|
|
|
|
@ -0,0 +1,229 @@
|
|
|
|
|
import os
|
|
|
|
|
import uuid
|
|
|
|
|
import time
|
|
|
|
|
import jieba
|
|
|
|
|
from fastapi import FastAPI, Form, HTTPException
|
|
|
|
|
from openai import OpenAI
|
|
|
|
|
from gensim.models import KeyedVectors
|
|
|
|
|
from contextlib import asynccontextmanager
|
|
|
|
|
from TtsConfig import *
|
|
|
|
|
from WxMini.OssUtil import upload_mp3_to_oss
|
|
|
|
|
from WxMini.TtsUtil import TTS
|
|
|
|
|
from WxMini.Milvus.Utils.MilvusCollectionManager import MilvusCollectionManager
|
|
|
|
|
from WxMini.Milvus.Utils.MilvusConnectionPool import *
|
|
|
|
|
from WxMini.Milvus.Config.MulvusConfig import *
|
|
|
|
|
|
|
|
|
|
import jieba.analyse
|
|
|
|
|
|
|
|
|
|
# 提取用户输入的关键词
|
|
|
|
|
def extract_keywords(text, topK=3):
|
|
|
|
|
"""
|
|
|
|
|
提取用户输入的关键词
|
|
|
|
|
:param text: 用户输入的文本
|
|
|
|
|
:param topK: 返回的关键词数量
|
|
|
|
|
:return: 关键词列表
|
|
|
|
|
"""
|
|
|
|
|
keywords = jieba.analyse.extract_tags(text, topK=topK)
|
|
|
|
|
return keywords
|
|
|
|
|
|
|
|
|
|
# 构建查询条件
|
|
|
|
|
def build_query_expr(session_id, keywords):
|
|
|
|
|
"""
|
|
|
|
|
构建查询条件
|
|
|
|
|
:param session_id: 用户会话 ID
|
|
|
|
|
:param keywords: 关键词列表
|
|
|
|
|
:return: 查询条件表达式
|
|
|
|
|
"""
|
|
|
|
|
# 基础条件：过滤 session_id
|
|
|
|
|
expr = f"session_id == '{session_id}'"
|
|
|
|
|
# 添加关键词条件
|
|
|
|
|
if keywords:
|
|
|
|
|
keyword_conditions = []
|
|
|
|
|
for keyword in keywords:
|
|
|
|
|
if len(keyword) > 1: # 过滤过短的关键词
|
|
|
|
|
# 使用前缀匹配
|
|
|
|
|
keyword_conditions.append(f"user_input like '{keyword}%'")
|
|
|
|
|
if keyword_conditions:
|
|
|
|
|
expr += " and (" + " or ".join(keyword_conditions) + ")"
|
|
|
|
|
return expr
|
|
|
|
|
|
|
|
|
|
# 初始化 Word2Vec 模型
|
|
|
|
|
model_path = MS_MODEL_PATH
|
|
|
|
|
model = KeyedVectors.load_word2vec_format(model_path, binary=False, limit=MS_MODEL_LIMIT)
|
|
|
|
|
print(f"模型加载成功，词向量维度: {model.vector_size}")
|
|
|
|
|
|
|
|
|
|
# 初始化 Milvus 连接池
|
|
|
|
|
milvus_pool = MilvusConnectionPool(host=MS_HOST, port=MS_PORT, max_connections=MS_MAX_CONNECTIONS)
|
|
|
|
|
|
|
|
|
|
# 初始化集合管理器
|
|
|
|
|
collection_name = MS_COLLECTION_NAME
|
|
|
|
|
collection_manager = MilvusCollectionManager(collection_name)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# 将文本转换为嵌入向量
|
|
|
|
|
def text_to_embedding(text):
|
|
|
|
|
words = jieba.lcut(text) # 使用 jieba 分词
|
|
|
|
|
print(f"文本: {text}, 分词结果: {words}")
|
|
|
|
|
embeddings = [model[word] for word in words if word in model]
|
|
|
|
|
print(f"有效词向量数量: {len(embeddings)}")
|
|
|
|
|
if embeddings:
|
|
|
|
|
avg_embedding = sum(embeddings) / len(embeddings)
|
|
|
|
|
