改用 httpx 实现所有 client 交互

This commit is contained in:
李如威 2025-12-19 11:12:29 +08:00
parent 4345dfb9cb
commit fcaac9c0f1
6 changed files with 258 additions and 158 deletions

View File

@ -8,5 +8,4 @@ scikit-learn
aiofiles
pillow
loguru
aiohttp
elasticsearch>=8.0.0,<9.0.0
httpx

191
src/pipeline/core/es.py Normal file
View File

@ -0,0 +1,191 @@
import httpx
from datetime import datetime, timezone
from src.pipeline.config import config
from src.pipeline.core.utils import logger
class AsyncES:
def __init__(
self,
timeout: float = 30.0,
max_connections: int = 50,
max_keepalive: int = 10,
):
self.base = f"{config['es_host']}:{config['es_port']}".rstrip("/")
self.auth = (config["es_user"], config["es_password"])
self.client = httpx.AsyncClient(
http2=False,
trust_env=False,
timeout=httpx.Timeout(timeout),
limits=httpx.Limits(
max_connections=max_connections,
max_keepalive_connections=max_keepalive,
),
auth=self.auth,
headers={
"Content-Type": "application/json",
"Accept": "application/json",
},
)
self.mapping = {
"settings": {
"analysis": {
"analyzer": {
"ik_smart": {
"tokenizer": "ik_smart",
}
}
}
},
"mappings": {
"properties": {
"type": {"type": "keyword"},
"doc_id": {"type": "keyword"},
"title": {"type": "text", "analyzer": "ik_smart"},
"created_at": {"type": "date"},
"content": {"type": "text", "analyzer": "ik_smart"},
"embedding": {
"type": "dense_vector",
"dims": config['embedding_dims'],
"index": True,
"similarity": "cosine",
},
}
},
}
# -------------------------
# Index
# -------------------------
async def create_index(self, index: str) -> bool:
url = f"{self.base}/{index}"
try:
resp = await self.client.put(url, json=self.mapping)
if resp.status_code in (200, 201):
return True
if resp.status_code == 400 and "resource_already_exists" in resp.text:
return True
logger.error(f"Create index failed: {resp.text}")
return False
except Exception:
logger.exception("Create index error")
return False
# -------------------------
# Index document
# -------------------------
async def add_doc(
self,
index: str,
es_id: str,
doc_id: str,
content: str,
embedding: list,
title: str = "",
doc_type: str = "",
created_at: datetime | None = None,
):
if created_at is None:
created_at = datetime.now(timezone.utc)
doc = {
"doc_id": doc_id,
"title": title,
"type": doc_type,
"content": content,
"embedding": embedding,
"created_at": created_at.isoformat(),
}
url = f"{self.base}/{index}/_doc/{es_id}"
try:
resp = await self.client.put(url, json=doc)
resp.raise_for_status()
except Exception:
logger.exception("Index document failed")
# -------------------------
# Hybrid search
# -------------------------
async def hybrid_search_es(
self,
index: str,
query_text: str,
query_vector: list,
top_k: int = 5,
doc_type: str | None = None,
):
must_clauses = [
{
"multi_match": {
"query": query_text,
"fields": ["content", "title"],
}
}
]
if doc_type:
must_clauses.append({"term": {"type": doc_type}})
body = {
"size": top_k,
"query": {
"function_score": {
"query": {"bool": {"must": must_clauses}},
"functions": [
{
"script_score": {
"script": {
"source": ("cosineSimilarity(params.query_vector, 'embedding') + 1.0"),
"params": {"query_vector": query_vector},
}
}
}
],
"boost_mode": "sum",
}
},
}
url = f"{self.base}/{index}/_search"
resp = await self.client.post(url, json=body)
resp.raise_for_status()
hits = resp.json()["hits"]["hits"]
return [
{
"es_id": h["_id"],
"doc_id": h["_source"].get("doc_id"),
"title": h["_source"].get("title"),
"type": h["_source"].get("type"),
"created_at": h["_source"].get("created_at"),
"score": h["_score"],
"content": h["_source"].get("content"),
}
for h in hits
]
async def close(self):
await self.client.aclose()
client: AsyncES | None = None
async def init_client():
global client
client = AsyncES()
async def close_client():
await client.close()

