ai-video-assistant

AI Video Assistant is an end-to-end meeting-intelligence agent that converts video such as recorded standups, YouTube lectures, or client calls into structured, searchable knowledge. It uses a Retrieval-Augmented Generation pipeline that combines automatic speech recognition with large language models to transcribe, summarize, extract, and chat with video content. The workflow accepts a YouTube URL or local video/audio file, downloads and normalizes audio to WAV, splits it into chunks, and transcribes locally with OpenAI Whisper. Whisper can also translate Hindi/Hinglish audio to English on demand.

The transcript is processed with Mistral's language model to generate a professional meeting title, a bullet-point summary, action items with owners and deadlines, key decisions, and open questions. Chunked transcription and a map-reduce summarizer are used to keep hour-long recordings fast and coherent. Results are presented in a dark-themed Streamlit interface with live pipeline status. The RAG layer embeds the transcript with a Hugging Face sentence-transformer and indexes it in ChromaDB. In chat, questions retrieve relevant transcript segments and answers are grounded in that context.

The system is built as modular Python components for audio acquisition, transcription, summarization, structured extraction, and retrieval, so components such as the transcription engine, embedding model, or LLM provider can be swapped. It provides a Streamlit web app and a command-line runner. The application runs locally; only LLM calls via the Mistral API require a key.