Merge pull request #2 from romantech/faeture/langchain-update

.env.example

@@ -1,11 +1,18 @@
+# Deployment Environment Ports
 PORT=3001
-NODE_ENV=development # koyeb 환경 변수 설정이랑 맞추기 위해 추가
+# Added to match koyeb environment variable settings
+NODE_ENV=development
+# Specify which origins to allow, separated by commas, e.g. https://google.com,http://localhost:3000
+CORS_ORIGIN=
+
 OPENAI_API_KEY=
+
 REDIS_HOST=
 REDIS_PORT=
 REDIS_USERNAME=
 REDIS_PASSWORD=
-CORS_ORIGIN=
-FINE_TUNED_ID=
 
+MODEL_GPT_3_5_FT=gpt-3.5-turbo
+MODEL_GPT_3_5=gpt-3.5-turbo
+MODEL_GPT_4=gpt-4
 

package.json

@@ -2,6 +2,7 @@
   "name": "project-server",
   "version": "1.0.0",
   "description": "Personal project server that provides a RESTful API",
+  "packageManager": "[email protected]",
   "main": "server.ts",
   "_moduleAliases": {
     "@": "dist"
@@ -20,6 +21,7 @@
   "author": "Romantech",
   "license": "ISC",
   "dependencies": {
+    "@langchain/openai": "^0.0.26",
     "@types/request-ip": "^0.0.41",
     "compression": "^1.7.4",
     "cookie-parser": "^1.4.6",
@@ -28,14 +30,14 @@
     "express-validator": "^7.0.1",
     "helmet": "^7.1.0",
     "ioredis": "^5.3.2",
-    "langchain": "^0.0.209",
+    "langchain": "^0.1.31",
     "module-alias": "^2.2.3",
     "morgan": "^1.10.0",
     "node-schedule": "^2.1.1",
-    "openai": "^4.23.0",
+    "openai": "^4.33.0",
     "request-ip": "^3.3.0",
     "winston": "^3.11.0",
-    "winston-daily-rotate-file": "^4.7.1"
+    "winston-daily-rotate-file": "^5.0.0"
   },
   "devDependencies": {
     "@types/compression": "^1.7.5",
@@ -45,10 +47,10 @@
     "@types/express-serve-static-core": "^4.17.41",
     "@types/module-alias": "^2.0.4",
     "@types/morgan": "^1.9.9",
-    "@types/node": "^20.10.5",
+    "@types/node": "^20.12.5",
     "@types/node-schedule": "^2.1.5",
-    "@typescript-eslint/eslint-plugin": "^6.15.0",
-    "@typescript-eslint/parser": "^6.15.0",
+    "@typescript-eslint/eslint-plugin": "^7.5.0",
+    "@typescript-eslint/parser": "^7.5.0",
     "dotenv": "^16.3.1",
     "eslint": "^8.56.0",
     "eslint-config-prettier": "^9.1.0",

src/config/createServer.ts

@@ -4,12 +4,12 @@ import morgan from 'morgan';
 import compression from 'compression';
 import cookieParser from 'cookie-parser';
 import cors, { CorsOptions } from 'cors';
-import { CORS_ORIGIN, isProd } from '@/config/environment';
+import { envConfig } from '@/config/environment';
 
-const morganFormat = isProd() ? 'combined' : 'dev';
+const morganFormat = envConfig.isProd ? 'combined' : 'dev';
 const corsOptions: CorsOptions = {
   /** Access-Control-Allow-Origin 응답 헤더 설정 (허용할 오리진 목록) */
-  origin: CORS_ORIGIN,
+  origin: envConfig.corsOrigins,
   /** Access-Control-Allow-Methods 응답 헤더 설정 (허용할 HTTP 메서드 목록) */
   methods: ['GET', 'POST'],
   /** Access-Control-Allow-Headers 응답 헤더 설정 (허용할 요청 헤더 목록) */

src/config/environment.ts

@@ -1,28 +1,50 @@
+/**
+ * 웹스톰에서 alias 설정을 위해 tsconfig.json 파일의 baseUrl 속성을 설정하면
+ * process.env 참조시 해결되지 않은 변수로 나오는 문제 있음
+ * import { env } from 'node:process' 구문으로 임포트해서 사용해서 임시 해결
+ * */
 import { env } from 'node:process';
 
