A local YouTube Q&A Engine using Llama.cpp and Microsoft Phi-3-Mini

Author(s): Vatsal Saglani

Originally published on Towards AI.

The cheapest and easiest way for Video Question Answering
Image by ChatGPT

In my last blog about Microsoft-Phi-3-Mini, I discussed how Small language models (SLMs) like the Phi-3-Mini help with quick experimentations on a user’s local machine. In this blog, we’ll look at how we can prototype a VideoQA engine that runs locally using the Microsoft Phi-3-Mini model and llama-cpp-python.

Some parts of the LLM invocation and context management logic will be taken from the older blog, so please go through it before moving forward.

Phi-3-Mini is a great local LLM (SLM) for developing compute-efficient GenAI-powered applications

pub.towardsai.net

Before diving into the code, let’s first understand what we plan to implement.

We’ll be implementing a question-answering bot over a YouTube video. To achieve this first we’ll get the transcript of the YouTube video using the Python youtube-transcript-api. After that, we’ll divide the transcript into chunks. When chunking, we won’t be using any token-based, character-based, or word-based approach. Instead, we’ll use something different, which we’ll look into when we reach there.

Once we’ve created the chunks, we’ll start embedding them into batches. For this, we’ll use the BGE-Small-v1.5 embedding model. We won’t be using any Vector databases. As we’re doing things locally, we’ll use NumPy. Now once we get the query, we embed it and get the… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI