From 3bf0446ef8c7d5fd97ebe4b765d19b598f2594c0 Mon Sep 17 00:00:00 2001 From: root <953994191@qq.com> Date: Thu, 30 Apr 2026 17:45:06 +0800 Subject: [PATCH] =?UTF-8?q?feat:=20=E4=BF=AE=E5=A4=8D=E6=95=B0=E6=8D=AE?= =?UTF-8?q?=E5=BA=93=E6=8C=81=E4=B9=85=E5=8C=96=EF=BC=8C=E5=AE=8C=E5=96=84?= =?UTF-8?q?=E6=9C=8D=E5=8A=A1=E9=99=8D=E7=BA=A7=E6=9C=BA=E5=88=B6?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - 恢复使用 AsyncPostgresSaver 持久化短期记忆 - 添加 LLM 作为 Rerank 服务的最后降级方案 - 完善降级链：Local llama.cpp → Zhipu Rerank → LLM Fallback --- backend/app/backend.py | 83 +++-- backend/app/model_services/rerank_services.py | 92 ++++- scripts/start.sh | 344 +----------------- 3 files changed, 148 insertions(+), 371 deletions(-) diff --git a/backend/app/backend.py b/backend/app/backend.py index 05825d4..c8ea409 100644 --- a/backend/app/backend.py +++ b/backend/app/backend.py @@ -4,8 +4,7 @@ FastAPI 后端 - 支持动态模型切换，使用 PostgreSQL 持久化记忆 """ import os -# from app.config import DB_URI, BACKEND_PORT -from app.config import BACKEND_PORT +from app.config import DB_URI, BACKEND_PORT import uuid import json from contextlib import asynccontextmanager @@ -15,7 +14,7 @@ from fastapi import FastAPI, HTTPException, WebSocket, WebSocketDisconnect, Depe from fastapi.middleware.cors import CORSMiddleware from fastapi.responses import StreamingResponse from pydantic import BaseModel -from langgraph.checkpoint.memory import MemorySaver +from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver from .agent.service import AIAgentService from .agent.history import ThreadHistoryService from app.core.human_review import ( @@ -27,42 +26,56 @@ from app.core.human_review import ( from app.subgraphs.contact.api_client import ContactAPIClient from app.subgraphs.dictionary.api_client import DictionaryAPIClient from app.subgraphs.news_analysis.api_client import NewsAPIClient -# from .db.init_db import init_subgraph_tables -# from .db.models import ContactRepository, DictionaryRepository, NewsRepository +from .db.init_db import init_subgraph_tables +from .db.models import ContactRepository, DictionaryRepository, NewsRepository from app.logger import info, error @asynccontextmanager async def lifespan(app: FastAPI): - """应用生命周期管理：创建并注入全局服务（临时用内存 checkpoint）""" - # 1. 创建内存 checkpointer（临时测试） - checkpointer = MemorySaver() - - # 2. 构建 AI Agent 服务 - agent_service = AIAgentService(checkpointer) - await agent_service.initialize() - - # 3. 创建历史查询服务（保持原有的 checkpointer 参数） - history_service = ThreadHistoryService(checkpointer) - - # 4. 创建审核管理器 - review_manager = ReviewManager(InMemoryReviewStore()) - - # 5. 将服务实例存入 app.state - app.state.agent_service = agent_service - app.state.history_service = history_service - app.state.review_manager = review_manager - app.state.contact_api = ContactAPIClient() - app.state.dictionary_api = DictionaryAPIClient() - app.state.news_api = NewsAPIClient() - app.state.contact_repo = None - app.state.dictionary_repo = None - app.state.news_repo = None - - # 应用运行中... - yield - - # 6. 关闭时清理 - info("🛑 应用关闭") + """应用生命周期管理：创建并注入全局服务""" + # 1. 