评审意见 §1.3 要求的修法。背景(双方实测共同确认):同一批集合名在两个开发环境里是两套不同 schema: 我方:knowledge_id / snippet(无 visibility),行数 106/177/73 架构师:doc_id / content / chapter / section / doc_no / visibility,行数 125/297/214 上一轮我把字段名硬编码成我方那套,在架构师环境会让 Milvus 报 field doc_id not exist → 三集合全失败 → 客服一律转人工(反向亦然)。硬编码任一套都会打挂另一套。 改法(采纳评审建议): - 新增 app/core/knowledge_schema.py:describe_collection → 逻辑名到物理名映射,按集合缓存; 缺必需字段的集合明确判为不可用并如实记 degraded,不静默零召回 - 检索服务改为逐集合探测:output_fields 只请求实际存在的字段;visibility 过滤有该字段才拼 - KnowledgeHit 对外形状不变,检索逻辑(字面召回/父子块/去重/置信判定)一行未改 测试:新增 17 个探测单测;架构师的关键词召回测试参数化为两套 schema 各跑一遍。
479 lines
23 KiB
Python
479 lines
23 KiB
Python
"""知识库检索:把客户问题向量化后在三个知识集合里检索(只读)。
|
||
|
||
为什么不复用记忆那套 `VectorMemoryAdapter`:它只把命中折叠成 `(memory_uuid, score)`,
|
||
会把知识块的标题与正文丢掉。而客服回答必须能把**原文与来源**一起交给客户——金融场景
|
||
里「答案出自哪份文件的哪一条」本身就是交付物的一部分,丢了正文等于没法给来源引用。
|
||
|
||
失败语义与基座一致:**任何一步失败都不抛异常给主链路**,而是返回 `degraded=True`
|
||
的空结果,由调用方(客服 Agent)据此走「引导客户致电人工客服」的兜底路径。
|
||
金融场景下"答不了"是可接受的结果,"答错"不是。
|
||
"""
|
||
|
||
from collections.abc import Awaitable, Callable, Sequence
|
||
from dataclasses import dataclass, field
|
||
from typing import Any, Protocol
|
||
|
||
from app.core.knowledge_schema import CollectionSchema, SchemaCache, detect_schema
|
||
|
||
# 三个知识集合(方案 §2.4.4 / §4.1)
|
||
FAQ_COLLECTION = "fin_faq_collection"
|
||
PRODUCT_COLLECTION = "fin_product_collection"
|
||
POLICY_COLLECTION = "fin_policy_collection"
|
||
DEFAULT_COLLECTIONS: tuple[str, ...] = (FAQ_COLLECTION, PRODUCT_COLLECTION, POLICY_COLLECTION)
|
||
|
||
# 检索输出字段:**运行时探测**,不硬编码任何一套 schema。
|
||
#
|
||
# 背景(实测):同一批集合名在不同环境下可能是两套 schema ——
|
||
# 环境甲:`knowledge_id` / `snippet`(无 visibility)
|
||
# 环境乙:`doc_id` / `content` / `visibility` / `chapter` / `source_file`
|
||
# 硬编码任一套都会把另一套打挂(Milvus 对不存在的字段直接报错 → 三集合全失败 →
|
||
# `degraded=True` → 客服一律"引导人工")。因此字段名一律经
|
||
# `app/core/knowledge_schema.py` 的 `detect_schema()` 探测得出:
|
||
# · 逻辑名 `doc_id` → 物理名 `doc_id` 或 `knowledge_id`(谁在就用谁)
|
||
# · 逻辑名 `content` → 物理名 `content` 或 `snippet`
|
||
# · `visibility` / `source_file` / `chapter` 等存在就用、不存在就跳过
|
||
# `KnowledgeHit` 的对外形状**保持不变**(下游 Agent 与测试依赖它),只改"怎么从库里取"。
|
||
|
||
# 关键字精确召回:客户问到产品名这类**专有名词**时,字面匹配比相似度更确定。
|
||
#
|
||
# 为什么需要:专有名词在 embedding 空间里不占优势。实测 160 条知识,客户问
|
||
# 「季季盈90天的起投金额是多少」向量 top1 = PROD-007 仅 0.6291(够不到 0.75 的硬门槛,
|
||
# 只能靠与次优的差值勉强通过);而同一次查询用 `title like "%季季盈%"` 是**唯一命中**
|
||
# PROD-007。既然客户已经明确说出了产品名,就不该再让相似度去赌。
|
||
KEYWORD_MATCH_SCORE = 1.0 # 字面命中的确定分,压过任何相似度分
|
||
# 与标题的最长公共子串至少要这么长,才算"客户确切提到了它"。
|
||
# 为什么不是 4:实测客户问「基金赎回几天到账」,与手册章节标题「5.2 基金赎回流程」
|
||
# 正好有 4 个字连续重合——但"基金赎回"是业务动作词,不是专有名词。产品名
|
||
# ("南方季季盈90天")通常比业务动作词长,取 6 字能同时保住产品名、挡住动作词。
|
||
MIN_KEYWORD_OVERLAP = 6
|
||
KEYWORD_SCAN_LIMIT = 500 # 一次最多扫描多少条标题;知识库到上千块后应改为倒排索引
|
||
|
||
# 行级子块命中时,其父块(整节)按子块分数的这个比例一并返回:排在子块之后,
|
||
# 既不抢"起投多少"这类聚焦答案,又不至于把差距压到转人工门槛之下。
|
||
PARENT_SCORE_RATIO = 0.9
|
||
|
||
