Generating replies with prompt chaining using Gemini API and NestJS

Connie Leung - Jun 10 - - Dev Community

Introduction

In this blog post, I demonstrated generating replies with prompt chaining. Buyers can provide ratings and comments on sales transactions on auction sites like eBay. When the feedback is negative, the seller has to respond promptly to resolve the dispute. This demo saves time by generating replies in the same language as the buyer according to the tone (positive, neutral, or negative) and topics. Previous prompts obtain answers from the Gemini models and become a new prompt's parameters. Similarly, the model receives the new prompt to generate the final reply to keep customers happy.

Generate Gemini API Key

Go to https://aistudio.google.com/app/apikey to generate an API key for a new or an existing Google Cloud project.

Create a new NestJS Project 

nest new nestjs-customer-feedback
Enter fullscreen mode Exit fullscreen mode

Install dependencies 

npm i --save-exact @nestjs/swagger @nestjs/throttler dotenv compression helmet @google/generative-ai class-validator class-transformer
Enter fullscreen mode Exit fullscreen mode

Generate a Feedback Module 

nest g mo advisoryFeedback
nest g co advisoryFeedback/presenters/http/advisoryFeedback --flat
nest g s advisoryFeedback/application/advisoryFeedback --flat
nest g s advisoryFeedback/application/advisoryFeedbackPromptChainingService --flat
Enter fullscreen mode Exit fullscreen mode

Create an AdvisoryFeedbackModule module, a controller, a service for the API, and another service to build chained prompts.

Define Gemini environment variables 

// .env.example

PORT=3000
GOOGLE_GEMINI_API_KEY=<google gemini api key>
GOOGLE_GEMINI_MODEL=gemini-1.5-pro-latest
Enter fullscreen mode Exit fullscreen mode

Copy .env.example to .env, and replace GOOGLE_GEMINI_API_KEY and GOOGLE_GEMINI_MODEL with the actual API Key and the Gemini model, respectively.

  • PORT - port number of the NestJS application
  • GOOGLE_GEMINI_API_KEY - API Key of Gemini
  • GOOGLE_GEMINI_MODEL - Google model and I used Gemini 1.5 Pro in this demo

Add .env to the .gitignore file to prevent accidentally committing the Gemini API Key to the GitHub repo.

Add configuration files

The project has 3 configuration files. validate.config.ts validates the payload is valid before any request can route to the controller to execute .

// validate.config.ts

import { ValidationPipe } from '@nestjs/common';

export const validateConfig = new ValidationPipe({
  whitelist: true,
  stopAtFirstError: true,
  forbidUnknownValues: false,
});
Enter fullscreen mode Exit fullscreen mode

env.config.ts extracts the environment variables from process.env and stores the values in the env object.

// env.config.ts

import dotenv from 'dotenv';

dotenv.config();

export const env = {
  PORT: parseInt(process.env.PORT || '3000'),
  GEMINI: {
    API_KEY: process.env.GOOGLE_GEMINI_API_KEY || '',
    MODEL_NAME: process.env.GOOGLE_GEMINI_MODEL || 'gemini-pro',
  },
};
Enter fullscreen mode Exit fullscreen mode

throttler.config.ts defines the rate limit of the Translation API

// throttler.config.ts

import { ThrottlerModule } from '@nestjs/throttler';

export const throttlerConfig = ThrottlerModule.forRoot([
  {
    ttl: 60000,
    limit: 10,
  },
]);
Enter fullscreen mode Exit fullscreen mode

Each route allows ten requests in 60,000 milliseconds or 1 minute.

