RetrievalService(FakeSession(), embedding=FakeEmbedder())
|
||||
|
||||
outcome = await svc.search("no ctx")
|
||||
|
||||
assert outcome.degraded is True
|
||||
assert stubbed_repo["gaps"] == [
|
||||
("no ctx", {"body_shape": None, "occasion": None, "entity_types": []})
|
||||
]
|
||||
|
||||
|
||||
def test_service_module_reexports_for_dependency_injection() -> None:
|
||||
"""service 与 repo 共享同一模块对象：monkeypatch 打桩路径就是生产调用路径."""
|
||||
|
||||
assert service_mod.repo is repo
|
||||
Reference in New Issue
Block a user