Looking for input on how to tackle my AI use case

[robiningelbrecht]

Hi 👋!

I'm the developer of https://github.com/robiningelbrecht/statistics-for-strava

a self-hosted web app designed to provide you with better stats

I want to add an AI agent to this app and I think I have kinda figured out what I need to do, but I'd like to have some feedback.
So the agent that I want to build needs to be aware of 2 different kinds of data:

• Structured data available in the database. For this I will use the tools and function calls to make the agent aware of that data and feed it with it.
• Wikipedia articles with general information about working out. If I understood correctly that is what RAG is for:

providing references to a knowledge base outside of the LLM training data sources before generating a response.

Does this mean I need to my agent to combine RAG functionality and tools? What do you think is the correct approach here?

[ilvalerione]

Technically Neuron provides you with the features needed for this implementation.

You can implement your agent extending the built-in \NeuronAI\RAG class to connect the embeddings provider and a vector store.

Since the RAG class is just an extension of the basic Agent class you can also attach Tools and define the system instructions as usual for your agent.

Maybe it's not so clear in the documentation.

This should give to your agent all the capabilities needed for your use case. Basically all inside a single Agent class.

https://docs.neuron-ai.dev/rag

How to feed the vector store

Before your agent can use relevant content to respond to your messages you have to get Wikipedia contents, transform this contents into embeddings, and save this information into the vector store.

Neuron provide you with data loaders, a mechanism to help you create this pipeline in a script that you can run to feed your database. All by using the same Neuron components.

Here is the guide on the documentation: https://docs.neuron-ai.dev/components/data-loader

Let us know if you stumble on any problems.

[ilvalerione]

@robiningelbrecht I added a new section in the RAG documentation to make this concept more clear: https://docs.neuron-ai.dev/rag#rag--tools

[robiningelbrecht]

@ilvalerione thanks!!!

[robiningelbrecht]

@ilvalerione I was wondering, are there any plans on adding a "DatabaseChatHistory" implementation to the library?
Or does it not make sense to store chat history in a database?
If this would make sense, I'd be willing to create PR for this

[ilvalerione]

I decided to not develop this implementation because I thought the database usually has such a custom needs. Also based it could be needed to use an ORM if you are inside a framework, or you could prefer to attach the chat create relations with the user table or something else.

Other developers already implemented it. Look at #31 as an example.

If you find a good starting point for a generalized implementation it could be great to share with the community.