print(f"生成的平均向量: {avg_embedding[:5]}...") # 打印前 5 维
|
|
|
|
|
return avg_embedding
|
|
|
|
|
else:
|
|
|
|
|
print("未找到有效词，返回零向量")
|
|
|
|
|
return [0.0] * model.vector_size
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# 使用 Lifespan Events 处理应用启动和关闭逻辑
|
|
|
|
|
@asynccontextmanager
|
|
|
|
|
async def lifespan(app: FastAPI):
|
|
|
|
|
# 应用启动时加载集合到内存
|
|
|
|
|
collection_manager.load_collection()
|
|
|
|
|
print(f"集合 '{collection_name}' 已加载到内存。")
|
|
|
|
|
yield
|
|
|
|
|
# 应用关闭时释放连接池
|
|
|
|
|
milvus_pool.close()
|
|
|
|
|
print("Milvus 连接池已关闭。")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# 初始化 FastAPI 应用
|
|
|
|
|
app = FastAPI(lifespan=lifespan)
|
|
|
|
|
|
|
|
|
|
# 初始化 OpenAI 客户端
|
|
|
|
|
client = OpenAI(
|
|
|
|
|
api_key=MODEL_API_KEY,
|
|
|
|
|
base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
# 设置相似度阈值
|
|
|
|
|
SIMILARITY_THRESHOLD = 0.5 # 距离小于 0.5 的结果被认为是高相似度
|
|
|
|
|
|
|
|
|
|
# 在 /reply 接口中优化查询逻辑和提示词构建方式
|
|
|
|
|
# 在 /reply 接口中优化提示词传递和系统提示词
|
|
|
|
|
# 在 /reply 接口中优化提示词传递和系统提示词
|
|
|
|
|
# 在 /reply 接口中优化查询逻辑和提示词构建方式
|
|
|
|
|
@app.post("/reply")
|
|
|
|
|
async def reply(session_id: str = Form(...), prompt: str = Form(...)):
|
|
|
|
|
"""
|
|
|
|
|
接收用户输入的 prompt，调用大模型并返回结果
|
|
|
|
|
:param session_id: 用户会话 ID
|
|
|
|
|
:param prompt: 用户输入的 prompt
|
|
|
|
|
:return: 大模型的回复
|
|
|
|
|
"""
|
|
|
|
|
try:
|
|
|
|
|
# 从连接池中获取一个连接
|
|
|
|
|
connection = milvus_pool.get_connection()
|
|
|
|
|
|
|
|
|
|
# 将用户输入转换为嵌入向量
|
|
|
|
|
current_embedding = text_to_embedding(prompt)
|
|
|
|
|
|
|
|
|
|
# 提取用户输入的关键词
|
|
|
|
|
keywords = extract_keywords(prompt)
|
|
|
|
|
print(f"提取的关键词: {keywords}")
|
|
|
|
|
|
|
|
|
|
# 构建查询条件
|
|
|
|
|
expr = f"session_id == '{session_id}'" # 只过滤 session_id
|
|
|
|
|
print(f"查询条件: {expr}")
|
|
|
|
|
|
|
|
|
|
# 查询与当前对话最相关的五条历史交互
|
|
|
|
|
search_params = {
|
|
|
|
|
"metric_type": "L2", # 使用 L2 距离度量方式
|
|
|
|
|
"params": {"nprobe": MS_NPROBE} # 设置 IVF_FLAT 的 nprobe 参数
|
|
|
|
|
}
|
|
|
|
|
start_time = time.time()
|
|
|
|
|
results = collection_manager.search(
|
|
|
|
|
data=current_embedding, # 输入向量
|
|
|
|
|
search_params=search_params, # 搜索参数
|
|
|
|
|
expr=expr, # 查询条件
|
|
|
|
|
limit=5 # 返回 5 条结果
|
|
|
|
|
)
|
|
|
|
|
end_time = time.time()
|
|
|
|
|
|
|
|
|
|
# 调试：输出查询结果
|
|
|
|
|
print(f"查询结果: {results}")
|
|
|
|
|
|
|
|
|
|
# 构建历史交互提示词
|
|
|
|
|
history_prompt = ""
|
|
|
|
|
if results:
|
|
|
|
|
for hits in results:
|
|
|
|
|
for hit in hits:
|
|
|
|
|
try:
|
|
|
|
|
# 过滤低相似度的结果
|
|
|
|
|
if hit.distance > SIMILARITY_THRESHOLD:
|
|
|
|
|
print(f"跳过低相似度结果，距离: {hit.distance}")
|
|
|
|
|
continue
|
|
|
|
|
|
|
|