# 字面匹配只在向量结果**不够确定**时介入。
|
||
#
|
||
# 这条门禁是实测逼出来的:客户问「基金赎回几天到账」,向量已给出正确答案
|
||
# (FAQ-0016「基金赎回到账需要多长时间?」得 0.8060),但产品手册里的章节标题
|
||
# 「五、申购赎回操作指南」与问句也有 6 个字连续重合,无条件字面匹配会把它顶到第一,
|
||
# 用操作步骤替换掉客户真正问的到账时间。所以字面匹配是**兜底**,不是优先。
|
||
#
|
||
# 门槛必须与客服 Agent 的高置信门槛一致,tests/unit 有断言锁定两者相等。
|
||
VECTOR_CONFIDENT_SCORE = 0.75
|
||
|
||
|
||
class VectorSearcher(Protocol):
|
||
"""只依赖用到的两个方法,便于测试替身注入。"""
|
||
|
||
def search(self, **kwargs: Any) -> Any: ...
|
||
|
||
|
||
Embedder = Callable[[str], Awaitable[list[float]]]
|
||
|
||
|
||
@dataclass(frozen=True)
|
||
class KnowledgeHit:
|
||
"""一条知识命中;`score` 为 COSINE 相似度(越大越相似)。"""
|
||
|
||
doc_id: str
|
||
title: str
|
||
content: str
|
||
score: float
|
||
source_file: str = ""
|
||
visibility: str = "public"
|
||
doc_no: str = ""
|
||
version: str = ""
|
||
chapter: str = ""
|
||
|
||
@property
|
||
def reference_title(self) -> str:
|
||
"""给客户看的来源标题:优先带内部文件编号,便于人工核对。"""
|
||
return f"{self.title}({self.doc_no})" if self.doc_no else self.title
|
||
|
||
|
||
@dataclass(frozen=True)
|
||
class KnowledgeSearchOutcome:
|
||
"""检索结果;`degraded=True` 表示检索链路故障,调用方必须走兜底而非当'没找到'。"""
|
||
|
||
hits: tuple[KnowledgeHit, ...] = ()
|
||
degraded: bool = False
|
||
reason: str = ""
|
||
searched_collections: tuple[str, ...] = field(default_factory=tuple)
|
||
|
||
@property
|
||
def best(self) -> KnowledgeHit | None:
|
||
return self.hits[0] if self.hits else None
|
||
|
||
@property
|
||
def top_score(self) -> float:
|
||
return self.hits[0].score if self.hits else 0.0
|
||
|
||
|
||
class KnowledgeSearchService:
|
||
def __init__(
|
||
self,
|
||
client: VectorSearcher | None,
|
||
embedder: Embedder | None,
|
||
*,
|
||
collections: Sequence[str] = DEFAULT_COLLECTIONS,
|
||
) -> None:
|
||
self._client = client
|
||
self._embedder = embedder
|
||
self._collections = tuple(collections)
|
||
# 字段探测结果按集合缓存(进程内一次):`describe_collection` 是元数据调用,
|
||
# 但检索是热路径,不该每次调用都打一次。集合重建后需重启或 `invalidate`。
|
||
self._schemas = SchemaCache()
|
||
|
||
@property
|
||
def available(self) -> bool:
|
||
"""向量库与向量化能力是否都在位;缺任一项都不做检索,直接走兜底。"""
|
||
return self._client is not None and self._embedder is not None
|
||
|
||
def _schema_for(self, client: Any, collection: str) -> CollectionSchema | None:
|
||
"""该集合的字段映射;探测失败或缺少必需字段时返回 None(调用方跳过该集合)。"""
|
||
schema = detect_schema(client, collection, cache=self._schemas)
|
||
return schema if schema.usable else None
|
||
|
||
def _visibility_filter(
|
||
self, schemas: dict[str, CollectionSchema], include_internal: bool
|
||
) -> str | None:
|
||
"""可见性过滤表达式:**只在真的存在该字段时**才拼。
|
||
|
||
没有 `visibility` 字段的集合拼上这个表达式会让 Milvus 报
|
||
`field visibility not exist`,整次检索失败(实测)。缺少该字段时返回 None,
|
||
即不做检索层过滤——此时内部资料隔离依赖"入库侧只放对外知识",
|
||
这一代价在 `app/core/knowledge_schema.py` 的模块说明里有记录。
|
||
"""
|
||
if include_internal:
|
||
return None
|
||
holders = [s for s in schemas.values() if s.has("visibility")]
