AINoon Tutorial 2

The goal of this tutorial is to make your own chatbot to answer questions based on information in a document.

• Chatbots can answer “general knowledge” questions based on information in the text they were trained on
• But we often want them to be able to answer questions about domain-specific or company knowledge sources, like:
  • Frequently asked questions and other help documents
  • Instruction manuals
  • Past problems and resolutions from issue tracking systems

We’re going to make a chatbot that can answer questions about TechNoon courses, like AINoon.

Create a chatbot

1. Open chatbase.co
2. Go to the Dashboard
3. Click New AI agent
   • If you already have one:
     1. Delete it (free accounts can only have one chatbot)
     2. Select Create agent from the agent list dropdown
4. If it forces you to upload a source:
   • We’re going to skip uploading files for now
   • Select Text from the sidebar, and fill in:
     • Title: TechNoon
     • Text: TechNoon delivers free lunchtime industry training courses.
     • Click Add text snippet
5. Select Create agent
   • Note that it tells us the total size of our sources compared to the limit - we won’t be getting anywhere near that limit today.

Playground and settings

• You should now be on the Playground for your agent.
• Let’s ask a simple question about TechNoon:

What is TechNoon?

AUDIENCE QUESTION: Did it give a helpful answer?

It will only give a good answer if you provided the Text source above.

Now let’s look at the options we have to configure our chatbot:

• We can choose what Model it uses
  • Lots of options from different providers
  • You can see a brief description of each model
  • Some are only available for a paid account
• We can set the temperature
  • With a higher temperature, you will get more variety in the results - which you may or may not want
  • Note that Chatbase may re-use answers to questions your chatbot has been asked previously
  • We’ll leave it at zero for our informative chatbot.
• We can customise the Instructions to the chatbot:
  • Every time a message is sent to the chatbot, the instructions will be included in the prompt to the underlying model.
  • You can choose from a list of predefined instructions
    • e.g. “Customer Service Agent”
  • Or write your own! Include instructions like:
    • Some services might refer to these as a role or persona
    • The chatbot’s goal - “Encourage the user to attend TechNoon”
    • Style - professional, relaxed, or a pirate :)
    • Specific instructions
      • E.g. Link the user to the TechNoon website
      • “Dos” have been found to work better than “don’ts”
    • Remember: The user can potentially get the chatbot to reveal any details in the prompt, so treat it like a public document.

Providing more knowledge

What if we ask it for information about TechNoon that it doesn’t have the answer for?

List the available TechNoon courses

It won’t give a very helpful response.

Let’s give it more detailed knowledge!

1. In a new tab, open technoon.org/ainoon
2. Download the technoon_knowledge.txt
   • Or share it with everyone in chat
3. Open the file and have a quick read
   • It contains lots of information about TechNoon courses
4. From Chatbase, open Sources -> Files from the sidebar
5. Click the big Upload button and select the txt file.
   • Remember, any file you upload can be seen by the service provider, or information from it extracted by users asking questions of the chatbot
   • Check with your organisation before uploading any private files
6. After uploading, click Retrain agent
   • It’s not really training the chatbot, it’s just preparing the file for the chatbot to use.
7. After it’s finished “training”, go back to the Playground
8. Ask it the question again:

List the available TechNoon courses

AUDIENCE QUESTION: Did it give a helpful answer?

But how is it using the file?

1. Every time a question is asked, it searches through the file for chunks that are relevant to the question
   • A bit like a Google search
2. Those chunks are added to the prompt so that the LLM can use them to answer the question!
   • You can click on Sources below the chatbot to see the chunks it used - some might be quite large!

This is a technique called Retrieval Augmented Generation or RAG which we’ll talk more about next week.

Deploying your chatbot

Let’s share our chatbot with the world!

1. First, let’s give the chatbot a better name

   1. Open Settings -> General from the sidebar
   2. Change the name to TechNoon Bot
      • We should make it clear to our users that this isn’t a human they’re chatting with.

2. Open Deploy from the sidebar

   • Lots of options for places we could deploy it, e.g.
     • As an account you can chat with on Slack
     • To reply to messages sent to your company’s Facebook page

3. Select Help Page

   • This will let us serve the chatbot on a web page
   • There are lots of options to customise how the page looks
   • We’ll keep the defaults

4. Select Deploy

5. Open the Domain setup tab, and click Visit Page

6. Try asking a question!

   What can you help me with?

7. You can now share the link to this page with anyone in the world!

   • This means the information in your instructions and sources is publicly accessible!

Testing

It’s important to test chatbots for their intended use case.

Refer to the Testing a Chatbot slide

AUDIENCE QUESTION: What happens if you ask it an off-topic question?

What is a large language model?

Conclusion

You’ve successfully built a chatbot that can answer questions about your own documents!

This is a great example of a pattern for using LLMs to build apps that extend their capabilities.

Next week we’re going to look more at Retrieval Augmented Generation and other patterns for building applications with LLMs.