创建数据库连接池并初始化表（仅 checkpointer） + async with AsyncPostgresSaver.from_conn_string(DB_URI) as checkpointer: + await checkpointer.setup() + + # 1.5 初始化子图表 + await init_subgraph_tables(checkpointer.conn) + + # 2. 构建 AI Agent 服务 + agent_service = AIAgentService(checkpointer) + await agent_service.initialize() + + # 3. 创建历史查询服务（保持原有的 checkpointer 参数） + history_service = ThreadHistoryService(checkpointer) + + # 3.5 创建子图 Repositories + contact_repo = ContactRepository(checkpointer.conn) + dictionary_repo = DictionaryRepository(checkpointer.conn) + news_repo = NewsRepository(checkpointer.conn) + + # 3.6 创建子图 API 客户端（真实数据库模式） + contact_api = ContactAPIClient(checkpointer.conn) + dictionary_api = DictionaryAPIClient(word_repository=dictionary_repo) + news_api = NewsAPIClient(news_repository=news_repo) + + # 4. 创建审核管理器 + review_manager = ReviewManager(InMemoryReviewStore()) + + # 5. 将服务实例存入 app.state + app.state.agent_service = agent_service + app.state.history_service = history_service + app.state.review_manager = review_manager + app.state.contact_api = contact_api + app.state.dictionary_api = dictionary_api + app.state.news_api = news_api + app.state.contact_repo = contact_repo + app.state.dictionary_repo = dictionary_repo + app.state.news_repo = news_repo + + # 应用运行中... + yield + + # 6. 关闭时自动清理数据库连接（async with 负责） + info("🛑 应用关闭，数据库连接池已释放") app = FastAPI(lifespan=lifespan) diff --git a/backend/app/model_services/rerank_services.py b/backend/app/model_services/rerank_services.py index c1fff27..ecd5855 100644 --- a/backend/app/model_services/rerank_services.py +++ b/backend/app/model_services/rerank_services.py @@ -136,6 +136,92 @@ class ZhipuRerankService(BaseRerankService): raise +class LLMFallbackRerankService(BaseRerankService): + """ + 使用 LLM 作为最后的降级方案进行重排 + 通过让 LLM 评估文档相关性并给出分数 + """ + + def __init__(self, llm=None): + from .chat_services import get_chat_service + self.llm = llm or get_chat_service() + + def compute_scores(self, query: str, documents: List[str]) -> List[float]: + """ + 使用 LLM 评估文档相关性并打分 + """ + if not documents: + return [] + + scores = [] + for doc in documents: + score = self._score_single_document(query, doc) + scores.append(score) + + return scores + + def _score_single_document(self, query: str, document: str) -> float: + """ + 让 LLM 为单个文档的相关性打分 (0.0-1.0) + """ + prompt = f"""你是一个文档相关性评分专家。请评估以下文档与查询的相关性，返回一个0到1之间的分数： +- 1.0表示完全相关 +- 0.0表示完全不相关 + +查询: {query} + +文档: {document} + +请只返回一个数字，不要解释。""" + + try: + result = self.llm.invoke(prompt) + content = result.content if hasattr(result, 'content') else str(result) + # 尝试提取数字 + import re + match = re.search(r'(\d+\.?\d*)', content) + if match: + score = float(match.group(1)) + # 确保在 0-1 之间 + return max(0.0, min(1.0, score)) + # 如果没有找到数字，返回0.5作为默认值 + return 0.5 + except Exception as e: + logger.warning(f"LLM 打分失败，返回默认分数 0.5: {e}") + return 0.5 + + +class LLMFallbackRerankProvider(BaseServiceProvider[BaseRerankService]): + """ + LLM 降级重排服务提供者 + """ + + def __init__(self, llm=None): + super().