# 基于RAG的智能客服Agent系统

**Repository Path**: leke_code/rag_agent

## Basic Information

- **Project Name**: 基于RAG的智能客服Agent系统
- **Description**: 面向电商/金融场景，构建可解释、高准确率的智能客服系统，需解决传统LLM幻觉问题，支持实时知识库更新（如商品规则/政策变更），并能在4060硬件上流畅运行。
- **Primary Language**: Python
- **License**: MulanPSL-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 2
- **Created**: 2026-03-18
- **Last Updated**: 2026-03-18

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Intelligent Customer Service Agent System Based on RAG

#### Introduction
Targeting e-commerce/financial scenarios, building an interpretable and highly accurate intelligent customer service system that addresses the traditional LLM illusion problem, supports real-time knowledge base updates (such as product rules/policy changes), and can run smoothly on 4060 hardware.

#### 1. Software architecture
1. RAG module: Chroma vector database (lightweight, GPU accelerated)+custom retrieval strategy (mixed semantics+keyword retrieval)

2. Agent framework: LangGraph (state machine driven multi round dialogue)+custom tool calls (such as querying order system APIs)

3. Model layer: Lightweight LLM (such as Llama-8B-Instruct-4bit quantized version, 4060 memory friendly)+Prompt engineering (few sample examples+thought chain guidance)

#### 2. Optimization points

1. Retrieval result relevance reordering (based on cross encoder)

2. Dialogue history compression (reducing context length)

#### 3. Key implementation steps
1. Knowledge base construction: Crawl/organize business documents → Text segmentation (SentencePiece) → Embedding model (BGE-M3) vectorization → Chroma storage (supports incremental updates)

2. Optimization of retrieval module: Implement HybridRetriever (semantic retrieval+BM25 keyword retrieval fusion), dynamically adjust weights through relevances_store

3. Agent process design: Use LangGraph to define the State (including chat_ history/retrieved.docs/current step), with nodes including: user input parsing → retrieval enhancement → LLM generation → tool invocation (such as calculating shipping costs) → response verification (regular verification of phone numbers/order numbers)

#### 4. Deployment optimization

Model 4-bit quantization (GPTQ)+vLLM inference service (supporting dynamic batch processing), 4060 actual QPS ≥ 15 (single round dialogue)

#### 5.architecture 

```
rag_agent_system/
│
├── knowledge_base/              # Knowledge base management
│   ├── documents/               # Raw documents storage
│   ├── vector_db/               # ChromaDB vector database storage
│   └── knowledge_builder.py     # Knowledge base construction module
│
├── retrieval/                   # Document retrieval components
│   ├── hybrid_retriever.py      # Hybrid retrieval (vector + keyword)
│   └── reranker.py              # Result reranking module
│
├── agent/                       # Agent core logic
│   ├── agent_state.py           # State management for conversation
│   ├── langgraph_workflow.py    # LangGraph conversation workflow
│   ├── tools.py                 # Agent tools and utilities
│   └── history.py               # Conversation history compression
│
├── prompt/                      # Prompt project
│   ├── onstraints.txt           # Constraint Settings
│   └── thought_instructions.txt # Thinking Process Design
│
├── models/                      # Model management
│   └── setup_models.py          # Model initialization and configuration
│
├── api/                         # API interface
│   └── main.py                  # FastAPI server implementation
│
├── utils/                       # Utility functions
│   └── document_loader.py       # Document loading utilities
│
└── requirements.txt             # Python dependencies
```