Personal AI Chat Bot

AI Discussions
,
1

We have written numerous papers, and have completed numerous projects. I wanna build a bot that reads through these papers and answers the questions in first person perspective.

What should happen:

John Doe visited my web application and uploaded all the stuff, about his interests, his past work, current projects, progress, and challenges. My web application will then generate a chat interface, that he can share with his circle, for example via LinkedIn.

Anyone who accesses the link should be able to ask questions like ā€œHey John, how did you achieve x accuracy on the ABC scale in your project Yā€ and the chatbot will reply in the first person ā€œI have used XYZ, which helped me in achieving x accuracyā€. ā€œI have used references from PQR paper for the Y projectā€. Any question we ask about the person John Doe, the chatbot should reply in the First person ā€˜I didā€™, ā€˜I usedā€™, ā€˜I thinkā€™, etc., This chatbot is John Doe himself. and is limited to reply as John Doe only. If anyone asks questions that the bot doesnā€™t have enough information, it gives a soft reply to send an email, and the bot will send the conversation as an email to John Doe.

Applications: The basic usage I can think of is, that this is like a resume reading bot, that can answer questions on the resume you provide. Celebrities, Comedians, Players, and Scholars, all have their websites. A chatbot that replies like the celebrity ā€œTo check on their ongoing projects, tours, ticket buying optionsā€ etc.,

Future: I wanna add controls like humor level, slang, mood, languages, attitude etc.,

What to use? How to do?

These are the biggest questions I have right now. I know Bert can be used to achieve a simple question-answering model. And I can use LangChain for certain tasks. But, how do I achieve

  • It needs to produce more generic replies, like human conversations.
  • It should always reply as the owner/ Author of the bot only. In a ā€œfirst personā€ only.
3 Likes
2

Here are the main considerations:

  1. Good Prompt Engineering - this is where you can instruct your model to answer always in the first person, with whatever tone you wish, and only to provide answers that it has the data to back up. Iā€™d recommend Deeplearningā€™s Prompt Engineering short course to help here.

  2. The data, upon which the answers will be based. Here you need Retrieval Augmented Generation or RAG to ground the answers in facts. The RAG data will consist of the ā€œKnowledge Baseā€ that John Doe uploads. Watch the Deeplearning course on RAG. Llamaindex can help you ingest different kinds of data into the Knowledge Base e.g. PDFs, documents, html pages and so on.

  3. You mentioned Bert, but you can use ChatGPT, or open source models like Llama instead. There are short courses on some of the different LLMs. To build a UI for it, you can use Azure OpenAI for example, or you can use something like Voiceflow, StackAI or other tools with various LLM option in the backend.

Hope this helps!
Mim

4

Hej @saikbodi - Iā€™ve been looking for a similar solution to help with initial client meetings. A digital clone able to engage with customers at the customerā€™s connivence. I havenā€™t made much headway. Iā€™m not sure if this is a good example of something youā€™re looking to create https://profg.ai/ . Seems like an effective way to expand the general publics access to the professors knowledge.

5

This is it. profg.ai is a perfect example of what I want to achieve, and make this feature available globally, Fill the forms provided and an AI that responds in the first person is ready. Thank you for providing the necessary stuff I need for my project.