peterh0323/streamer-sales

Streamer-Sales is an AI system designed to act as a live-stream sales host. You give it product information, and it generates the kind of persuasive spoken commentary that a human host would deliver during a live shopping broadcast. The README is written in Chinese and the project is aimed at the Chinese e-commerce live-streaming market, though the codebase is open source. The system combines several AI components working together. A large language model (fine-tuned from InternLM2) generates the sales script based on product features. A retrieval component (RAG) can pull from a product manual or specification document so the host's answers stay accurate for each specific item. A text-to-speech module converts the generated script into spoken audio. On top of that, there is a digital human generator that produces a video avatar speaking the words, so the output is not just text or audio but an animated presenter. An agent component can also perform live web searches, for example to look up current shipping information when a viewer asks. The voice input side is covered too: ASR (automatic speech recognition) lets users speak questions to the AI host rather than typing them. The web application has a Vue-based frontend and a FastAPI backend connected to a PostgreSQL database, with JWT-based authentication. The whole stack can be deployed in one step using Docker Compose. The README includes screenshots of an admin dashboard where you can upload products, generate scripts, and manage sessions. The underlying model was fine-tuned with roughly 400,000 tokens of training data and is available in both a standard and a 4-bit quantized version on ModelScope and OpenXLab. LMDeploy is used to accelerate inference, which the changelog notes improved throughput by around three times compared to the earlier version. The project won first place in a 2024 Chinese large-model competition.