View File

@ -1,125 +0,0 @@
from codecs import ignore_errors
from datetime import datetime, timezone
from venv import logger
from src.pipeline.config import config
from elasticsearch import AsyncElasticsearch, ApiError
class AsyncES:
def __init__(self):
self.base = f"{config['es_host']}:{config['es_port']}"
self.auth = (config["es_user"], config["es_password"])
self.client = AsyncElasticsearch(self.base, basic_auth=self.auth)
self.mapping = {
"settings": {
"analysis": {
"analyzer": {
"ik_smart": {
"tokenizer": "ik_smart",
},
}
}
},
"mappings": {
"properties": {
"type": {"type": "keyword"}, # 分类/类型
"doc_id": {"type": "keyword"}, # 业务文档 ID便于聚合
"title": {"type": "text", "analyzer": "ik_smart"}, # 文档标题
"created_at": {"type": "date"}, # 创建时间
"content": {"type": "text", "analyzer": "ik_smart"}, # 切片内容
"embedding": { # 向量字段
"type": "dense_vector",
"dims": 1024, # 必须和 embedding 模型的纬度一样
"index": True,
"similarity": "cosine",
},
},
},
}
async def create_index(self, index: str):
"""
创建 index知识库
"""
try:
await self.client.indices.create(
index=index,
body=self.mapping,
)
return True
except ApiError as e:
logger.error(e.message)
return False
async def add_doc(
self,
index: str,
es_id: str,
doc_id: str,
content: str,
embedding: list,
title: str = "",
doc_type: str = "",
created_at: datetime = None,
):
"""
写入文档普通文本
"""
if created_at is None:
created_at = datetime.now(timezone.utc)
doc = {
"doc_id": doc_id,
"title": title,
"type": doc_type,
"content": content,
"embedding": embedding,
"created_at": created_at.isoformat(),
}
await self.client.index(index=index, id=es_id, document=doc)
async def hybrid_search_es(self, index: str, query_text: str, query_vector: list, top_k: int = 5, doc_type: str = None):
"""
Elasticsearch 内执行 BM25 + 向量混合检索
"""
# 基础查询:中文 BM25 匹配 content 或 title
must_clauses = [{"multi_match": {"query": query_text, "fields": ["content", "title"]}}]
# 可选类型过滤
if doc_type:
must_clauses.append({"term": {"type": doc_type}})
body = {
"size": top_k,
"query": {
"function_score": {
"query": {"bool": {"must": must_clauses}},
"functions": [
{
"script_score": {
"script": {
# cosineSimilarity 返回 [-1,1],加 1 保证非负
"source": "cosineSimilarity(params.query_vector, 'embedding') + 1.0",
"params": {"query_vector": query_vector},
}
}
}
],
"boost_mode": "sum", # BM25 得分 + 向量相似度相加
}
},
}
res = await self.client.search(index=index, body=body)
return [
{
"es_id": hit["_id"],
"doc_id": hit["_source"].get("doc_id"),
"title": hit["_source"].get("title"),
"type": hit["_source"].get("type"),
"created_at": hit["_source"].get("created_at"),
"score": hit["_score"],
"content": hit["_source"].get("content"),
}
for hit in res["hits"]["hits"]
]
es_client = AsyncES()

