May I create learning chat bot like Notebooklm of Google with Pickaxe

General
,
1

I want to create AI bot to help user easy create learning assitant like Noteboolm that only use knowledge-base to answer?

How to build it in studio?

Brgds,
James

3 Likes
2

I would love to create NotebookLM type responses.
I watched this resource but would appreciate more troubleshooting tips.

I’m trying to create a bot that is a guru about a specific fiction book I wrote - helping me to find key scenes, etc. I have a bigger long-term plan but this is the first step. The problem is it seems to struggle to search the Knowledge base and find those scenes. And when it fails to find something it will occassionally just make it up.

1 Like
3

Without knowing more about what you’re trying to do, it’s hard to answer. But generally speaking, yes many users create learning assistants built on top of Knowledge Bases. It’s a popular use case.

4

Thanks Mike.

Here’s a bit more insight into the bot purpose and instructions.

I created a bot with the purpose of it being a FAQ (NotebookLM style) resource for a fiction book series that consists of 4 titles, ~200pages each.

User Input: “can you find top 3 scenes with the highest emotion and conflict in [Book XYZ]?”

Bot Role:
Persona:
You are a genius marketer who has read every page of every book in the [series XYZ]. You have a keen understanding of key scenes that evoke strong emotions such as anger, sadness, and confusion. Your goal is to identify impactful scenes that can be used for social media marketing.

Rules:

  • Only use information from the Knowledge base
  • Provide precise and accurate key scenes from the books for social media use.
  • Each response must include dialogue that showcases personal conflict or drama.
  • Identify the name of the book the scene is from
  • Format responses to identify each key scene separately.
  • Include a brief statement explaining why the scene was chosen.
  • Specify the page number in the book for easy reference.
  • Provide the entire scene, starting with the dialogue and avoiding internal thoughts or scene descriptions.
  • End at a logical conclusion to maintain coherence.

Model: GPT 4

Configuration Tab: Default settings
Knowledge Base: All four books uploaded plus a series overview document
– Context: Use the content in this section (if relevant) to better inform your answer.

Result (based on Output and ‘Message Insights’ feature):
The ChatBot inconsistently identifies scenes from the book.
Here are some of those inconsistencies:

  1. Pulls scenes from other books in the series.
  2. Combines two different irrelevant scenes.
  3. Doesn’t find any scene and just makes it up (or pulls it from somewhere outside the knowledge base?)

This capability is the vital foundation for many different automations I have planned - ability to look through a set knowledge base and pull out the most relevant information based on user request. But haven’t yet ‘cracked the code’ on getting consistent and accurate results.

I’d greatly appreciate any tips, resources or best practices.