🏭 Industrial-Intelligent-QA: High-Precision Intelligent Q&A System for General Industry

Based on a RAG + FAQ dual-engine architecture, integrating BGE-M3 hybrid retrieval and the RAGAS automated evaluation system to solve LLM hallucinations and knowledge latency.

📖 Project Introduction

This project is an intelligent Q&A system specifically designed for the General Industrial vertical domain. Addressing the issues of "knowledge latency" and "hallucinations" common in general large models within professional fields, the system innovatively adopts a FAQ (MySQL) + RAG (Milvus) dual-retrieval architecture.

By introducing BGE-M3 hybrid retrieval, dynamic intent recognition routing, and a RAGAS automated evaluation closed-loop, the system transitions from experience-based tuning to data-driven optimization. It ensures the rigor and timeliness of answers while utilizing Redis caching and multi-level indexing strategies to achieve millisecond-level responses for high-frequency queries and second-level responses for complex reasoning.

✨ Core Highlights

• 🚀 Dual-Engine Retrieval Architecture: Combines MySQL FAQ (BM25 exact matching) with Milvus RAG (semantic vector retrieval). High-frequency questions are answered in <100ms, while complex questions automatically degrade to the RAG process.
• 🧠 Adaptive Multi-Strategy Retrieval: Features advanced strategies like HyDE and Sub-query Decomposition to effectively solve multi-hop reasoning problems.
• 🎯 High-Precision Hybrid Retrieval: Implements Dense + Sparse vector hybrid retrieval based on BGE-M3, coupled with BGE-Reranker-Large re-ranking to filter noise significantly, boosting accuracy to 90%+.
• 📊 RAGAS Automated Evaluation: Constructs a standardized evaluation dataset to automatically calculate four key metrics (Faithfulness, Answer Relevancy, etc.), enabling a visualized closed-loop for algorithm iteration.
• 🐳 Full-Chain Containerized Deployment: Uses Docker Compose for one-click orchestration of Milvus, Redis, MySQL, and backend services, supporting rapid private deployment.

🛠️ Tech Stack

🏗️ System Architecture

The system adopts a layered design, primarily consisting of offline data processing and online services:

graph TD
    User[User Query] --> Router{Intent Recognition Router}
    
    subgraph "Online Service Layer"
    Router -- High Freq / Simple --> FAQ_Module[FAQ Module (MySQL + Redis)]
    Router -- Complex / Long-tail --> RAG_Module[RAG Module (Milvus)]
    
    FAQ_Module -- Similarity > 0.85 --> Response[Direct Answer]
    
    RAG_Module --> Retriever[Hybrid Retrieval (Dense+Sparse)]
    Retriever --> Rerank[BGE-Reranker Re-rank]
    Rerank --> Context[Context Augmentation]
    Context --> LLM[Qwen2.5 Generation]
    LLM --> Response
    end
    
    subgraph "Offline Processing Layer"
    Docs[Industrial Docs/PDF] --> Splitter[Document Splitting]
    Splitter --> Embedder[Vectorization]
    Embedder --> Milvus[(Milvus Vector DB)]
    end

Core Module Descriptions

1. Smart Router: Identifies user intent based on a fine-tuned BERT model and dynamically dispatches requests.
2. AQ Module: Uses BM25 algorithm for keyword matching, hitting Redis cache for ultra-fast responses.
3. RAG Retrieval Engine:

• Hybrid Retrieval: Utilizes both semantic and keyword vectors to compensate for the shortcomings of single retrieval methods.
• Re-ranking: Fine-tunes scores for initial results to ensure high-quality Top-K context.
• Evaluation System (RAGAS): Runs automated tests regularly to monitor and optimize system performance.

🚀 Quick Start

1. Environment Preparation Ensure Docker and Docker Compose are installed.

git clone https://github.com/your-username/Industrial-Intelligent-QA.git
cd Industrial-Intelligent-RAG

2. Start Services Use the provided script to start all dependencies (Milvus, Redis, MySQL) and backend services:

chmod +x start_compose.sh
./start_compose.sh

Or manually using docker-compose:

docker-compose up -d --build

3. Initialize Knowledge Base Place industrial general documents into the rag_qa/data directory and run the ingestion script:

python rag_qa/core/ingest_data.py

4. Access Interfaces Once the service is started, access via: API Docs: http://localhost:8000/docs Health Check: http://localhost:8000/health

📂 Project Structure

integrated_qa_system/
├── base/               # Base config & logging
├── mysql_qa/           # FAQ Module (MySQL + Redis + BM25)
│   ├── cache/          # Redis caching logic
│   └── retrieval/      # Similarity calculation
├── rag_qa/             # RAG Core Module
│   ├── core/           # Full RAG flow orchestration
│   ├── project_document_loaders/ # Document loaders
│   ├── models/         # Model definitions (Embedding, Rerank)
│   └── rag_assessment/ # RAGAS evaluation scripts
├── app.py              # Flask/FastAPI entry
├── docker-compose.yml  # Container orchestration
└── requirements.txt    # Dependencies

📦 Models & Resources

1. Model Downloads The required models can be downloaded from the link below or from their official repositories:

Baidu Pan Link: https://pan.baidu.com/s/1qUqdoHzKVa8cq2TJ4AVYIg?pwd=h588 Extraction Code: h588
Required Models:
bert_query_classifier
bert-base-chinese
bge-m3
bge-reranker-large
nlp_bert_document-segmentation_chinese-base

2. Related Blog For a step-by-step tutorial on building and learning this system, please refer to:

https://blog.csdn.net/zhoupenghui168/article/details/162518685

3. 🤝 Contributing Issues and Pull Requests are welcome! If you have suggestions for optimizing Industrial RAG, feel free to contact us.

=====================================================================================

工业通用领域高精度智能问答系统

基于 RAG + FAQ 双引擎架构，融合 BGE-M3 混合检索与 RAGAS 自动化评估体系，解决大模型幻觉与知识滞后痛点。

📖 项目简介

本项目是一个专为工业通用垂直领域设计的智能问答系统。针对通用大模型在专业领域存在的“知识滞后”和“幻觉”问题，系统创新性地采用了 FAQ (MySQL) + RAG (Milvus) 的双路召回架构。 通过引入 BGE-M3 混合检索、动态意图识别路由以及 RAGAS 自动化评估闭环，实现了从经验调优到数据驱动优化的转变。系统不仅保证了回答的严谨性与时效性，还通过 Redis 缓存与多级索引策略，实现了毫秒级的高频响应与秒级的复杂推理回复。

✨ 核心亮点

• 🚀 双引擎召回架构：结合 MySQL FAQ（BM25 精确匹配）与 Milvus RAG（语义向量检索），高频问题 <100ms 响应，复杂问题自动降级至 RAG 流程。
• 🧠 自适应多策略检索：内置 HyDE、子查询分解（Sub-query Decomposition）等高级策略，有效解决多跳推理难题。
• 🎯 高精度混合检索：基于 BGE-M3 实现 Dense + Sparse 向量混合检索，配合 BGE-Reranker-Large 重排序，大幅过滤噪声，准确率提升至 90%+。
• 📊 RAGAS 自动化评估：构建标准化评估数据集，自动化计算 Faithfulness, Answer Relevancy 等四大指标，实现算法迭代的可视化闭环。
• 🐳 全链路容器化部署：基于 Docker Compose 一键编排 Milvus、Redis、MySQL 及后端服务，支持私有化快速落地。

🛠️ 技术栈

🏗️ 系统架构

系统采用分层设计，主要包含离线数据处理与在线服务两大部分：

graph TD
    User[用户 Query] --> Router{意图识别路由}
    
    subgraph "在线服务层"
    Router -- 高频/简单问题 --> FAQ_Module[FAQ 模块 (MySQL + Redis)]
    Router -- 复杂/长尾问题 --> RAG_Module[RAG 模块 (Milvus)]
    
    FAQ_Module -- 相似度 > 0.85 --> Response[直接返回答案]
    
    RAG_Module --> Retriever[混合检索 (Dense+Sparse)]
    Retriever --> Rerank[BGE-Reranker 精排]
    Rerank --> Context[上下文增强]
    Context --> LLM[Qwen2.5 生成]
    LLM --> Response
    end
    
    subgraph "离线处理层"
    Docs[工业通用文档/PDF] --> Splitter[文档切分]
    Splitter --> Embedder[向量化]
    Embedder --> Milvus[(Milvus 向量库)]
    end

核心模块说明

• 1. 智能路由 (Smart Router)：基于微调后的 BERT 模型识别用户意图，动态分发请求。
• 2. FAQ 模块：利用 BM25 算法进行关键词匹配，命中 Redis 缓存实现极速响应。
• 3. RAG 检索引擎：
• 混合检索：同时利用语义向量与关键词向量，弥补单一检索短板。
• 重排序：对初筛结果进行精细打分，确保 Top-K 上下文的高质量。
• 评估体系 (RAGAS)：定期运行自动化测试，监控并优化系统表现。

🚀 快速开始

1. 环境准备 确保已安装 Docker 和 Docker Compose。

git clone https://github.com/your-username/Medical-Intelligent-QA.git
cd Medical-Intelligent-QA

2. 一键启动服务 使用提供的脚本启动所有依赖组件（Milvus, Redis, MySQL）及后端服务：

chmod +x start_compose.sh
./start_compose.sh

或者手动使用 docker-compose：

docker-compose up -d --build

3. 初始化知识库 将工业通用文档放入 rag_qa/data 目录，运行数据入库脚本：

python rag_qa/core/ingest_data.py

4. 访问接口 服务启动后，可通过以下地址访问： API 文档: http://localhost:8000/docs 健康检查: http://localhost:8000/health

📂 项目结构

integrated_qa_system/
├── base/               # 基础配置与日志模块
├── mysql_qa/           # FAQ 模块 (MySQL + Redis + BM25)
│   ├── cache/          # Redis 缓存逻辑
│   └── retrieval/      # 相似度计算
├── rag_qa/             # RAG 核心模块
│   ├── core/           # 完整 RAG 流程编排
│   ├── project_document_loaders/ # 文档加载器
│   ├── models/         # 模型定义 (Embedding, Rerank)
│   └── rag_assessment/ # RAGAS 评估脚本
├── app.py              # Flask/FastAPI 入口
├── docker-compose.yml  # 容器编排文件
└── requirements.txt    # 依赖列表

项目相关

1. model下载地址

通过网盘分享的文件：models
链接: https://pan.baidu.com/s/1qUqdoHzKVa8cq2TJ4AVYIg?pwd=h588 提取码: h588

所需要的模型如下, 也可以自行在官网下载:
bert_query_classifier
bert-base-chinese
bge-m3
bge-reranker-large
nlp_bert_document-segmentation_chinese-base

2. 相关博客地址 可以根据博客一步步搭建学习

https://blog.csdn.net/zhoupenghui168/article/details/162518685