View File

@ -1,29 +1,60 @@
import httpx
from src.pipeline.config import config
import aiohttp
from src.pipeline.core.utils import logger
_embedding_session: aiohttp.ClientSession | None = None
async def get_embedding(text: str, timeout: int = 30):
global _embedding_session
if _embedding_session is None or _embedding_session.closed:
_embedding_session = aiohttp.ClientSession(
timeout=aiohttp.ClientTimeout(total=timeout)
)
async with _embedding_session.post(
config["embedding_api_host"],
json={
"model": config["embedding_model"],
"input": text,
},
class AsyncLLm:
def __init__(
self,
timeout: float = 30.0,
max_connections: int = 100,
max_keepalive: int = 20,
):
self.embedding_api = config["embedding_api_host"].rstrip("/") + "/embeddings"
self.embedding_model = config["embedding_model"]
self.api_key = config["embedding_api_key"]
logger.debug(self.embedding_api)
self.embedding_client = httpx.AsyncClient(
http2=False,
trust_env=False,
timeout=httpx.Timeout(timeout),
limits=httpx.Limits(
max_connections=max_connections,
max_keepalive_connections=max_keepalive,
),
headers={
"Content-Type": "application/json",
"Authorization": f"Bearer {config['embedding_api_key']}",
"Authorization": f"Bearer {self.api_key}",
},
) as resp:
if resp.status >= 400:
)
async def embedding(self, text: str) -> list[float]:
try:
resp = await self.embedding_client.post(
self.embedding_api,
json={"model": self.embedding_model, "input": text},
)
resp.raise_for_status()
data = resp.json()
return data["data"][0]["embedding"]
except httpx.HTTPStatusError as e:
logger.error(e)
logger.error(f"Embedding HTTP error: {e.response.text}")
except Exception as e:
logger.exception("Embedding request failed")
return []
data = await resp.json()
return data["data"][0]["embedding"]
async def close(self):
await self.embedding_client.aclose()
client: AsyncLLm | None = None
async def init_client():
global client
client = AsyncLLm()
async def close_client():
await client.close()

View File

@ -1,8 +1,8 @@
import uuid
from src.pipeline.core.pocket_flow import AsyncBatchNode
from src.pipeline.core.utils import fixed_size_chunk, load_document, logger
from src.pipeline.core.llm import get_embedding
from src.pipeline.core.es_client import es_client
from src.pipeline.core import llm
from src.pipeline.core import es
import re
@ -49,7 +49,7 @@ class EmbeddingDocumentsNode(AsyncBatchNode):
:param document: {text, file_name}
"""
logger.debug(f"开始 embedding: {document["text"].strip()[:10]}...")
res = {**document, "embedding": await get_embedding(document["text"])}
res = {**document, "embedding": await llm.client.embedding(document["text"])}
logger.debug(f"结束 embedding: {document["text"].strip()[:10]}...")
return res
@ -95,7 +95,7 @@ class ReadDocumentNode(AsyncBatchNode):
class WriteToElasticsearchNode(AsyncBatchNode):
async def prep_async(self, shared):
index = shared["index"]
await es_client.create_index(index)
await es.client.create_index(index)
return [
{
"index": index,
@ -110,7 +110,7 @@ class WriteToElasticsearchNode(AsyncBatchNode):
]
async def exec_async(self, prep_res):
await es_client.add_doc(**prep_res)
await es.client.add_doc(**prep_res)
return True
async def post_async(self, shared, prep_res, exec_res):

View File

@ -3,19 +3,20 @@ import json
from src.pipeline.core.nodes import ReadDocumentNode, ChunkDocumentsNode, EmbeddingDocumentsNode, WriteToElasticsearchNode
from src.pipeline.core.pocket_flow import AsyncFlow
from src.pipeline.core.utils import logger
from src.pipeline.core import llm, es
@pytest.mark.asyncio
async def test_embedding():
await llm.init_client()
await es.init_client()
print("\n\ntest_embedding:\n")
shared = {
"files": [
# "./files/带图片的.pdf",
# "./files/大白智问-API接入文档-V1.2.2.pdf",
"./files/山海经01.txt",
# "./files/deepsearch状态.txt",
],
"documents": [],
"index": "test_kb",
@ -31,3 +32,6 @@ async def test_embedding():
await flow.run_async(shared)
logger.debug(json.dumps([{**x, "embedding":x["embedding"][:4]} for x in shared["documents"]], indent=4, ensure_ascii=False))
await llm.close_client()
await es.close_client()