__init__("llm_fallback_rerank") + self._llm = llm + + def is_available(self) -> bool: + """ + LLM 降级方案总是可用（只要 LLM 服务可用） + """ + try: + from .chat_services import get_chat_service + get_chat_service() + logger.info("LLM 降级重排服务可用") + return True + except Exception as e: + logger.warning(f"LLM 降级重排服务不可用: {e}") + return False + + def get_service(self) -> BaseRerankService: + """ + 获取 LLM 降级重排服务 + """ + if self._service_instance is None: + self._service_instance = LLMFallbackRerankService(self._llm) + return self._service_instance + + class LocalLlamaCppRerankProvider(BaseServiceProvider[BaseRerankService]): """ 本地 llama.cpp 重排服务提供者 @@ -221,13 +307,15 @@ def get_rerank_service() -> BaseRerankService: """ 获取重排服务（带自动降级）- 纯服务层 + 降级链: Local llama.cpp -> Zhipu Rerank -> LLM Fallback + Returns: BaseRerankService: 重排服务实例 """ def _create_chain(): primary = LocalLlamaCppRerankProvider() - fallback = ZhipuRerankProvider() - return FallbackServiceChain(primary, [fallback]) + fallbacks = [ZhipuRerankProvider(), LLMFallbackRerankProvider()] + return FallbackServiceChain(primary, fallbacks) chain = SingletonServiceManager.get_or_create("rerank_service_chain", _create_chain) return chain.get_available_service() diff --git a/scripts/start.sh b/scripts/start.sh index e109d91..a093e39 100755 --- a/scripts/start.sh +++ b/scripts/start.sh @@ -1,7 +1,7 @@ #!/bin/bash # ============================================================================= -# AI Agent 启动与管理脚本 -# 用法: ./backend/scripts/start.sh both [check|backend|frontend|both|docker-up|docker-down] +# AI Agent 启动与管理脚本 - 简化版 +# 用法: ./scripts/start.sh [check|backend|frontend|both] # ============================================================================= set -e @@ -14,284 +14,18 @@ YELLOW='\033[1;33m' NC='\033[0m' # No Color # 项目根目录 -PROJECT_DIR="/home/huang/Study/AIProject/Agent1" +PROJECT_DIR="/root/projects/ailine" echo -e "${BLUE}========================================${NC}" echo -e "${BLUE} AI Agent - 个人生活助手${NC}" echo -e "${BLUE}========================================${NC}" echo "" -# ============================================================================= -# 配置检查函数 -# ============================================================================= -check_config() { - echo -e "${BLUE}📋 开始环境配置检查...${NC}" - echo "" - - # 加载 .env 文件 - set -a - source "$PROJECT_DIR/.env" 2>/dev/null || true - set +a - - PASS=0 - FAIL=0 - WARN=0 - - # 辅助函数 - check_pass() { - echo -e "${GREEN}✓${NC} $1" - ((PASS++)) - } - - check_fail() { - echo -e "${RED}✗${NC} $1" - ((FAIL++)) - } - - check_warn() { - echo -e "${YELLOW}⚠${NC} $1" - ((WARN++)) - } - - # 1. 检查 .env 文件 - echo "🔍 检查配置文件..." - if [ -f "$PROJECT_DIR/.env" ]; then - check_pass ".env 文件存在" - - # 检查文件格式 - if grep -q "^EOF" .env 2>/dev/null; then - check_fail ".env 文件格式错误：发现多余的 EOF 标记" - else - check_pass ".env 文件格式正确" - fi - else - check_fail ".env 文件不存在" - echo " 提示: 请创建 .env 文件并配置环境变量" - return 1 - fi - - # 2. 检查必需的环境变量 - echo "" - echo "🔑 检查环境变量..." - - # 检查 ZHIPUAI_API_KEY - if grep -q "^ZHIPUAI_API_KEY=" "$PROJECT_DIR/.env" 2>/dev/null; then - API_KEY=$(grep "^ZHIPUAI_API_KEY=" "$PROJECT_DIR/.env" | head -1 | cut -d'=' -f2- | tr -d '[:space:]') - if [ ${#API_KEY} -gt 10 ]; then - check_pass "ZHIPUAI_API_KEY 已配置（长度: ${#API_KEY}）" - else - check_fail "ZHIPUAI_API_KEY 配置可能无效（过短）" - fi - else - check_fail "ZHIPUAI_API_KEY 未配置或格式错误" - fi - - # 检查 LLAMACPP_API_KEY - if grep -q "^LLAMACPP_API_KEY=" "$PROJECT_DIR/.env" 2>/dev/null; then - check_pass "LLAMACPP_API_KEY 已配置" - else - check_warn "LLAMACPP_API_KEY 未配置（如不使用本地模型可忽略）" - fi - - # 检查 DB_URI (远程服务器) - if grep -q "^DB_URI=.