|
||
if not holders:
|
||
return None
|
||
# 只要有一个集合带该字段就过滤;不带该字段的集合由各自调用处跳过表达式的拼接
|
||
return 'visibility == "public"'
|
||
|
||
async def search(
|
||
self,
|
||
query: str,
|
||
*,
|
||
collections: Sequence[str] | None = None,
|
||
top_k: int = 5,
|
||
include_internal: bool = False,
|
||
) -> KnowledgeSearchOutcome:
|
||
"""检索知识库。
|
||
|
||
`include_internal=False`(默认)时在 Milvus 侧就过滤掉 `visibility=internal` 的块,
|
||
内部资料不进入面向客户的答案——这是检索层的硬隔离,不依赖提示词约束。
|
||
"""
|
||
text = query.strip()
|
||
if not text:
|
||
return KnowledgeSearchOutcome(reason="empty_query")
|
||
if not self.available:
|
||
return KnowledgeSearchOutcome(degraded=True, reason="vector_backend_unavailable")
|
||
|
||
client, embedder = self._client, self._embedder
|
||
if client is None or embedder is None:
|
||
# 与 available 重复,但这里需要类型收窄(mypy 不跨属性判断 Optional)
|
||
return KnowledgeSearchOutcome(degraded=True, reason="vector_backend_unavailable")
|
||
try:
|
||
vector = await embedder(text)
|
||
except Exception:
|
||
return KnowledgeSearchOutcome(degraded=True, reason="embedding_failed")
|
||
if not vector:
|
||
return KnowledgeSearchOutcome(degraded=True, reason="embedding_empty")
|
||
|
||
targets = tuple(collections or self._collections)
|
||
# 每个集合**各自探测**字段名:不同环境(甚至同环境不同集合)可能是不同 schema,
|
||
# 用统一的一套字段名去查会让整次检索失败(Milvus 对不存在的字段直接报错)。
|
||
schemas: dict[str, CollectionSchema] = {}
|
||
unusable: list[str] = []
|
||
for collection in targets:
|
||
schema = self._schema_for(client, collection)
|
||
if schema is None:
|
||
unusable.append(collection)
|
||
continue
|
||
schemas[collection] = schema
|
||
if not schemas:
|
||
# 一个集合都用不了:是链路/配置故障,不是"知识库里没有"——如实标记降级
|
||
return KnowledgeSearchOutcome(
|
||
degraded=True, reason="collections_unusable", searched_collections=targets
|
||
)
|
||
expression = self._visibility_filter(schemas, include_internal)
|
||
collected: list[KnowledgeHit] = []
|
||
failures = len(unusable)
|
||
for collection, schema in schemas.items():
|
||
try:
|
||
raw = client.search(
|
||
collection_name=collection,
|
||
data=[vector],
|
||
limit=max(1, min(top_k, 20)),
|
||
output_fields=list(schema.output_fields),
|
||
filter=expression,
|
||
)
|
||
except Exception:
|
||
failures += 1
|
||
continue
|
||
collected.extend(self._parse(raw, collection, schema))
|
||
|
||
# 第二路召回:客户确切说出的产品名按字面取回。只在向量结果不够确定时介入,
|
||
# 否则会把向量已经答对的题顶掉(见 VECTOR_CONFIDENT_SCORE 的说明)。
|
||
best_vector_score = max((hit.score for hit in collected), default=0.0)
|
||
if best_vector_score < VECTOR_CONFIDENT_SCORE:
|
||
collected.extend(
|
||
self._product_keyword_hits(client, schemas, text, expression)
|
||
)
|
||
|
||
# 命中行级子块时把父块(整节)一并带回,供调用方按问句选粒度:
|
||
# 「起投多少」要那一行,「介绍一下」要整节。
|
||
collected.extend(self._parent_hits(client, schemas, collected, expression))
|
||
|
||
if not collected and failures == len(targets) and targets:
