phoenix-zhou/agent-sweep-engine

What it solves

This project provides an intelligent customer service system specifically for robot vacuum cleaner users. It addresses the limitations of traditional customer service—such as slow response times and fragmented knowledge—by offering a one-stop service hub that handles everything from pre-sales inquiries to post-sales maintenance and personalized device optimization.

How it works

The system is built on an Agentic RAG (Retrieval-Augmented Generation) architecture using the ReAct (Reasoning + Acting) framework. It functions as a decision center that understands user intent and decides when to call specific tools:

• RAG Retrieval: Pulls accurate information from a vector database (ChromaDB) to answer technical or product questions without hallucinations.
• Report Generation: Analyzes actual device usage data (cleaning frequency, consumable status, error logs) to create personalized usage reports and optimization suggestions.
• Model Factory: A decoupled design that allows the system to switch between different LLMs (like Qwen or Llama) and embedding models via LangChain.

Who it’s for

• Consumer electronics brands looking to automate their customer support for robotic home appliances.
• Developers interested in implementing Agentic RAG and ReAct patterns for specific hardware product ecosystems.

Highlights

• Full Lifecycle Support: Covers pre-sales, post-sales, and proactive maintenance.
• Data-Driven Insights: Generates custom reports based on real device logs rather than just answering static questions.
• Multi-Interface Support: Includes API endpoints via Flask and web interfaces via Gradio and Streamlit.
• Integrated Evaluation: Features a dedicated evaluation module using BERT to measure accuracy, recall, and relevance of the RAG system.