Bootstrap the application 

// bootstrap.ts

export class Bootstrap {
  private app: NestExpressApplication;

  async initApp() {
    this.app = await NestFactory.create(AppModule);
  }

  enableCors() {
    this.app.enableCors();
  }

  setupMiddleware() {
    this.app.use(express.json({ limit: '1000kb' }));
    this.app.use(express.urlencoded({ extended: false }));
    this.app.use(compression());
    this.app.use(helmet());
  }

  setupGlobalPipe() {
    this.app.useGlobalPipes(validateConfig);
  }

  async startApp() {
    await this.app.listen(env.PORT);
  }

  setupSwagger() {
    const config = new DocumentBuilder()
      .setTitle('ESG Advisory Feedback with Prompt Chaining ')
      .setDescription('Integrate with Gemini to improve ESG advisory feebacking by prompt chaining')
      .setVersion('1.0')
      .addTag('Gemini API, Gemini 1.5 Pro Model, Prompt Chaining')
      .build();
    const document = SwaggerModule.createDocument(this.app, config);
    SwaggerModule.setup('api', this.app, document);
  }
}
Enter fullscreen mode Exit fullscreen mode

Added a Bootstrap class to set up Swagger, middleware, global validation, CORS, and finally, application start.

// main.ts

import { env } from '~configs/env.config';
import { Bootstrap } from '~core/bootstrap';

async function bootstrap() {
  const bootstrap = new Bootstrap();
  await bootstrap.initApp();
  bootstrap.enableCors();
  bootstrap.setupMiddleware();
  bootstrap.setupGlobalPipe();
  bootstrap.setupSwagger();
  await bootstrap.startApp();
}

bootstrap()
  .then(() => console.log(`The application starts successfully at port ${env.PORT}`))
  .catch((error) => console.error(error));
Enter fullscreen mode Exit fullscreen mode

The bootstrap function enabled CORS, registered middleware to the application, set up Swagger documentation, and validated payloads using a global pipe.

I have laid down the groundwork, and the next step is to add an endpoint to receive a payload for generating replies with prompt chaining.

Define Feedback DTO 

// feedback.dto.ts

import { IsNotEmpty, IsString } from 'class-validator';

export class FeedbackDto {
  @IsString()
  @IsNotEmpty()
  prompt: string;
}
Enter fullscreen mode Exit fullscreen mode

FeedbackDto accepts a prompt, which is customer feedback.

Construct Gemini Models 

// gemini.config.ts

import { GenerationConfig, HarmBlockThreshold, HarmCategory, SafetySetting } from '@google/generative-ai';

export const SAFETY_SETTINGS: SafetySetting[] = [
  {
    category: HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
    threshold: HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
  },
  {
    category: HarmCategory.HARM_CATEGORY_HARASSMENT,
    threshold: HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
  },
  {
    category: HarmCategory.HARM_CATEGORY_HATE_SPEECH,
    threshold: HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
  },
  {
    category: HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT,
    threshold: HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
  },
];

export const GENERATION_CONFIG: GenerationConfig = {
  temperature: 0.5,
  topK: 10,
  topP: 0.5,
  maxOutputTokens: 2048,
};
Enter fullscreen mode Exit fullscreen mode 
// gemini.constant.ts

export const GEMINI_SENTIMENT_ANALYSIS_MODEL = 'GEMINI_SENTIMENT_ANALYSIS_MODEL';
export const GEMINI_REPLY_MODEL = 'GEMINI_REPLY_MODEL';
export const GEMINI_FIND_LANGUAGE_MODEL = 'GEMINI_FIND_LANGUAGE_MODEL';
Enter fullscreen mode Exit fullscreen mode 
// mode.factory.ts

import { GoogleGenerativeAI } from '@google/generative-ai';
import { env } from '~configs/env.config';
import { GENERATION_CONFIG, SAFETY_SETTINGS } from '../configs/genimi.config';

export function modelFactory(systemInstruction: string, toJson = false) {
  const genAI = new GoogleGenerativeAI(env.GEMINI.API_KEY);
  const generationConfig = toJson ? { ...GENERATION_CONFIG, responseMimeType: 'application/json' } : GENERATION_CONFIG;
  return genAI.getGenerativeModel({
    model: env.GEMINI.MODEL_NAME,
    systemInstruction,
    generationConfig,
    safetySettings: SAFETY_SETTINGS,
  });
}
Enter fullscreen mode Exit fullscreen mode 
// gemini-find-language.provider.ts

import { GenerativeModel } from '@google/generative-ai';
import { Provider } from '@nestjs/common';
import { GEMINI_FIND_LANGUAGE_MODEL } from '../constants/gemini.constant';
import { modelFactory } from './model-factory';