|
||
# 所有集合全查失败:是链路故障,不是"知识库里没有"
|
||
return KnowledgeSearchOutcome(
|
||
degraded=True, reason="search_failed", searched_collections=targets
|
||
)
|
||
|
||
collected.sort(key=lambda hit: hit.score, reverse=True)
|
||
# 同一内容可能同时存在于产品手册与问答对里,按 doc_id 去重保留最高分
|
||
deduped: list[KnowledgeHit] = []
|
||
seen: set[str] = set()
|
||
for hit in collected:
|
||
if hit.doc_id in seen:
|
||
continue
|
||
seen.add(hit.doc_id)
|
||
deduped.append(hit)
|
||
|
||
# "整节块"= **有行级子块挂在它下面**的块。FAQ/政策/公司信息的块没有子块,
|
||
# 它们本身就是细粒度答案,不能和产品手册的整节块混为一谈:第一版用"doc_id 不含
|
||
# 两位数字后缀"判断,把 FAQ 块全当成整节块排到最后,直接害得「基金赎回几天到账」
|
||
# 转人工(正确答案被挤出了 top1)。
|
||
section_ids = {
|
||
parent for parent in (self._parent_of(hit.doc_id) for hit in deduped) if parent
|
||
}
|
||
plain = [hit for hit in deduped if hit.doc_id not in section_ids]
|
||
sections = [hit for hit in deduped if hit.doc_id in section_ids]
|
||
selected = plain[: max(1, top_k)]
|
||
if sections:
|
||
# 整节块保底占最后一个名额:它按分数容易被 top_k 截掉,而「介绍一下」这类
|
||
# 概括问句只能靠它拿到整节(实测被截后客户只收到一行"产品期限 90天封闭期")。
|
||
# 放末尾是为了不让它参与 top1/top2 判定:实测它挤到第 2 位时 gap 会从
|
||
# 0.090 掉到 0.076,几乎跌破 0.07 的转人工门槛。
|
||
selected = [*selected[: max(0, top_k - 1)], sections[0]]
|
||
return KnowledgeSearchOutcome(
|
||
hits=tuple(selected),
|
||
degraded=failures > 0,
|
||
reason="partial_collection_failure" if failures else "",
|
||
searched_collections=targets,
|
||
)
|
||
|
||
# ---- 关键字精确召回(字面匹配) ----
|
||
|
||
@staticmethod
|
||
def _overlap_length(left: str, right: str) -> int:
|
||
"""两段文本的最长公共子串长度。
|
||
|
||
用最长公共子串而不是分词:知识库标题是「南方科技有限公司 个人理财产品手册 ·
|
||
2.1 南方季季盈90天」这种没有词边界的长串,任何分词器都得先养一份自定义词典,
|
||
而词典会和手册一起过期。子串匹配不需要词典,手册改版也不会失效。
|
||
"""
|
||
if not left or not right:
|
||
return 0
|
||
previous = [0] * (len(right) + 1)
|
||
best = 0
|
||
for i in range(1, len(left) + 1):
|
||
current = [0] * (len(right) + 1)
|
||
for j in range(1, len(right) + 1):
|
||
if left[i - 1] == right[j - 1]:
|
||
current[j] = previous[j - 1] + 1
|
||
if current[j] > best:
|
||
best = current[j]
|
||
previous = current
|
||
return best
|
||
|
||
def _product_keyword_hits(
|
||
self, client: Any, schemas: dict[str, CollectionSchema], query: str,
|
||
expression: str | None,
|
||
) -> list[KnowledgeHit]:
|
||
"""客户确切说出某个产品名时,按字面把它取出来(兜底用)。
|
||
|
||
两处边界都是实测逼出来的:
|
||
|
||
1. **只对产品集合做**。客户问「季季盈90天的起投金额是多少」,与通用 FAQ 标题
|
||
「基金起投金额是多少?」的最长公共子串有 7 个字;若把 FAQ 也纳入字面匹配,
|
||
它会和真正的产品块一起拿到满分、差距归零,反而又退化成"转人工"。
|
||
2. **命中不能无条件优先**。产品手册的标题里不只有产品名,还有章节名:客户问
|
||
「基金赎回几天到账」时,「五、申购赎回操作指南」那一块与问句也有 6 个字连续
|
||
重合。所以本方法产出什么是一回事,是否采用由调用方按"向量是否已经足够确定"
|
||
决定(见 VECTOR_CONFIDENT_SCORE)。
|
||
|
||
失败一律返回空:关键字路径是**增益**,它坏了不能让整个检索变成故障。
|
||
"""
|
||
lookup = getattr(client, "query", None)
|
||
product_schema = schemas.get(PRODUCT_COLLECTION)
|
||
doc_field = product_schema.resolve("doc_id") if product_schema else None
|
||