-export enum EnvVars {
+export enum RequiredEnv { // Required Environment Variables
   PORT = 'PORT',
   NODE_ENV = 'NODE_ENV',
   OPENAI_API_KEY = 'OPENAI_API_KEY',
+}
+
+export enum OptionalEnv { // Optional Environment Variables
+  CORS_ORIGIN = 'CORS_ORIGIN',
+
   REDIS_HOST = 'REDIS_HOST',
   REDIS_PORT = 'REDIS_PORT',
   REDIS_PASSWORD = 'REDIS_PASSWORD',
   REDIS_USERNAME = 'REDIS_USERNAME',
-  CORS_ORIGIN = 'CORS_ORIGIN',
-  FINE_TUNED_ID = 'FINE_TUNED_ID',
+
+  MODEL_GPT_3_5_FT = 'MODEL_GPT_3_5_FT',
+  MODEL_GPT_3_5 = 'MODEL_GPT_3_5',
+  MODEL_GPT_4 = 'MODEL_GPT_4',
 }
 
-/**
- * 웹스톰에서 alias 설정을 위해 tsconfig.json 파일의 baseUrl 속성을 설정하면
- * process.env 참조시 해결되지 않은 변수로 나오는 문제 있음
- * import { env } from 'node:process' 구문으로 임포트해서 사용해서 임시 해결
- * */
-export const isProd = () => env[EnvVars.NODE_ENV] === 'production';
-export const PORT = isProd() ? env[EnvVars.PORT] : 3001;
-export const REDIS_HOST = env[EnvVars.REDIS_HOST];
-export const REDIS_PORT = Number(env[EnvVars.REDIS_PORT]);
-export const REDIS_PASSWORD = env[EnvVars.REDIS_PASSWORD];
-export const REDIS_USERNAME = env[EnvVars.REDIS_USERNAME];
-export const OPENAI_API_KEY = env[EnvVars.OPENAI_API_KEY];
-export const CORS_ORIGIN = env[EnvVars.CORS_ORIGIN]?.split(',') ?? [];
-export const FINE_TUNED_ID = env[EnvVars.FINE_TUNED_ID];
+const loadEnvironment = () => {
+  const isProd = env[RequiredEnv.NODE_ENV] === 'production';
+  const DEFAULT_SERVER_PORT = 3001;
+
+  return {
+    isProd,
+    port: isProd ? Number(env[RequiredEnv.PORT]) : DEFAULT_SERVER_PORT,
+    corsOrigins: env[OptionalEnv.CORS_ORIGIN]?.split(',') ?? [],
+    openAIKey: env[RequiredEnv.OPENAI_API_KEY],
+    redis: {
+      host: env[OptionalEnv.REDIS_HOST],
+      port: Number(env[OptionalEnv.REDIS_PORT]),
+      password: env[OptionalEnv.REDIS_PASSWORD],
+      username: env[OptionalEnv.REDIS_USERNAME],
+    },
+    modelNames: {
+      gpt_3_5_FT: env[OptionalEnv.MODEL_GPT_3_5_FT] ?? 'gpt-3.5-turbo',
+      gpt_3_5: env[OptionalEnv.MODEL_GPT_3_5] ?? 'gpt-3.5-turbo',
+      gpt_4: env[OptionalEnv.MODEL_GPT_4] ?? 'gpt-4',
+    },
+  } as const;
+};
+
+export const envConfig = loadEnvironment();

src/config/logger.ts

@@ -12,12 +12,26 @@ const DATE_FORMAT = 'YYYY-MM-DD HH:mm:ss';
 const ignorePrivate = format((info) => (info.private ? false : info));
 
 /** Custom logging format */
-const customFormat = printf(({ level, message, timestamp, ...meta }) => {
-  let result = `[${level}] ${timestamp}: ${message}`;
-  if (Object.keys(meta).length) {
-    result += ` ${JSON.stringify(meta)}`;
+const formatStackTrace = (stack: string, limit = 3) => {
+  const [errorMessage, ...restLines] = stack.split('\n');
+  return [
+    errorMessage.replace(': ', '-'),
+    ...restLines.slice(0, limit).map((line) => `-> ${line}`),
+  ].join('\n');
+};
+
+const customFormat = printf(({ level, message, timestamp, stack, ...meta }) => {
+  let logMessage = `[${timestamp}] [${level}]: ${message}`;
+
+  if (stack?.trim()) logMessage += `\nStack: ${formatStackTrace(stack)}`;
+
+  const metaKeys = Object.keys(meta);
+  if (metaKeys.length > 0) {
+    const metaInfo = JSON.stringify(meta, null, 2);
+    logMessage += `\nMeta: ${metaInfo}`;
   }
-  return result;
+
+  return logMessage;
 });
 
