rag_server/applications/utils/milvus/search.py

import asyncio
from typing import List, Optional, Dict, Any, Union


class MilvusSearcher:

    output_fields = [
        "doc_id",
        "chunk_id",
        "summary",
        "topic",
        "domain",
        "task_type",
        "keywords",
        "concepts",
        "questions",
        "entities",
        "tokens",
        "topic_purity",
    ]

    def __init__(self, milvus_pool):
        self.milvus_pool = milvus_pool

    # 通过向量匹配
    async def vector_search(
        self,
        query_vec: List[float],
        anns_field: str = "vector_text",
        limit: int = 5,
        expr: Optional[str] = None,
        search_params: Optional[Dict[str, Any]] = None,
    ):
        """向量搜索，可选过滤"""
        if search_params is None:
            search_params = {"metric_type": "COSINE", "params": {"ef": 64}}
        return await asyncio.to_thread(
            self.milvus_pool.search,
            data=[query_vec],
            anns_field=anns_field,
            param=search_params,
            limit=limit,
            expr=expr,
            output_fields=self.output_fields,
        )

    # 通过doc_id + chunk_id 获取数据
    async def get_by_doc_and_chunk(self, doc_id: str, chunk_id: int):
        expr = f'doc_id == "{doc_id}" and chunk_id == {chunk_id}'
        return await asyncio.to_thread(
            self.milvus_pool.query,
            expr=expr,
            output_fields=self.output_fields,
        )

    # 只按 metadata 条件查询
    async def filter_search(self, filters: Dict[str, Union[str, int, float]]):
        exprs = []
        for k, v in filters.items():
            if isinstance(v, str):
                exprs.append(f'{k} == "{v}"')
            else:
                exprs.append(f"{k} == {v}")
        expr = " and ".join(exprs)
        return await asyncio.to_thread(
            self.milvus_pool.query,
            expr=expr,
            output_fields=self.output_fields,
        )

    # 混合搜索（向量 + metadata）
    async def hybrid_search(
        self,
        query_vec: List[float],
        anns_field: str = "vector_text",
        limit: int = 5,
        filters: Optional[Dict[str, Union[str, int, float]]] = None,
    ):
        expr = None
        if filters:
            parts = []
            for k, v in filters.items():
                if isinstance(v, str):
                    parts.append(f'{k} == "{v}"')
                else:
                    parts.append(f"{k} == {v}")
            expr = " and ".join(parts)

        return await self.vector_search(
            query_vec=query_vec, anns_field=anns_field, limit=limit, expr=expr
        )

    # 通过主键获取milvus数据
    async def get_by_id(self, pk: int):
        return await asyncio.to_thread(
            self.milvus_pool.query,
            expr=f"id == {pk}",
            output_fields=self.output_fields,
        )

    async def search_by_strategy(
        self,
        query_vec: List[float],
        weight_map: Dict,
        limit: int = 5,
        expr: Optional[str] = None,
        search_params: Optional[Dict[str, Any]] = None,
    ):
        async def _sub_search(vec, field):
            return await asyncio.to_thread(
                self.milvus_pool.search,
                data=[vec],
                anns_field=field,
                param={"metric_type": "COSINE", "params": {"ef": 64}},
                limit=limit,
                expr=expr,
                output_fields=self.output_fields,
            )

        tasks = {field: _sub_search(query_vec, field) for field in weight_map.keys()}
        results = await asyncio.gather(*tasks.values())

        scores = {}
        for (field, weight), res in zip(weight_map.items(), results):
            for hit in res[0]:
                key = (hit.id, hit.entity.get("doc_id"), hit.entity.get("chunk_id"))
                sim_score = 1 - hit.distance
                scores[key] = scores.get(key, 0) + weight * sim_score

        ranked = sorted(scores.items(), key=lambda x: x[1], reverse=True)[:limit]

        return [
            {"pk": k[0], "doc_id": k[1], "chunk_id": k[2], "score": v}
            for k, v in ranked
        ]