content_field = product_schema.resolve("content") if product_schema else None
|
||
if lookup is None or product_schema is None or doc_field is None or content_field is None:
|
||
return [] # 客户端不支持标量查询(如测试替身),或本次没查/不能查产品集合
|
||
try:
|
||
rows = lookup(
|
||
collection_name=PRODUCT_COLLECTION,
|
||
filter=expression,
|
||
output_fields=[doc_field, "title"],
|
||
limit=KEYWORD_SCAN_LIMIT,
|
||
)
|
||
except Exception:
|
||
return []
|
||
|
||
matched_ids = [
|
||
str(row.get(doc_field) or "")
|
||
for row in (rows if isinstance(rows, list) else [])
|
||
if isinstance(row, dict)
|
||
and self._overlap_length(query, str(row.get("title") or "")) >= MIN_KEYWORD_OVERLAP
|
||
]
|
||
matched_ids = [doc_id for doc_id in matched_ids if doc_id]
|
||
if not matched_ids:
|
||
return []
|
||
|
||
quoted = ", ".join(f'"{doc_id}"' for doc_id in matched_ids)
|
||
try:
|
||
details = lookup(
|
||
collection_name=PRODUCT_COLLECTION,
|
||
filter=f"{doc_field} in [{quoted}]",
|
||
output_fields=list(product_schema.output_fields),
|
||
limit=len(matched_ids),
|
||
)
|
||
except Exception:
|
||
return []
|
||
|
||
hits: list[KnowledgeHit] = []
|
||
for row in details if isinstance(details, list) else []:
|
||
if not isinstance(row, dict):
|
||
continue
|
||
content = str(row.get(content_field) or "")
|
||
if not content:
|
||
continue
|
||
hits.append(self._hit_from_row(row, score=KEYWORD_MATCH_SCORE,
|
||
schema=product_schema))
|
||
# 一个产品名可能命中多个块(产品概览、费率表各一块):全都保留,
|
||
# 是不是"只有一个明确候选"交给上层的 gap 判定,这里不替它做选择。
|
||
return hits
|
||
|
||
def _parent_hits(
|
||
self, client: Any, schemas: dict[str, CollectionSchema], hits: list[KnowledgeHit],
|
||
expression: str | None,
|
||
) -> list[KnowledgeHit]:
|
||
"""把命中到的行级子块的**父块**一并带回来。
|
||
|
||
行级子块让「起投多少」拿到了聚焦答案,但客户问「介绍一下」时会被某一行抢答
|
||
(实测返回了"产品期限 90天封闭期",而客户要的是整个产品的介绍)。父块是同一
|
||
产品的整节内容,一并带回来,由调用方按问句自己选粒度——检索层不猜客户想听多细。
|
||
|
||
分数按子块的 0.9 折算:既排在子块之后(不抢聚焦答案),又不会把差距压到转人工
|
||
门槛之下(0.869 折算成 0.782,与子块差 0.087,仍在 0.07 之上)。
|
||
|
||
失败一律返回空:父块是**补充**,取不到不影响子块结果。
|
||
"""
|
||
# 只取"得分最高的那个子块"的父块:整节只可能来自一个产品,把命中到的子块的父块
|
||
# 全带上只会挤占 top_k 名额(实测带上多个后,父块反而被截断在门外、整节拿不到)。
|
||
best_child = max(
|
||
(hit for hit in hits if self._parent_of(hit.doc_id) is not None),
|
||
key=lambda hit: hit.score,
|
||
default=None,
|
||
)
|
||
lookup = getattr(client, "query", None)
|
||
if best_child is None:
|
||
return []
|
||
parent_id = self._parent_of(best_child.doc_id)
|
||
if not parent_id or lookup is None:
|
||
return []
|
||
parent_scores = {parent_id: best_child.score * PARENT_SCORE_RATIO}
|
||
|
||
quoted = ", ".join(f'"{parent_id}"' for parent_id in parent_scores)
|
||
found: list[KnowledgeHit] = []
|
||
for collection, schema in schemas.items():
|
||
doc_field = schema.resolve("doc_id")
|
||
content_field = schema.resolve("content")
|
||
if doc_field is None or content_field is None:
|
||
continue # 该集合缺少必需字段(探测阶段已记为不可用)