 /** Transform log level to uppercase */

src/constants/analyzer.ts

@@ -1,8 +1,8 @@
-import { GPTModel } from '@/services';
+import { AIModelKey } from '@/services';
 
 export const ANALYSIS_DECREMENT_COUNT = {
-  [GPTModel.GPT_3]: 1,
-  [GPTModel.GPT_4]: 5,
+  [AIModelKey.GPT_3_5_FT]: 1,
+  [AIModelKey.GPT_4]: 5,
 };
 
 export enum RandomSentenceParam {

src/controllers/analyzer/createAnalysis.ts

@@ -1,19 +1,20 @@
 import { ANALYSIS_DECREMENT_COUNT, ERROR_MESSAGES } from '@/constants';
 import { asyncHandler, throwCustomError } from '@/utils';
 import {
-  ANALYSIS_MODEL_OPTIONS,
+  AIModelKey,
+  ANALYSIS_MODEL_OPTION,
+  AnalysisModel,
   ANALYZER_REDIS_SCHEMA,
   decrementRedisCounters,
-  GPTModel,
   redis,
   validateAndRepairJSON,
 } from '@/services';
 import { checkModelField, checkSentenceField } from '@/validators';
 import { handleValidationErrors, validateAnalysisCount } from '@/middlewares';
-import { HumanMessage, SystemMessage } from 'langchain/schema';
-import { ChatOpenAI } from 'langchain/chat_models/openai';
+import { HumanMessage, SystemMessage } from '@langchain/core/messages';
+import { ChatOpenAI, type ChatOpenAICallOptions } from '@langchain/openai';
 
-type RequestBody = { sentence: string[]; model: GPTModel };
+type RequestBody = { sentence: string[]; model: AnalysisModel };
 
 const { RETRIEVE_FAILED, GENERATE_FAILED } = ERROR_MESSAGES;
 const { KEYS, FIELDS } = ANALYZER_REDIS_SCHEMA;
@@ -39,22 +40,28 @@ export const createAnalysis = [
   }),
 ];
 
-const retrieveAnalysisPrompt = async (model: GPTModel) => {
-  const { promptField } = ANALYSIS_MODEL_OPTIONS[model];
+const retrieveAnalysisPrompt = async (model: AnalysisModel) => {
+  const { promptField } = ANALYSIS_MODEL_OPTION[model];
   const prompt = await redis.hget(KEYS.PROMPT, promptField);
 
   if (!prompt) return throwCustomError(RETRIEVE_FAILED('prompt'), 500);
 
   return prompt;
 };
 
-const executeAnalysis = async (sentence: string, model: GPTModel) => {
-  const { temperature, modelName } = ANALYSIS_MODEL_OPTIONS[model];
+const executeAnalysis = async (sentence: string, model: AnalysisModel) => {
+  const isGPT4 = model === AIModelKey.GPT_4;
+
+  const { temperature, modelName } = ANALYSIS_MODEL_OPTION[model];
   const prompt = await retrieveAnalysisPrompt(model);
   const messages = [new SystemMessage(prompt), new HumanMessage(sentence)];
 
-  const chat = new ChatOpenAI({ modelName, temperature });
-  const { content } = await chat.call(messages);
+  const callOptions: ChatOpenAICallOptions = {};
+  // json_object response_format only support GPT-4-Turbo
+  if (isGPT4) callOptions.response_format = { type: 'json_object' };
+
+  const chat = new ChatOpenAI({ modelName, temperature }).bind(callOptions);
+  const { content } = await chat.invoke(messages);
 
   if (!content) return throwCustomError(GENERATE_FAILED('analysis'), 500);
   return await validateAndRepairJSON(`${content}`);

src/controllers/analyzer/getRandomSentences.ts

@@ -1,9 +1,10 @@
 import { asyncHandler, throwCustomError } from '@/utils';
 import { ParamsDictionary } from 'express-serve-static-core';
 import {
+  AI_MODEL,
+  AIModelKey,
   ANALYZER_REDIS_SCHEMA,
   decrementRedisCounters,
-  GPTModel,
   redis,
 } from '@/services';
 import { ERROR_MESSAGES, RandomSentenceParams } from '@/constants';
@@ -13,8 +14,8 @@ import {
   checkTopicsField,
 } from '@/validators';
 import { handleValidationErrors, validateAnalysisCount } from '@/middlewares';
-import { PromptTemplate } from 'langchain/prompts';
-import { OpenAI } from 'langchain/llms/openai';
+import { PromptTemplate } from '@langchain/core/prompts';
+import { OpenAI } from '@langchain/openai';
 
 const { RETRIEVE_FAILED, GENERATE_FAILED } = ERROR_MESSAGES;
 const { KEYS, FIELDS } = ANALYZER_REDIS_SCHEMA;
@@ -55,9 +56,12 @@ const retrieveRandomSentencePrompt = async (query: RandomSentenceParams) => {
 
 const generateRandomSentences = async (query: RandomSentenceParams) => {
   const prompt = await retrieveRandomSentencePrompt(query);
-  const llm = new OpenAI({ temperature: 1, modelName: GPTModel.GPT_3 });
+  const llm = new OpenAI({
+    temperature: 1,
+    modelName: AI_MODEL[AIModelKey.GPT_3_5],
+  });
 
