Proposal for Enhancing Istio Community Engagement with AI-Integrated Q&A Bot about community

[irisdingbj]

Thanks @rootsongjc for this initiative and proposal!

@howardjohn @linsun @craigbox @kfaseela @rcernich @justinpettit @ctrath @hzxuzhonghu please take a look for the proposal.

[linsun]

Hi @rootsongjc thanks for proposing this. I saw the cost and complexity worry from other steering members which makes sense to me. Is it possible to start with a simple version where we simply alert users to check out a list of things before writing up the post in discuss?

[rootsongjc]

I want to adapt and utilize the existing bots. What is available now? I've heard of some bots no longer planned to be maintained.

[craigbox]

Any further thought Jimmy? I think that having a model trained on Istio documentation attempting to manually answer some user queries would be a good first step. You have such a model?

[rootsongjc]

@craigbox I don't have a direct model available for open source yet. But we can train one, and there's also the technology or platform to consider using, and there's the cost involved when we have to integrate the model to call the APIs, and I haven't figured out how to do all that yet.

[terrchen]

@rootsongjc is there still interest in doing this?

@craigbox has provided some feedback for my tool (https://devboard.gitsense.com/istio) and I'm looking to convert all my data into embeddings for future AI features and I would be interested in working with the Istio to gather requirements. I'm currently capturing comments (up to 30,000 per repository due to GitHub limitations), issues, pull request, commits, etc. so experimenting with different types of embedding models and chunking methods will be trivial.

My only concern right now is performance (not my indexing engine but rather the process of generating embeddings). My indexing engine can scale horizontally and you can rent GPUs by the hour, so I'm hoping there will only be a one time initial cost hit and from then on, commodity hardware can be used

[craigbox]

What Jimmy is proposing is effectively training a transformer model on the Istio and Envoy documentation, Q&A etc, and then using that to answer user questions in the first instance.

Some GitHub data might be useful in this, but I think that is more likely to trend towards developer questions/answers.

[terrchen]

I'm actually interested in generating embeddings that can support multiple personas (customers, developers, managers, executives, etc.) so what Jimmy is proposing does interest me, but I can see how including more development related data, can increase complexity.

I'm not sure what stage things are at, but I am very interested in learning more about your findings and knowing what sources (documents, questions, etc.) you are planning on training on.

[rootsongjc]

@terrchen I haven't started yet. Your tool is very useful for showing the contribution data but do you have any bot to answer questions in the the GIthub issue or discussions?

[terrchen]

@rootsongjc Not yet. The goal is to get to a point where I can create agents/bots to answer questions and to perform tasks for maintainers, developers, team leads, etc.

The conclusion that I've come to is, in order to create a useful automated Q and A system, you'll need an easy way to do the following:

1. Classify and clean data.
2. Generate Q and A pairs.
3. Review and iterate on generated Q and A pairs

The part that I will be tackling in the near future is creating a system to classify and clean data, which is the most critical, since garbage in = garbage out. Once you have clean data (with good metadata), generating Q and A pairs should be pretty straight forward as you can use LLMs to generate them.

If you have a system or thoughts on how to clean/organize the data, I'd be interested to hear them. I'm currently planning on creating a system that will leverage LLM to help classify and prep data.