|
||
try:
|
||
rows = lookup(
|
||
collection_name=collection,
|
||
filter=self._anded(expression, f"{doc_field} in [{quoted}]"),
|
||
output_fields=list(schema.output_fields),
|
||
limit=KEYWORD_SCAN_LIMIT,
|
||
)
|
||
except Exception:
|
||
continue # 某个集合查不到不影响其它集合
|
||
for row in rows if isinstance(rows, list) else []:
|
||
if not isinstance(row, dict):
|
||
continue
|
||
doc_id = str(row.get(doc_field) or "")
|
||
content = str(row.get(content_field) or "")
|
||
if not content or doc_id not in parent_scores:
|
||
continue
|
||
found.append(self._hit_from_row(row, score=parent_scores[doc_id], schema=schema))
|
||
return found
|
||
|
||
@staticmethod
|
||
def _parent_of(doc_id: str) -> str | None:
|
||
"""行级子块的父块 doc_id;整节块返回 None。
|
||
|
||
判据是"末段恰为 2 位数字"(PROD-007-04 → PROD-007),**不是**"含连字符":
|
||
整节块自己的编号就形如 PROD-901,用连字符判断会把整节块误判成子块。
|
||
"""
|
||
head, _, tail = doc_id.rpartition("-")
|
||
if head and tail.isdigit() and len(tail) == 2:
|
||
return head
|
||
return None
|
||
|
||
@staticmethod
|
||
def _anded(expression: str | None, extra: str) -> str:
|
||
"""把可见性过滤与 doc_id 过滤合成一个 Milvus 表达式。"""
|
||
return f"({expression}) and ({extra})" if expression else extra
|
||
|
||
@staticmethod
|
||
def _hit_from_row(row: Any, *, score: float, schema: CollectionSchema) -> KnowledgeHit:
|
||
"""把一行 Milvus 结果折成 `KnowledgeHit`。
|
||
|
||
字段名**按该集合探测出来的映射**取(可能是 `doc_id`/`content`,也可能是
|
||
`knowledge_id`/`snippet`);该集合没有的字段留空字符串——**不编造内容**,
|
||
`reference_title` 会因此退回纯标题,来源引用退化但不失真。
|
||
"""
|
||
def value(logical: str) -> str:
|
||
field_name = schema.resolve(logical)
|
||
if field_name is None:
|
||
return ""
|
||
return str(row.get(field_name) or "")
|
||
|
||
return KnowledgeHit(
|
||
doc_id=value("doc_id"),
|
||
title=value("title"),
|
||
content=value("content"),
|
||
score=score,
|
||
source_file=value("source_file"),
|
||
visibility=value("visibility") or "public",
|
||
doc_no=value("doc_no"),
|
||
version=value("version"),
|
||
chapter=value("chapter"),
|
||
)
|
||
|
||
@staticmethod
|
||
def _parse(raw: Any, collection: str, schema: CollectionSchema) -> list[KnowledgeHit]:
|
||
"""把 pymilvus 的 `[[{id, distance, entity}]]` 折叠成命中列表(纯函数,不抛异常)。"""
|
||
hits: list[KnowledgeHit] = []
|
||
groups = raw if isinstance(raw, (list, tuple)) else [raw]
|
||
for group in groups:
|
||
rows = group if isinstance(group, (list, tuple)) else [group]
|
||
for row in rows:
|
||
entity = row.get("entity") if isinstance(row, dict) else None
|
||
if not isinstance(entity, dict):
|
||
continue
|
||
content_field = schema.resolve("content")
|
||
content = str(entity.get(content_field) or "") if content_field else ""
|
||
if not content:
|
||
continue # 没有正文的命中无法作为答案来源,直接丢弃而不是猜造
|
||
hits.append(KnowledgeSearchService._hit_from_row(
|
||
entity, score=float(row.get("distance") or 0.0), schema=schema))
|
||
return hits
|