*115.190.121.151" "$PROJECT_DIR/.env" 2>/dev/null; then - check_pass "DB_URI 已配置（远程服务器）" - elif grep -q "^DB_URI=" "$PROJECT_DIR/.env" 2>/dev/null; then - check_warn "DB_URI 已配置（非远程服务器地址）" - else - check_fail "DB_URI 未配置" - fi - - # 检查 QDRANT_URL (远程服务器) - if grep -q "^QDRANT_URL=.*115.190.121.151" "$PROJECT_DIR/.env" 2>/dev/null; then - check_pass "QDRANT_URL 已配置（远程服务器）" - elif grep -q "^QDRANT_URL=" "$PROJECT_DIR/.env" 2>/dev/null; then - check_warn "QDRANT_URL 已配置（非远程服务器地址）" - else - check_fail "QDRANT_URL 未配置" - fi - - # 3. 检查 Docker 环境 - echo "" - echo "🐳 检查 Docker 环境..." - - if command -v docker &> /dev/null; then - check_pass "Docker 已安装" - - if docker info &> /dev/null; then - check_pass "Docker 守护进程正在运行" - else - check_fail "Docker 守护进程未运行" - echo " 提示: sudo systemctl start docker" - fi - else - check_fail "Docker 未安装" - fi - - if command -v docker compose version &> /dev/null || command -v docker-compose &> /dev/null; then - check_pass "Docker Compose 已安装" - else - check_fail "Docker Compose 未安装" - fi - - # 4. 检查端口占用 - echo "" - echo "🔌 检查端口占用..." - - for port in 8081 8082 8083 8501; do - if lsof -i :$port &> /dev/null; then - check_warn "端口 $port 已被占用" - else - check_pass "端口 $port 可用" - fi - done - - # 5. 检查远程服务连接 - echo "" - echo "🌐 检查远程服务连接..." - - # 测试 PostgreSQL 连接（从环境变量读取密码） - if command -v psql &> /dev/null; then - if [ -n "$DB_PASSWORD" ]; then - if PGPASSWORD="$DB_PASSWORD" psql -h 115.190.121.151 -U postgres -d langgraph_db -c "SELECT 1;" &> /dev/null; then - check_pass "PostgreSQL 远程连接正常 (115.190.121.151:5432)" - else - check_fail "PostgreSQL 远程连接失败" - echo " 提示: 检查网络连接和防火墙设置" - fi - else - check_warn "DB_PASSWORD 未配置，跳过 PostgreSQL 连接测试" - fi - else - check_warn "psql 客户端未安装，跳过 PostgreSQL 连接测试" - fi - - # 测试 Qdrant 连接 - if curl -s http://115.190.121.151:6333/collections &> /dev/null; then - check_pass "Qdrant 远程连接正常 (115.190.121.151:6333)" - else - check_fail "Qdrant 远程连接失败" - echo " 提示: 检查网络连接和防火墙设置" - fi - - # 总结 - echo "" - echo "==========================================" - echo " 检查结果汇总" - echo "==========================================" - echo -e "${GREEN}通过: $PASS${NC}" - echo -e "${RED}失败: $FAIL${NC}" - echo -e "${YELLOW}警告: $WARN${NC}" - echo "" - - if [ $FAIL -eq 0 ]; then - echo -e "${GREEN}✅ 配置检查通过！${NC}" - return 0 - else - echo -e "${RED}❌ 发现 $FAIL 个错误，请修复后重试${NC}" - return 1 - fi -} - -# ============================================================================= -# Docker 容器检查函数（仅检查 llama.cpp 服务） -# ============================================================================= -check_llamacpp() { - echo -e "${BLUE}🔍 检查 llama.cpp LLM 容器...${NC}" - if ! docker ps --format '{{.Names}}' | grep -q "^gemma4-llamacpp-server$"; then - echo -e "${YELLOW}⚠️ llama.cpp LLM 容器未运行${NC}" - return 1 - else - echo -e "${GREEN}✓ llama.cpp LLM 容器正在运行 (gemma4-llamacpp-server)${NC}" - return 0 - fi -} - -check_embedding() { - echo -e "${BLUE}🔍 检查 llama.cpp Embedding 容器...