|
|
# 查询非向量字段
|
|
|
|
|
record = collection_manager.query_by_id(hit.id)
|
|
|
|
|
if record:
|
|
|
|
|
print(f"查询到的记录: {record}")
|
|
|
|
|
# 只添加与当前问题高度相关的历史交互
|
|
|
|
|
if any(keyword in record['user_input'] for keyword in keywords):
|
|
|
|
|
history_prompt += f"用户: {record['user_input']}\n大模型: {record['model_response']}\n"
|
|
|
|
|
except Exception as e:
|
|
|
|
|
print(f"查询失败: {e}")
|
|
|
|
|
|
|
|
|
|
print(f"历史交互提示词: {history_prompt}")
|
|
|
|
|
|
|
|
|
|
# 调用大模型，将历史交互作为提示词
|
|
|
|
|
response = client.chat.completions.create(
|
|
|
|
|
model=MODEL_NAME,
|
|
|
|
|
messages=[
|
|
|
|
|
{"role": "system", "content": "你是一个私人助理，负责回答用户的问题。请根据用户的历史对话和当前问题，提供准确且简洁的回答。不要提及你是通义千问或其他无关信息。"},
|
|
|
|
|
{"role": "user", "content": f"{history_prompt}用户: {prompt}"} # 将历史交互和当前输入一起发送
|
|
|
|
|
],
|
|
|
|
|
max_tokens=500
|
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
# 提取生成的回复
|
|
|
|
|
if response.choices and response.choices[0].message.content:
|
|
|
|
|
result = response.choices[0].message.content.strip()
|
|
|
|
|
|
|
|
|
|
# 记录用户输入和大模型反馈到向量数据库
|
|
|
|
|
timestamp = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
|
|
|
|
|
entities = [
|
|
|
|
|
[session_id], # session_id
|
|
|
|
|
[prompt], # user_input
|
|
|
|
|
[result], # model_response
|
|
|
|
|
[timestamp], # timestamp
|
|
|
|
|
[current_embedding] # embedding
|
|
|
|
|
]
|
|
|
|
|
collection_manager.insert_data(entities)
|
|
|
|
|
print("用户输入和大模型反馈已记录到向量数据库。")
|
|
|
|
|
|
|
|
|
|
# 调用tts进行生成mp3
|
|
|
|
|
uuid_str = str(uuid.uuid4())
|
|
|
|
|
tts_file = "audio/" + uuid_str + ".mp3"
|
|
|
|
|
t = TTS(tts_file)
|
|
|
|
|
t.start(result)
|
|
|
|
|
# 文件上传到oss
|
|
|
|
|
upload_mp3_to_oss(tts_file, tts_file)
|
|
|
|
|
# 删除临时文件
|
|
|
|
|
try:
|
|
|
|
|
os.remove(tts_file)
|
|
|
|
|
print(f"临时文件 {tts_file} 已删除")
|
|
|
|
|
except Exception as e:
|
|
|
|
|
print(f"删除临时文件失败: {e}")
|
|
|
|
|
# 完整的url
|
|
|
|
|
url = 'https://ylt.oss-cn-hangzhou.aliyuncs.com/' + tts_file
|
|
|
|
|
return {
|
|
|
|
|
"success": True,
|
|
|
|
|
"url": url,
|
|
|
|
|
"search_time": end_time - start_time, # 返回查询耗时
|
|
|
|
|
"response": result # 返回大模型的回复
|
|
|
|
|
}
|
|
|
|
|
else:
|
|
|
|
|
raise HTTPException(status_code=500, detail="大模型未返回有效结果")
|
|
|
|
|
except Exception as e:
|
|
|
|
|
raise HTTPException(status_code=500, detail=f"调用大模型失败: {str(e)}")
|
|
|
|
|
finally:
|
|
|
|
|
# 释放连接
|
|
|
|
|
milvus_pool.release_connection(connection)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
# 运行 FastAPI 应用
|
|
|
|
|
if __name__ == "__main__":
|
|
|
|
|
import uvicorn
|
|
|
|
|
|
|
|
|
|
uvicorn.run(app, host="0.0.0.0", port=5600)
|