const FIND_LANGUAGE_SYSTEM_INSTRUCTION = `You are a multilingual expert that can identify the language used in this piece of text. Give me the language name, and nothing else.
  If the text is written in Chinese, please differentiate Traditional Chinese and Simplified Chinese. 
  `;

export const GeminiFindLanguageProvider: Provider<GenerativeModel> = {
  provide: GEMINI_FIND_LANGUAGE_MODEL,
  useFactory: () => modelFactory(FIND_LANGUAGE_SYSTEM_INSTRUCTION),
};
Enter fullscreen mode Exit fullscreen mode

The Gemini 1.5 Pro model accepts system instruction, which gives the model context. GeminiFindLanguageProvider is a model that can detect written language. When the language is Chinese, it should distinguish between Traditional Chinese and Simplified Chinese.

// sentiment-model.provider.ts

const SENTIMENT_ANALYSIS_SYSTEM_INSTRUCTION = `
    You are a sentiment analysis assistant who can identify the sentiment and topic of feedback and return the JSON output { "sentiment": string, "topic": string }.
    When the sentiment is positive, return 'POSITIVE', is neutral, return 'NEUTRAL', is negative, return 'NEGATIVE'.
`;

export const GeminiSentimentAnalysisProvider: Provider<GenerativeModel> = {
  provide: GEMINI_SENTIMENT_ANALYSIS_MODEL,
  useFactory: () => modelFactory(SENTIMENT_ANALYSIS_SYSTEM_INSTRUCTION, true),
};
Enter fullscreen mode Exit fullscreen mode

GeminiSentimentAnalysisProvider is a model that can identify the sentiment and topics of the feedback, and return the result in JSON. The JSON schema of the result is { sentiment: string; topic: string }.

// advisory-feedback-model.provider.ts

const REPLY_SYSTEM_INSTRUCTION =
  "You are a professional ESG advisor, please give a short reply to customer's response and in the same language.";

export const GeminiReplyProvider: Provider<GenerativeModel> = {
  provide: GEMINI_REPLY_MODEL,
  useFactory: () => modelFactory(REPLY_SYSTEM_INSTRUCTION),
};
Enter fullscreen mode Exit fullscreen mode

GeminiReplyProvider is a model that writes a short reply in the same language as the feedback.

Implement Reply Service 

// sentiment-analysis.type.ts

export type SentimentAnalysis = {
  sentiment: 'POSITIVE' | 'NEUTRAL' | 'NEGATIVE';
  topic: string;
};
Enter fullscreen mode Exit fullscreen mode 
// advisory-feedback-prompt-chaining.service.ts

// Omit the import statements 

@Injectable()
export class AdvisoryFeedbackPromptChainingService {
  private readonly logger = new Logger(AdvisoryFeedbackPromptChainingService.name);

  constructor(
    @Inject(GEMINI_SENTIMENT_ANALYSIS_MODEL) private analysisModel: GenerativeModel,
    @Inject(GEMINI_REPLY_MODEL) private replyModel: GenerativeModel,
    @Inject(GEMINI_FIND_LANGUAGE_MODEL) private findLanguageModel: GenerativeModel,
  ) {}

  async generateReply(prompt: string): Promise<string> {
    try {
      const [analysis, language] = await Promise.all([this.analyseSentinment(prompt), this.findLanguage(prompt)]);
      const { sentiment, topic } = analysis;
      const chainedPrompt = `
        The customer wrote a ${sentiment} feedback about ${topic} in ${language}. Provided feedback: ${prompt}.
        Feedback: 
      `;

      this.logger.log(chainedPrompt);
      const result = await this.replyModel.generateContent(chainedPrompt);
      const response = await result.response;
      const text = response.text();
      this.logger.log(text);
      return text;
    } catch (ex) {
      console.error(ex);
      throw ex;
    }
  }

  private async analyseSentinment(prompt: string): Promise<SentimentAnalysis> {
    try {
      const result = await this.analysisModel.generateContent(prompt);
      const response = await result.response;
      return JSON.parse(response.text()) as SentimentAnalysis;
    } catch (ex) {
      console.error(ex);
      throw ex;
    }
  }

  private async findLanguage(prompt: string): Promise<string> {
    try {
      const languageResult = await this.findLanguageModel.generateContent(prompt);
      const languageResponse = await languageResult.response;
      return languageResponse.text();
    } catch (ex) {
      console.error(ex);
      throw ex;
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

AdvisoryFeedbackPromptChainingService injects three chat models in the constructor.

