@motorro/firebase-ai-chat-vertexai

Firebase VertexAI chat library

OpenAI chat library. See top-level documentation for complete reference.

VertexAI setup

We also need to set up the VertexAI API. To do this we need prepare the model, wrapper and instructions bound to the ID or your chat assistant config:

import {projectID} from "firebase-functions/params";
import {VertexAI} from "@google-cloud/vertexai";
import {factory} from "@motorro/firebase-ai-chat-vertexai";

// Chat component factory
const chatFactory = factory(firestore(), getFunctions(), region);

// Create VertexAI adapter
const ai = chatFactory.ai(model, "/threads")

// Your assistant system instructions bound to assistant ID in chat config
const instructions: Readonly<Record<string, VertexAiSystemInstructions<any>>> = {
    "yourAssistantId": {
        instructions: instructions2,
        tools: {
            dispatcher: (dataSoFar, name, args) => dataSoFar,
            definition: [
                {functionDeclarations: [{name: "someFunction"}]}
            ]
        }
    }
}

const options: CallableOptions = {
  secrets: [openAiApiKey],
  region: region,
  invoker: "public"
};

Optional custom message mapper

By default, the library uses only text messages as-is. But if you want custom message processing, say image processing or adding some metadata, you could create your own AI message processor and supply it to chat factory worker method. Default message mapper could be found here.

const myMessageMapper: VertexAiMessageMapper = {
    toAi(message: NewMessage): Array<Part> {
        throw new Error("TODO: Implement mapping from Chat to OpenAI");
    },

    fromAi(message: Content): NewMessage | undefined {
        throw new Error("TODO: Implement OpenAI to Chat message mapping")
    }
}

Command dispatcher configuration

The requests to front-facing functions return to user as fast as possible after changing the chat state in Firestore. As soon as the AI run could take a considerable time, we run theme in a Cloud Task "offline" from the client request. To execute the Assistant run we use the second class from the library - the VertexAiChatWorker. To create it, use the AiChat factory we created as described in the main documentation.

To register the Cloud Task handler you may want to create the following function:

import {onTaskDispatched} from "firebase-functions/v2/tasks";
import {firestore} from "firebase-admin";
import {getFunctions} from "firebase-admin/functions";

// Function region
const region = "europe-west1";
// Collection path to store threads
const VERTEXAI_THREADS = "treads";

export const calculator = onTaskDispatched(
    {
      retryConfig: {
        maxAttempts: 1,
        minBackoffSeconds: 30
      },
      rateLimits: {
        maxConcurrentDispatches: 6
      },
      region: region
    },
    async (req) => {
      // Create and run a worker
      // See the `dispatchers` definitions below
      const vertexAi = new VertexAI({
          project: projectID.value(),
          location: region
      });
      const model = vertexAi.getGenerativeModel(
          {
              model: "gemini-1.0-pro",
              generationConfig: {
                  candidateCount: 1
              }
          },
          {
              timeout: 30 * 1000
          }
      );

      // Dispatch request  
      await chatFactory.worker(model, VERTEXAI_THREADS, instructions, myMessageMapper).dispatch(
          req,
          (chatDocumentPath: string, meta: Meta) => {
             // Optional task completion handler
             // Meta - some meta-data passed to chat operation
          }   
      );
    }
);

The VertexAiChatWorker handles the VertexAiChatCommand and updates VertexAiChatState with the results.

Full example is available in the sample Firebase project.