-  const sentences = await llm.predict(prompt);
+  const sentences = await llm.invoke(prompt);
   if (!sentences) return throwCustomError(GENERATE_FAILED('sentence'), 500);
 
   return sentences;

src/server.ts

@@ -1,5 +1,5 @@
 import { setupRoutes } from '@/routes';
-import { createServer, logger, PORT } from '@/config';
+import { createServer, envConfig, logger } from '@/config';
 import { initRedisKeys, redis } from '@/services';
 import { errorHandler, notFoundHandler } from '@/middlewares';
 import { initSchedulers } from '@/schedulers';
@@ -31,8 +31,8 @@ const initServer = async () => {
   app.use(notFoundHandler);
   app.use(errorHandler);
 
-  app.listen(PORT, () => {
-    logger.info(`Server is running on port ${PORT}`);
+  app.listen(envConfig.port, () => {
+    logger.info(`Server is running on port ${envConfig.port}`);
   });
 };
 

src/services/openAI/openAI.ts

@@ -1,26 +1,34 @@
-import { FINE_TUNED_ID } from '@/config';
+import { envConfig } from '@/config';
 import { ANALYZER_REDIS_SCHEMA } from '@/services/redis';
 
 const { FIELDS } = ANALYZER_REDIS_SCHEMA;
 
 const FINE_TUNE_SUFFIX = '_fine_tuned';
 
-export enum GPTModel {
-  GPT_3 = 'gpt-3.5-turbo',
+export enum AIModelKey {
+  GPT_3_5 = 'gpt-3.5',
+  GPT_3_5_FT = 'gpt-3.5-ft',
   GPT_4 = 'gpt-4',
 }
 
-export const GPT_MODEL_LIST = Object.values(GPTModel);
+export const AI_MODEL = {
+  [AIModelKey.GPT_3_5]: envConfig.modelNames.gpt_3_5,
+  [AIModelKey.GPT_3_5_FT]: envConfig.modelNames.gpt_3_5_FT,
+  [AIModelKey.GPT_4]: envConfig.modelNames.gpt_4,
+};
 
-export const ANALYSIS_MODEL_OPTIONS = {
-  [GPTModel.GPT_3]: {
+export const ANALYSIS_MODEL_OPTION = {
+  [AIModelKey.GPT_3_5_FT]: {
     promptField: FIELDS.ANALYSIS + FINE_TUNE_SUFFIX,
     temperature: 0.6,
-    modelName: FINE_TUNED_ID ?? GPTModel.GPT_3,
+    modelName: AI_MODEL[AIModelKey.GPT_3_5_FT],
   },
-  [GPTModel.GPT_4]: {
+  [AIModelKey.GPT_4]: {
     promptField: FIELDS.ANALYSIS,
     temperature: 0.4,
-    modelName: GPTModel.GPT_4,
+    modelName: AI_MODEL[AIModelKey.GPT_4],
   },
-} as const;
+};
+
+export const ANALYSIS_MODEL = Object.keys(ANALYSIS_MODEL_OPTION);
+export type AnalysisModel = keyof typeof ANALYSIS_MODEL_OPTION;

src/services/openAI/openAIUtils.ts

@@ -1,7 +1,7 @@
 import { logger } from '@/config';
-import { OpenAI } from 'langchain/llms/openai';
+import { OpenAI } from '@langchain/openai';
 import { throwCustomError } from '@/utils/customError';
-import { GPTModel } from '@/services';
+import { AI_MODEL, AIModelKey } from '@/services';
 
 /**
  * When using fine-tuned model, there's a recurring issue where unnecessary ']' and '}' are added.
@@ -21,10 +21,13 @@ const repairJSONManually = (jsonString: string, errorMessage: string) => {
 const repairJSONWithOpenAI = async (jsonString: string) => {
   logger.info('Trying to repair JSON using OpenAI');
 
-  const llm = new OpenAI({ temperature: 0, modelName: GPTModel.GPT_3 });
+  const llm = new OpenAI({
+    temperature: 0,
+    modelName: AI_MODEL[AIModelKey.GPT_3_5],
+  });
   const prompt = `Fix JSON format and the results should be returned in JSON: ${jsonString}`;
 
-  const repaired = await llm.predict(prompt);
+  const repaired = await llm.invoke(prompt);
   return JSON.parse(repaired);
 };