  • findLanguageModel - A chat model for detecting the language of the feedback.
  • analysisModel - A chat model declared for determining the sentiment (POSITIVE, NEUTRAL, NEGATIVE) and the topics of the feedback.

  • replyModel - A chat model declared for generating replies with prompt chaining.

  • findLanguage - a private method that uses the findLanguageModel to detect the language of the feedback.

  • analyseSentinment - a private method that uses the analysisModel to determine the sentiment and the topics of the feedback.

  • generateReply - this method uses the results of findLanguage and analyseSentinment to construct a chained prompt. Then, it uses the replyModel to generate replies in the same language based on sentiment and topics.

The process for generating replies with prompt chaining ended by producing the text output from generateReply. The method asked questions iteratively and wrote a clear prompt for the LLM to draft a reply that was polite and addressed the need of the customer.

// advisory-feedback.service.ts

// Omit the import statements to save space

@Injectable()
export class AdvisoryFeedbackService {
  constructor(private promptChainingService: AdvisoryFeedbackPromptChainingService) {}

  generateReply(prompt: string): Promise<string> {
    return this.promptChainingService.generateReply(prompt);
  }
}
Enter fullscreen mode Exit fullscreen mode

AdvisoryFeedbackService injects AdvisoryFeedbackPromptChainingService and derives chained prompts to ask chat models to generate a reply.

Implement Advisory Feedback Controller 

// advisory-feedback.controller.ts

// Omit the import statements to save space

@Controller('esg-advisory-feedback')
export class AdvisoryFeedbackController {
  constructor(private service: AdvisoryFeedbackService) {}

  @Post()
  generateReply(@Body() dto: FeedbackDto): Promise<string> {
    return this.service.generateReply(dto.prompt);
  }
}
Enter fullscreen mode Exit fullscreen mode

The AdvisoryFeedbackController injects AdvisoryFeedbackService using Gemini API and Gemini 1.5 Pro model. The endpoint invokes the method to generate a reply from the prompt.

  • /esg-advisory-feedback - generate a reply from a prompt Module Registration

The AdvisoryFeedbackModule provides AdvisoryFeedbackPromptChainingService, AdvisoryFeedbackService, GeminiSentimentAnalysisProvider, GeminiAdvisoryFeedbackProvider and GeminiFindLanguageProvider. The module has one controller that is AdvisoryFeedbackController.

// advisory-feedback.module.ts

// Omit the import statements due to brevity reason 

@Module({
  controllers: [AdvisoryFeedbackController],
  providers: [
    AdvisoryFeedbackPromptChainingService,
    AdvisoryFeedbackService,
    GeminiSentimentAnalysisProvider,
    GeminiReplyProvider,,
    GeminiFindLanguageProvider,
  ],
})
export class AdvisoryFeedbackModule {}
Enter fullscreen mode Exit fullscreen mode

Import AdvisoryFeedbackModule into AppModule. 

// app.module.ts

@Module({
  imports: [throttlerConfig, AdvisoryFeedbackModule],
  controllers: [AppController],
  providers: [
    {
      provide: APP_GUARD,
      useClass: ThrottlerGuard,
    },
  ],
})
export class AppModule {}
Enter fullscreen mode Exit fullscreen mode

Test the endpoints

I can test the endpoints with cURL, Postman or Swagger documentation after launching the application.

npm run start:dev
Enter fullscreen mode Exit fullscreen mode

The URL of the Swagger documentation is http://localhost:3000/api.

