Hello everyone!
I’m embarking on an intriguing project to develop a bot specifically designed to assist users in replying on Telegram. The core idea is for the bot to learn and replicate each user’s unique communication style, including their tone, writing style, and other nuances. This will involve acquiring a certain level of knowledge about each user, such as their typical writing patterns and preferences.
To date, I’ve experimented with prompt engineering. This involved defining a specific author tone, creating a persona for each user to provide context, and embedding all of their messages. This allows for vector search analysis when receiving a message, to generate example responses. However, the results so far haven’t been entirely satisfying.
I’m considering two primary approaches and would appreciate your insights:
- Fine-Tuning a Small Model for Each User: This would mean creating a separate, tailored model for each user. While this could yield high accuracy in mimicking individual styles, I’m concerned about the practicality and resource demands, especially for users with a message history of around 100-200 messages. Is fine-tuning for each user too time-consuming or resource-intensive in such cases?
- Advanced Prompt Engineering: Continuing with a larger, more general model and refining prompt engineering techniques. The aim is to achieve efficient yet personalized interactions, but I’m unsure if it can match the specificity of individualized models.
I’m seeking your advice on which approach might be more effective or if there are other strategies I should explore. Additionally, if you have any resources, papers, or examples of similar projects that could guide my decision, I’d be grateful for the share.
Thank you so much for your input and help! 