feat: 接入 BGE-M3 本地嵌入默认配置 + 新增 Qwen3-Reranker 重排
- .env.example 嵌入默认改为本地 BGE-M3（EMBEDDING_PROVIDER=local, dim=1024），NAS 全新部署即用 - 新增 app/services/reranker.py：基于 Ollama /api/rerank 的 cross-encoder 重排服务 - retriever 在 chunk 候选阶段接入语义精排，调用失败优雅降级为 RRF 顺序 - docker-compose 拉取 qwen3-reranker:0.6b，OLLAMA_MAX_LOADED_MODELS 提至 3 - .gitignore 排除 .workbuddy/ 与 .trae/（防止误提交项目记忆与 IDE spec） - 新增 reranker 单元与集成测试，全量测试 529 passed
This commit is contained in:
@@ -21,6 +21,12 @@ class Settings(BaseSettings):
|
||||
ollama_model: str = "qwen2.5:1.5b" # 备选: qwen2.5:3b
|
||||
ollama_embedding_model: str = "bge-m3" # embedding_provider=local 时使用的嵌入模型
|
||||
|
||||
# 重排模型（cross-encoder，经 Ollama /api/rerank，对 chunk 候选做最终精排）
|
||||
# 关闭时检索链路退化为纯 RRF 融合结果，行为不变
|
||||
reranker_enabled: bool = False
|
||||
reranker_model: str = "qwen3-reranker:0.6b"
|
||||
reranker_timeout: float = 30.0
|
||||
|
||||
# Qdrant
|
||||
qdrant_host: str = "localhost"
|
||||
qdrant_port: int = 6333
|
||||
|
||||
+26
-3
@@ -16,7 +16,7 @@ from qdrant_client import models
|
||||
from app.config import settings
|
||||
from app.core.embeddings import EmbeddingService, create_embedding_service
|
||||
from app.core.query_parser import QueryParser, RouteDecision
|
||||
from app.core.ranker import finalize, rrf_fuse
|
||||
from app.core.ranker import rrf_fuse
|
||||
from app.core.result_summarizer import ResultSummarizer
|
||||
from app.core.sparse import SparseEncoder
|
||||
from app.models.knowledge import load_taxonomy
|
||||
@@ -30,6 +30,7 @@ from app.services.qdrant import (
|
||||
SPARSE_COLLECTIONS,
|
||||
QdrantService,
|
||||
)
|
||||
from app.services.reranker import RerankerService, create_reranker_service
|
||||
|
||||
logger = structlog.get_logger()
|
||||
|
||||
@@ -49,6 +50,7 @@ class Retriever:
|
||||
embedding: EmbeddingService | None = None,
|
||||
sparse_encoder: SparseEncoder | None = None,
|
||||
result_summarizer: ResultSummarizer | None = None,
|
||||
reranker: RerankerService | None = None,
|
||||
) -> None:
|
||||
self.qdrant = qdrant or QdrantService()
|
||||
self.query_parser = query_parser or QueryParser(
|
||||
@@ -58,6 +60,7 @@ class Retriever:
|
||||
self.embedding = embedding or create_embedding_service()
|
||||
self.sparse_encoder = sparse_encoder or SparseEncoder()
|
||||
self.result_summarizer = result_summarizer or ResultSummarizer()
|
||||
self.reranker = reranker if reranker is not None else create_reranker_service()
|
||||
|
||||
async def search(self, request: SearchRequest) -> SearchResponse:
|
||||
"""分层检索主流程"""
|
||||
@@ -77,7 +80,7 @@ class Retriever:
|
||||
# L1 无候选文档 → 全库 chunk 兜底
|
||||
chunk_hits = await self._search_collection(COLLECTION_CHUNKS, dense, sparse, settings.retrieval_top_k, None)
|
||||
logger.info("L1 无命中，全库 chunk 兜底", hits=len(chunk_hits))
|
||||
hits = self._to_hits(finalize(chunk_hits, self._final_k(request)))
|
||||
hits = self._to_hits(await self._rerank(query_text, chunk_hits, self._final_k(request)))
|
||||
return await self._build_response(request, route, hits, fallback=True)
|
||||
|
||||
doc_ids = _unique((p.payload or {}).get("doc_id") for p in l1_hits)
|
||||
@@ -123,7 +126,7 @@ class Retriever:
|
||||
)
|
||||
logger.info("chunk 检索完成", hits=len(chunk_hits))
|
||||
|
||||
final_points = finalize(rrf_fuse([chunk_hits]), self._final_k(request))
|
||||
final_points = await self._rerank(query_text, chunk_hits, self._final_k(request))
|
||||
hits = self._to_hits(final_points)
|
||||
return await self._build_response(request, route, hits, fallback=route.fallback)
|
||||
|
||||
@@ -145,6 +148,26 @@ class Retriever:
|
||||
"""最终返回数：请求指定优先，否则用配置默认值"""
|
||||
return request.top_k or settings.retrieval_final_k
|
||||
|
||||
async def _rerank(
|
||||
self, query_text: str, points: list[models.ScoredPoint], final_k: int
|
||||
) -> list[models.ScoredPoint]:
|
||||
"""对 chunk 候选池做语义精排（cross-encoder）
|
||||
|
||||
重排未启用、候选为空或调用异常时，退化为 RRF 融合的原始顺序（取前 final_k），
|
||||
保证检索链路在任何情况下都不因重排失败而中断。
|
||||
"""
|
||||
if self.reranker is None or not points:
|
||||
return points[:final_k]
|
||||
documents = [(p.payload or {}).get("text", "") for p in points]
|
||||
try:
|
||||
scores = await self.reranker.rerank(query_text, documents)
|
||||
except Exception:
|
||||
logger.warning("重排调用失败，退化为 RRF 融合顺序", exc_info=True)
|
||||
return points[:final_k]
|
||||
# 按相关性分数降序重排，写回 score 字段为相关性分数
|
||||
ordered = sorted(range(len(points)), key=lambda i: scores[i], reverse=True)
|
||||
return [points[i].model_copy(update={"score": scores[i]}) for i in ordered[:final_k]]
|
||||
|
||||
@staticmethod
|
||||
def _to_hits(points: list[models.ScoredPoint]) -> list[SearchHit]:
|
||||
"""chunk 点组装为 SearchHit，字段取自 chunk payload（doc_summary 仅上下文标注）"""
|
||||
|
||||
@@ -0,0 +1,79 @@
|
||||
"""语义重排服务（cross-encoder reranker）
|
||||
|
||||
调用 Ollama 的 `/api/rerank` 端点（Qwen3-Reranker 等重排模型），对检索召回的
|
||||
chunk 候选池做精排，提升最终 Top-K 的相关性。相比纯向量/RRF 融合，cross-encoder
|
||||
以 (query, doc) 联合编码，能捕捉字词不匹配的语义关联，典型带来约 10% 的精度提升。
|
||||
|
||||
工厂 `create_reranker_service()` 按 `settings.reranker_enabled` 返回实例或 None：
|
||||
关闭时上层检索链路退化为原有的 RRF 融合结果，行为完全不变。
|
||||
"""
|
||||
|
||||
from typing import Protocol, runtime_checkable
|
||||
|
||||
import httpx
|
||||
import structlog
|
||||
|
||||
from app.config import settings
|
||||
|
||||
logger = structlog.get_logger()
|
||||
|
||||
|
||||
@runtime_checkable
|
||||
class RerankerService(Protocol):
|
||||
"""重排服务统一接口"""
|
||||
|
||||
async def rerank(self, query: str, documents: list[str]) -> list[float]:
|
||||
"""对候选文档按与 query 的相关性打分
|
||||
|
||||
Args:
|
||||
query: 查询文本
|
||||
documents: 候选文档文本列表（按输入顺序对齐）
|
||||
|
||||
Returns:
|
||||
与 documents 等长的相关性分数列表（分数越高越相关）
|
||||
"""
|
||||
...
|
||||
|
||||
|
||||
class OllamaRerankerService:
|
||||
"""基于 Ollama /api/rerank 的重排服务"""
|
||||
|
||||
def __init__(self, base_url: str, model: str, timeout: float = 30.0) -> None:
|
||||
self.base_url = base_url.rstrip("/")
|
||||
self.model = model
|
||||
self.timeout = timeout
|
||||
|
||||
async def rerank(self, query: str, documents: list[str]) -> list[float]:
|
||||
if not documents:
|
||||
return []
|
||||
url = f"{self.base_url}/api/rerank"
|
||||
payload = {"model": self.model, "query": query, "documents": documents}
|
||||
async with httpx.AsyncClient(timeout=self.timeout) as client:
|
||||
resp = await client.post(url, json=payload)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
|
||||
results = data.get("results", [])
|
||||
# 按 Ollama 返回的 index 映射 relevance_score；缺失项补 0，保证输出与输入等长
|
||||
scores: list[float] = [0.0] * len(documents)
|
||||
for item in results:
|
||||
idx = item.get("index")
|
||||
score = float(item.get("relevance_score", 0.0))
|
||||
if idx is not None and 0 <= idx < len(documents):
|
||||
scores[idx] = score
|
||||
logger.debug("重排完成", model=self.model, candidates=len(documents))
|
||||
return scores
|
||||
|
||||
|
||||
def create_reranker_service() -> RerankerService | None:
|
||||
"""按 settings.reranker_enabled 创建重排服务实例
|
||||
|
||||
关闭时返回 None，上层据此跳过精排、退化为 RRF 融合结果。
|
||||
"""
|
||||
if not settings.reranker_enabled:
|
||||
return None
|
||||
return OllamaRerankerService(
|
||||
base_url=settings.ollama_base_url,
|
||||
model=settings.reranker_model,
|
||||
timeout=settings.reranker_timeout,
|
||||
)
|
||||
Reference in New Issue
Block a user