In cURL

curl --location 'http://localhost:3000/esg-advisory-feedback' \
--header 'Content-Type: application/json' \
--data '{
    "prompt": "Looking ahead, the needs of our customers will increasingly be defined by sustainable choices. ESG reporting through diginex has brought us uniformity, transparency and direction. It provides us with a framework to be able to demonstrate to all stakeholders - customers, employees, and investors - what we are doing and to be open and transparent."
}'
Enter fullscreen mode Exit fullscreen mode

Dockerize the application 

// .dockerignore

.git
.gitignore
node_modules/
dist/
Dockerfile
.dockerignore
npm-debug.log
Enter fullscreen mode Exit fullscreen mode

Create a .dockerignore file for Docker to ignore some files and directories.

// Dockerfile

# Use an official Node.js runtime as the base image
FROM node:20-alpine

# Set the working directory in the container
WORKDIR /app

# Copy package.json and package-lock.json to the working directory
COPY package*.json ./

# Install the dependencies
RUN npm install

# Copy the rest of the application code to the working directory
COPY . .

# Expose a port (if your application listens on a specific port)
EXPOSE 3000

# Define the command to run your application
CMD [ "npm", "run", "start:dev"]
Enter fullscreen mode Exit fullscreen mode

I added the Dockerfile that installs the dependencies, builds the NestJS application, and starts it at port 3000.

// docker-compose.yaml

version: '3.8'

services:
  backend:
    build:
      context: .
      dockerfile: Dockerfile
    environment:
      - PORT=${PORT}
      - GOOGLE_GEMINI_API_KEY=${GOOGLE_GEMINI_API_KEY}
      - GOOGLE_GEMINI_MODEL=${GOOGLE_GEMINI_MODEL}
    ports:
      - "${PORT}:${PORT}"
    networks:
      - ai
    restart: unless-stopped
networks:
  ai:
Enter fullscreen mode Exit fullscreen mode

I added the docker-compose.yaml in the current folder, which was responsible for creating the NestJS application container.

Launch the Docker application 

docker-compose up
Enter fullscreen mode Exit fullscreen mode

Navigate to http://localhost:3000/api to read and execute the API.

This concludes my blog post about using Gemini API and Gemini 1.5 Pro model to tackle generating replies regardless the written languages. Generating replies with prompt chaining reduces the efforts that a writer needs to compose a polite reply to any customer. I only scratched the surface of the capability of Gemini API and Gemini 1.5 Pro model because Gemini 1.5 Pro model can not only understand text inputs but also multimedia such as images and audios. I hope you like the content and continue to follow my learning experience in Angular, NestJS, Generative AI, and other technologies.

Resources:

Image
A Time-Travelling Journey A Conversation With People About Compensation For Asbestos Exposure 20 Years Ago
Image
Hal-Hal yang Harus Menjadi perhatian Buat Pemain Pemula Judi Casino Online
Image
14 Savvy Ways To Spend Leftover Asbestos Payout Budget
Image
teknik untung besar di dalam permainan slots online
Image
How to Learn TypeScript in One Day?

javascript, typescript, beginners, tutorial
Image
Wow, pnpm, You’re Really Fast

devops, architecture, cloud, developer
Image
Building an AI Cron Builder with Cloudflare Pages and Next.js

ai, cloudflare, pages, nextjs
Image
A Brief History Of Injury Compensation In 10 Milestones
Image
Learn This Concept to be proficient in React Js.

react, webdev, javascript, beginners
Image
sejumlah metoda deposit pada website slots online
Image
What is Sepolia? A Beginner’s Guide to Ethereum Test Networks

blockchain, cryptocurrency, sepolia, nodes
Image
Teknik Simpel Temui Situs Permainan Bandar Judi Online Sah
Image
blabla yoyo ofgisjfioafj
Image
How Does Machine Learns (Intro to Machine Learning)

machinelearning, ai, datascience, beginners
Image
Lawyers For Asbestos Cases: The Ultimate Guide To Lawyers For Asbestos Cases
Image
15 Best Documentaries About Asbestos Lawsuit Settlement
Image
Kotlin Type Inference vs. Java: A Deductive Dance (Where Kotlin Takes the Lead!)

kotlin, java
Image
24/365 Days | ¥10M Japan Job Challenge

career, japan
Image
Automating SQL Script Execution Across Multiple Databases
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .