FIAI Multi-Agent System: RAG and Local Inference Practice for Catering Assistants

The fiai-llm-test project developed by Benlaptrinh is a multi-agent AI assistant system for the catering industry. Combining Retrieval-Augmented Generation (RAG) technology with local Ollama inference, it builds an intelligent system that both protects data privacy and provides professional services, demonstrating a practical implementation path for AI applications in vertical domains.

The catering industry faces multiple challenges: standardizing customer service, improving operational efficiency, and delivering personalized experiences. Traditional customer service relies on preset rules and limited knowledge bases, making it difficult to handle complex inquiries. Large language models bring new possibilities to this field, and this project aims to solve these problems through multi-agent + RAG + local inference.

The system adopts a multi-agent architecture, decomposing complex tasks into subtasks: the intent recognition agent parses the user's input intent; the knowledge retrieval agent retrieves relevant information from the catering knowledge base; the response generation agent generates natural replies based on context; the coordination agent is responsible for task allocation, result integration, and dialogue state management to ensure coherent multi-turn conversations.

• RAG technology: Solves the problems of model knowledge lag and hallucinations through document processing (parsing, chunking, vectorization), vector database construction (open-source vector libraries such as Chroma/Milvus), and hybrid retrieval strategies (vector + keyword matching).
• Advantages of local Ollama inference: Data security (processing sensitive information locally, compliant), controllable costs (predictable long-term operation costs), and fast response (no network latency).

Tech stack: Backend based on Python, using LangChain/LlamaIndex to simplify RAG processes, multi-agent coordination with AutoGen/CrewAI; Ollama interacts via HTTP API, supporting model switching; frontend is a chat component that maintains session states. Deployment challenges and solutions: Knowledge base maintenance requires an automated synchronization mechanism; model domain adaptation through fine-tuning or prompt engineering; hallucination issues handled via fact-checking or confidence evaluation.

Industry value: The solution is suitable for customer service-intensive industries such as retail and hotels; local deployment + domain RAG is suitable for data-sensitive or network-restricted scenarios, lowering the threshold for small and medium-sized enterprises to build AI assistants. Future directions: Introduce multimodality (image recognition), voice interaction, and recommendation systems; optimize exception handling mechanisms (automatically escalate complex issues to humans).