shop-api/src/modules/chatbot/providers/llm.service.ts

import { inject, injectable } from "inversify";
import OpenAI from "openai";

import { DataContextService } from "./data-context.service";
import { Logger } from "../../../core/logging/logger";
import { IOCTYPES } from "../../../IOC/ioc.types";
import { CHATBOT_CONSTANTS } from "../constants/chatbot.constants";
import { IChatContext, IChatbotResponse, ILLMConfig } from "../interfaces/chatbot.interface";

@injectable()
export class LLMService {
  private readonly logger: Logger;
  private openai: OpenAI;
  private config: ILLMConfig;

  constructor(@inject(IOCTYPES.ChatbotDataContextService) private dataContextService: DataContextService) {
    this.logger = new Logger("LLMService");
    this.config = {
      model: process.env.OPENAI_MODEL || "openai/gpt-4o-mini",
      temperature: Number(process.env.OPENAI_TEMPERATURE || "0.7"),
      maxTokens: Number(process.env.OPENAI_MAX_TOKENS || "2000"),
      topP: Number(process.env.OPENAI_TOP_P || "0.95"),
      apiKey:
        process.env.OPENAI_API_KEY ||
        "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJrZXkiOiI2OTFlMzBjOGI1NDRmYzJkY2JlMTQ2MGUiLCJ0eXBlIjoiYWlfa2V5IiwiaWF0IjoxNzYzNTg2MjQ4fQ.XRMe-TVV9FXCMdueI_xbxycBLj7KJEACXYlkxv3QZrE",
    };

    if (!this.config.apiKey) {
      throw new Error("OPENAI_API_KEY is required");
    }

    const baseURL = process.env.OPENAI_BASE_URL || "https://ai.liara.ir/api/v1/691e30a204c9b93ad278578b";

    this.openai = new OpenAI({
      apiKey: this.config.apiKey,
      baseURL: baseURL,
    });
  }

  async generateResponse(message: string, context: IChatContext): Promise<IChatbotResponse> {
    try {
      // Get relevant data from database
      const relevantData = await this.dataContextService.getRelevantContext(message, context);

      // Build system instruction with context
      const systemInstruction = this.buildSystemInstruction(relevantData);

      // Build conversation history
      const conversationHistory = this.buildConversationHistory(context);

      // Build messages array for OpenAI
      const messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[] = [
        {
          role: "system",
          content: systemInstruction,
        },
        ...conversationHistory,
        {
          role: "user",
          content: message,
        },
      ];

      // Generate response using OpenAI
      const completion = await this.openai.chat.completions.create({
        model: this.config.model,
        messages,
        temperature: this.config.temperature,
        max_tokens: this.config.maxTokens,
        top_p: this.config.topP,
      });

      const botMessage =
        completion.choices[0]?.message?.content || "متأسفم، نمی‌توانم پاسخی تولید کنم. لطفاً سوال خود را دوباره مطرح کنید. 🤖";

      // Calculate token usage from OpenAI response
      const tokensUsed = (completion.usage?.total_tokens || 0) + this.estimateTokens(systemInstruction);

      return {
        message: botMessage,
        confidence: this.calculateConfidence(completion),
        sources: relevantData.sources,
        tokensUsed,
        context: {
          model: this.config.model,
          relevantDataFound: relevantData.data.length > 0,
          finishReason: completion.choices[0]?.finish_reason || "unknown",
          usage: completion.usage,
        },
      };
    } catch (error) {
      this.logger.error("Failed to generate OpenAI response", error);
      throw new Error("Failed to generate response from AI service");
    }
  }

  async generateStreamResponse(message: string, context: IChatContext): Promise<AsyncIterable<string>> {
    try {
      // Get relevant data from database
      const relevantData = await this.dataContextService.getRelevantContext(message, context);

      // Build system instruction with context
      const systemInstruction = this.buildSystemInstruction(relevantData);

      // Build conversation history
      const conversationHistory = this.buildConversationHistory(context);

      // Build messages array for OpenAI
      const messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[] = [
        {
          role: "system",
          content: systemInstruction,
        },
        ...conversationHistory,
        {
          role: "user",
          content: message,
        },
      ];

      // Generate streaming response using OpenAI
      const stream = await this.openai.chat.completions.create({
        model: this.config.model,
        messages,
        temperature: this.config.temperature,
        max_tokens: this.config.maxTokens,
        top_p: this.config.topP,
        stream: true,
      });

      return this.createAsyncIterableFromStream(stream);
    } catch (error) {
      this.logger.error("Failed to generate streaming OpenAI response", error);
      throw new Error("Failed to generate streaming response from AI service");
    }
  }

  private buildSystemInstruction(relevantData: { data: string[]; sources: string[] }): string {
    let instruction = CHATBOT_CONSTANTS.SYSTEM_PROMPT;

    // Add relevant data context
    if (relevantData.data.length > 0) {
      instruction += "\n\nRelevant information from our database:\n";
      instruction += relevantData.data.map((data, index) => `${index + 1}. ${data}`).join("\n");
      instruction += "\n\nUse this information to provide accurate and specific answers.";
    }

    return instruction;
  }

  private buildConversationHistory(context: IChatContext): OpenAI.Chat.Completions.ChatCompletionMessageParam[] {
    if (!context.conversationHistory || context.conversationHistory.length === 0) {
      return [];
    }

    const history: OpenAI.Chat.Completions.ChatCompletionMessageParam[] = [];
    const recentHistory = context.conversationHistory
      .slice(-CHATBOT_CONSTANTS.MAX_CONVERSATION_HISTORY)
      .filter((msg) => msg.type !== "system");

    for (const msg of recentHistory) {
      history.push({
        role: msg.type === "user" ? "user" : "assistant",
        content: msg.content,
      });
    }

    return history;
  }

  private async *createAsyncIterableFromStream(stream: AsyncIterable<OpenAI.Chat.Completions.ChatCompletionChunk>): AsyncIterable<string> {
    try {
      for await (const chunk of stream) {
        const content = chunk.choices[0]?.delta?.content;
        if (content) {
          yield content;
        }
      }
    } catch (error) {
      this.logger.error("Error in streaming response", error);
      throw new Error("Failed to process streaming response");
    }
  }

  private calculateConfidence(completion: OpenAI.Chat.Completions.ChatCompletion): number {
    const text = completion.choices[0]?.message?.content || "";

    // Check finish reason - lower confidence if stopped early
    const finishReason = completion.choices[0]?.finish_reason;
    if (finishReason === "length" || finishReason === "content_filter") {
      return 0.4;
    }

    // Check for uncertainty indicators
    if (
      text.includes("I don't know") ||
      text.includes("I'm not sure") ||
      text.includes("uncertain") ||
      text.includes("نمی‌دانم") ||
      text.includes("مطمئن نیستم")
    ) {
      return 0.3;
    }

    // Check response length and quality
    if (text.length < 50) {
      return 0.6;
    }

    // Check if response uses provided data
    if (text.includes("based on") || text.includes("according to") || text.includes("بر اساس") || text.includes("طبق")) {
      return 0.95;
    }

    return 0.8;
  }

  private estimateTokens(text: string): number {
    // Rough estimation: 1 token ≈ 4 characters for English text
    // OpenAI uses similar tokenization to other models
    return Math.ceil(text.length / 4);
  }
}