Getting started with Google Gemini with Go

I am curious what it would take to build an AI chat agent in Go. Google’s Gemini is currently free and has a Go API binding. I looked at OpenLlama however I need to experiment with the hardware I have to see if I can get a performant model for it.

I am envisioning a chat bot who can pitch my virtues to potential employers on my resume site. I quickly sketched a WebSocket solution using github.com/gorilla/websocket for the conversation. Now onto the real challenge!

I frankensteined the example at github.com/gorilla/websocket/tree/main/examples/chat to hijack the chat request-response cycle.

I grabbed a Google API key from Google AI Studio.

Integration of the agent

My first instinct is to pull this into a separate struct:

type agent struct {
    client *genai.Client
}


func (a *agent) query(ctx context.Context, input string) (string, error) {
    model := a.client.GenerativeModel("gemini-pro")
    
    resp, err := model.GenerateContent(ctx, genai.Text(input))
    if err != nil {
        return "", err
    }
    if len(resp.Candidates) == 0 {
        return "I am sorry, I did not understand", nil
    }
    p := resp.Candidates[0].Content.Parts[0]
    if txt, ok := p.(genai.Text); ok {
        return string(txt), nil
    }
    return "Returned another type", nil
}

Definitely alpha quality, making a lot of assumptions with zero context. Testing shows a total round trip time around 4-7 seconds for most queries. I was hoping for something sub 2 second but beggars can not be choosers. To accommodate the slow processing time I double writeWait to 20 seconds and resolve the occasional timeout.

Context Matters

With a connection to the AI unit, I now need context. I loaded up a bunch of info into a data.txt file containing my resume and some commonly asked questions. Normally I would place this in a system role for the AI. Google’s 1.0 model does not support a system role, and they are about to charge for 1.5 which does.

In order to give gemini-pro instructions the data.txt Part looked like the following:

You are the best recruiter in the world representing Mark Eschbach to potential clients.  Given the resume of Mark,
answer the question below as professionally as possible.

Resume:  
<data>

Question:

This provided contextually aware results. It was a fun experiment for an hour to get it up and running. If anyone uses it then I will return to the problem.

Building expert AI systems in Prolog have the same problem as these units: they want closed universes to reason within. Google does offer various forms of pull based context however I am not sure the pricing model on this right now.

TLDR

Checkout the example at github.com/meschbach/xp-resume-agent. Particularly the agent.go file for interacting with the Gemini.