Discussions

Hi everyone,

I'm currently integrating HeyGen's Knowledge Base functionality into my AI language learning application. While the feature works well in generating responses based on custom knowledge, I’d love to understand more about how it actually works behind the scenes.

Some Key Questions:
1️⃣ Which LLM powers the Knowledge Base?
Does HeyGen use a proprietary model, or is it built on OpenAI, Claude, or another third-party LLM?

2️⃣ How does retrieval work?
Does it function like a traditional RAG (Retrieval-Augmented Generation) system?
Does it pre-process and index uploaded data for faster retrieval?

3️⃣ How much context does the model retain?
Does it only respond based on the most relevant chunks of the knowledge base?
Can it maintain long-form memory over multiple interactions?

4️⃣ Customization & Fine-tuning:
Can we fine-tune responses to ensure more structured or domain-specific outputs?
Are there any built-in prompt engineering techniques being applied behind the scenes?

5️⃣ Limitations & Best Practices:
Are there any word/character limits on knowledge base entries?
Any best practices to improve the response accuracy and relevance?

Understanding this would greatly help in optimizing the experience for my users and ensuring that the AI Tutor provides structured and accurate learning guidance.

Looking forward to any insights from the HeyGen team or community members who have experimented with this! 🚀

Thanks in advance! 😊