${NC}" - if ! docker ps --format '{{.Names}}' | grep -q "^embedding-server$"; then - echo -e "${YELLOW}⚠️ llama.cpp Embedding 容器未运行${NC}" - return 1 - else - echo -e "${GREEN}✓ llama.cpp Embedding 容器正在运行 (embedding-server)${NC}" - return 0 - fi -} - -# ============================================================================= -# 启动 Docker 依赖服务（llama.cpp） -# ============================================================================= -start_llamacpp() { - echo -e "${BLUE}🚀 启动 llama.cpp LLM 容器...${NC}" - - # 检查模型文件 - if [ ! -f "/home/huang/Study/AIModel/GGUF/Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q6_K_P.gguf" ]; then - echo -e "${RED}✗ 错误：LLM 模型文件不存在${NC}" - exit 1 - fi - - if [ ! -f "/home/huang/Study/AIModel/GGUF/mmproj-Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-f16.gguf" ]; then - echo -e "${RED}✗ 错误：多模态投影文件不存在${NC}" - exit 1 - fi - - docker run -d \ - --name gemma4-llamacpp-server \ - --restart=unless-stopped \ - --group-add=video \ - --device=/dev/kfd \ - --device=/dev/dri \ - -v /home/huang/Study/AIModel/GGUF:/models \ - -p 8081:8080 \ - ghcr.io/ggml-org/llama.cpp:server-rocm \ - -m /models/Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-Q6_K_P.gguf \ - --mmproj /models/mmproj-Gemma-4-E2B-Uncensored-HauhauCS-Aggressive-f16.gguf \ - --host 0.0.0.0 \ - --port 8080 \ - -ngl 99 - - echo -e "${GREEN}✓ llama.cpp LLM 容器已启动 (端口 8081)${NC}" - echo -e "${YELLOW}⏳ 等待模型加载（可能需要几分钟）...${NC}" - sleep 15 -} - -start_embedding() { - echo -e "${BLUE}🚀 启动 llama.cpp Embedding 容器...${NC}" - - # 检查模型文件 - if [ ! -f "/home/huang/Study/AIModel/GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf" ]; then - echo -e "${RED}✗ 错误：Embedding 模型文件不存在${NC}" - exit 1 - fi - - docker run -d \ - --name embedding-server \ - --restart=unless-stopped \ - --group-add=video \ - --device=/dev/kfd \ - --device=/dev/dri \ - -v /home/huang/Study/AIModel/GGUF:/models \ - -p 8082:8080 \ - -e LLAMA_ARG_CTX_SIZE=16384 \ - -e LLAMA_ARG_N_PARALLEL=1 \ - -e LLAMA_ARG_BATCH=512 \ - -e LLAMA_ARG_N_GPU_LAYERS=99 \ - -e LLAMA_ARG_API_KEY="$LLAMACPP_API_KEY" \ - ghcr.io/ggml-org/llama.cpp:server-rocm \ - -m /models/Qwen3-Embedding-0.6B-Q8_0.gguf \ - --host 0.0.0.0 \ - --port 8080 \ - --embeddings - - echo -e "${GREEN}✓ llama.cpp Embedding 容器已启动 (端口 8082)${NC}" - sleep 5 -} - # ============================================================================= # 启动 Python 服务 # ============================================================================= start_backend() { - echo -e "\n${BLUE}🚀 启动后端服务 (端口 8079)...${NC}" + echo -e "\n${BLUE}🚀 启动后端服务 (端口 10079)...${NC}" cd "$PROJECT_DIR" # 加载 .env 文件中的环境变量 @@ -300,6 +34,7 @@ start_backend() { set +a export PYTHONPATH="$PROJECT_DIR/backend" + export BACKEND_PORT=10079 python -m app.backend & BACKEND_PID=$! echo -e "${GREEN}✓ 后端服务已启动 (PID: $BACKEND_PID)${NC}" @@ -307,7 +42,7 @@ start_backend() { } start_frontend() { - echo -e "\n${BLUE}🎨 启动前端界面 (端口 8501)...${NC}" + echo -e "\n${BLUE}🎨 启动前端界面 (端口 10501)...${NC}" cd "$PROJECT_DIR" # 加载 .env 文件中的环境变量 @@ -316,41 +51,12 @@ start_frontend() { set +a export PYTHONPATH="$PROJECT_DIR/frontend/src" - streamlit run frontend/src/frontend_main.py --server.port 8501 --server.address 0.0.0.0 & + export API_URL="http://127.0.0.1:10079/chat" + streamlit run frontend/src/frontend_main.py --server.port 10501 --server.address 0.0.0.0 & FRONTEND_PID=$! echo -e "${GREEN}✓ 前端服务已启动 (PID: $FRONTEND_PID)${NC}" echo -e "${GREEN}✓ 访问地址:${NC}" - echo -e " 本地开发: http://127.0.0.1:8501" -} - -# ============================================================================= -# Docker Compose 管理 -# ============================================================================= -docker_up() { - echo -e "${BLUE}🐳 使用 Docker Compose 启动所有服务...${NC}" - cd "$PROJECT_DIR/docker" - - # 检查 .env 文件 - if [ ! -f "../.env" ]; then - echo -e "${RED}✗ 错误：.env 文件不存在${NC}" - echo " 请先复制配置文件:" - echo " cp .env.docker .env # 服务器部署" - exit 1 - fi - - docker compose up -d --build - - echo -e "\n${GREEN}✓ Docker Compose 服务已启动${NC}" - echo -e "${BLUE}📊 查看服务状态:${NC} docker compose ps" - echo -e "${BLUE}📝 查看日志:${NC} docker compose logs -f" - echo -e "${BLUE}🌐 访问应用:${NC} http://127.0.0.1:8501" -} - -docker_down() { - echo -e "${BLUE}🛑 停止 Docker Compose 服务...${NC}" - cd "$PROJECT_DIR/docker" - docker compose down - echo -e "${GREEN}✓ 服务已停止${NC}" + echo -e " 本地开发: http://127.0.0.1:10501" } # ============================================================================= @@ -366,10 +72,6 @@ cleanup() { kill $FRONTEND_PID 2>/dev/null || true echo -e "${GREEN}✓ 前端服务已停止${NC}" fi - echo -e "${YELLOW}💡 提示：Docker 容器需要手动停止${NC}" - echo -e " 停止 llama.cpp LLM: docker stop gemma4-llamacpp-server" - echo -e " 停止 llama.cpp Embedding: docker stop embedding-server" - echo -e " 或使用: $0 docker-down" exit 0 } @@ -380,62 +82,36 @@ trap cleanup SIGINT SIGTERM # 主逻辑 # ============================================================================= case "${1:-help}" in - check) - check_config - ;; - backend) - check_config || exit 1 - check_llamacpp || start_llamacpp - check_embedding || start_embedding start_backend echo -e "\n${GREEN}后端服务正在运行，按 Ctrl+C 停止${NC}" wait $BACKEND_PID ;; frontend) - check_config || exit 1 start_frontend echo -e "\n${GREEN}前端服务正在运行，按 Ctrl+C 停止${NC}" wait $FRONTEND_PID ;; both) - check_config || exit 1 - check_llamacpp || start_llamacpp - check_embedding || start_embedding start_backend sleep 3 start_frontend - echo -e "\n${GREEN}所有服务正在运行，按 Ctrl+C 停止 Python 服务${NC}" - echo -e "${YELLOW}注意：Docker 容器会在后台继续运行${NC}" + echo -e "\n${GREEN}所有服务正在运行，按 Ctrl+C 停止${NC}" wait ;; - docker-up) - check_config || exit 1 - docker_up - ;; - - docker-down) - docker_down - ;; - help|*) echo -e "${BLUE}用法:${NC} $0 [command]" echo "" echo -e "${BLUE}命令:${NC}" - echo " check 检查环境配置" echo " backend 仅启动后端服务" echo " frontend 仅启动前端服务" echo " both 启动前后端服务（默认）" - echo " docker-up 使用 Docker Compose 启动所有服务" - echo " docker-down 停止 Docker Compose 服务" echo " help 显示此帮助信息" echo "" echo -e "${BLUE}示例:${NC}" - echo " $0 check # 检查配置" echo " $0 both # 启动本地开发环境" - echo " $0 docker-up # 启动 Docker 部署环境" ;; esac