feat(tests): Add Tools to the chat

[ronal2do]

The idea is to catch up with other ai sdk allowing for tooling use, this implementation are using openai only, other models still implementing it with different apis

For example

import { RAGChat } from "@/lib/rag-chat";
import { aiUseChatAdapter } from "@upstash/rag-chat/nextjs";
import type { Message } from "ai";
import { z } from "zod";
import { tool } from 'ai';
import { openai } from "@ai-sdk/openai";

// 👇 allow streaming responses up to 30 seconds
export const maxDuration = 30;

const ragChat = new RAGChat({
  model: openai("gpt-4-turbo"), // <- import { openai } from "@ai-sdk/openai";
});

export async function POST(request: Request) {
  const { messages } = await request.json();

  const question = (messages as Message[]).at(-1)?.content;
  if (!question) throw new Error("No question in the request");

  const myAbortSignal = new AbortController().signal;

  const response = await ragChat.chat(question, 
    { 
      streaming: true,

      toolingOptions: {
        abortSignal: myAbortSignal,
        tools: {
          weather: tool({
            description: 'Get the weather in a location',
            parameters: z.object({ location: z.string() }),
            
            execute: async ({ location }, { abortSignal }) => {
              return fetch(
                // you might need an API key
                `https://api.weatherapi.com/v1/current.json?q=${location}`,
                { signal: abortSignal }, // forward the abort signal to fetch
              );
            },
          }),
        },
        maxSteps: 5,
        // Force the model to attempt tool usage (for testing)
        toolChoice: "auto",
        // Attach context data
        // Handle step completion
        onStepFinish: ({ text, toolCalls, toolResults, finishReason }) => {
          console.log("onStepFinish", {
            text,
            toolCalls,
            toolResults,
            finishReason,
          });
        },
        // Handle completion
        onFinish: ({ text, toolCalls, toolResults }) => {
          console.log("onFinish", {
            text,
            toolCalls,
            toolResults,
          });
        },
      }
    }
  );

  return aiUseChatAdapter(response);
}

• Add streaming and non-streaming mode tests
• Implement error handling tests for both modes
• Add tooling options support with integration tests
• Update type definitions for ToolingOptions integration

The test suite now covers core functionality, error cases, and tooling features, ensuring robust handling of LLM interactions across different scenarios.

@rkt/app:dev:  POST /api/public/leads 200 in 1735ms
@rkt/app:dev: Step finished: {
@rkt/app:dev:   text: '',
@rkt/app:dev:   toolCalls: [
@rkt/app:dev:     {
@rkt/app:dev:       type: 'tool-call',
@rkt/app:dev:       toolCallId: 'call_L8V9K27KegppPNexhCBbcO6F',
@rkt/app:dev:       toolName: 'discountCupomCode',
@rkt/app:dev:       args: [Object]
@rkt/app:dev:     }
@rkt/app:dev:   ],
@rkt/app:dev:   toolResults: [
@rkt/app:dev:     {
@rkt/app:dev:       type: 'tool-result',
@rkt/app:dev:       toolCallId: 'call_L8V9K27KegppPNexhCBbcO6F',
@rkt/app:dev:       toolName: 'discountCupomCode',
@rkt/app:dev:       args: [Object],
@rkt/app:dev:       result: [Object]
@rkt/app:dev:     }
@rkt/app:dev:   ],
@rkt/app:dev:   finishReason: 'tool-calls'
@rkt/app:dev: }

Limitations:

onToolCall has not been called from the frontend, we might need to update out streamText wrapper and the nextjs connector
https://github.com/vercel/ai/blob/b9cba49c0db07890e46aa9ea6e53023f1f051a2a/content/cookbook/20-rsc/21-stream-text-with-chat-prompt.mdx#L4

[ronal2do]

I'm not adding all the tooling support, just the basics, in case you can ping me, that I can add more options

[ronal2do]

those are the main changes, a 3rd parameter

[ronal2do]

I've tried to make few tests here, but kinda hard without fully understanding of the service

[ronal2do]

Ai do not export, I've tried other ways to import this interface without duplicating code, if you have any ideas

[ronal2do]

Last comment, just notice, not sure if is because of the openai from ai-sdk, but when the chat load tools it no longer takes the context as part of the response. maybe this is an expected behavior, but I think this can be a reason to close this PR.

[ronal2do]

Update, this is the intended behavior, when we are using tools, we pass messages, and the context is retrieved by calling it as a tool as well