anything-llm/server/utils/AiProviders/openAi/index.js

const { OpenAiEmbedder } = require("../../EmbeddingEngines/openAi");
const {
  handleDefaultStreamResponseV2,
} = require("../../helpers/chat/responses");

class OpenAiLLM {
  constructor(embedder = null, modelPreference = null) {
    if (!process.env.OPEN_AI_KEY) throw new Error("No OpenAI API key was set.");
    const { OpenAI: OpenAIApi } = require("openai");

    this.openai = new OpenAIApi({
      apiKey: process.env.OPEN_AI_KEY,
    });
    this.model =
      modelPreference || process.env.OPEN_MODEL_PREF || "gpt-3.5-turbo";
    this.limits = {
      history: this.promptWindowLimit() * 0.15,
      system: this.promptWindowLimit() * 0.15,
      user: this.promptWindowLimit() * 0.7,
    };

    if (!embedder)
      console.warn(
        "No embedding provider defined for OpenAiLLM - falling back to OpenAiEmbedder for embedding!"
      );
    this.embedder = !embedder ? new OpenAiEmbedder() : embedder;
    this.defaultTemp = 0.7;
  }

  #appendContext(contextTexts = []) {
    if (!contextTexts || !contextTexts.length) return "";
    return (
      "\nContext:\n" +
      contextTexts
        .map((text, i) => {
          return `[CONTEXT ${i}]:\n${text}\n[END CONTEXT ${i}]\n\n`;
        })
        .join("")
    );
  }

  streamingEnabled() {
    return "streamGetChatCompletion" in this;
  }

  promptWindowLimit() {
    switch (this.model) {
      case "gpt-3.5-turbo":
      case "gpt-3.5-turbo-1106":
        return 16_385;
      case "gpt-4-turbo":
      case "gpt-4-1106-preview":
      case "gpt-4-turbo-preview":
        return 128_000;
      case "gpt-4":
        return 8_192;
      case "gpt-4-32k":
        return 32_000;
      default:
        return 4_096; // assume a fine-tune 3.5?
    }
  }

  // Short circuit if name has 'gpt' since we now fetch models from OpenAI API
  // via the user API key, so the model must be relevant and real.
  // and if somehow it is not, chat will fail but that is caught.
  // we don't want to hit the OpenAI api every chat because it will get spammed
  // and introduce latency for no reason.
  async isValidChatCompletionModel(modelName = "") {
    const isPreset = modelName.toLowerCase().includes("gpt");
    if (isPreset) return true;

    const model = await this.openai.models
      .retrieve(modelName)
      .then((modelObj) => modelObj)
      .catch(() => null);
    return !!model;
  }

  constructPrompt({
    systemPrompt = "",
    contextTexts = [],
    chatHistory = [],
    userPrompt = "",
  }) {
    const prompt = {
      role: "system",
      content: `${systemPrompt}${this.#appendContext(contextTexts)}`,
    };
    return [prompt, ...chatHistory, { role: "user", content: userPrompt }];
  }

  async isSafe(input = "") {
    const { flagged = false, categories = {} } = await this.openai.moderations
      .create({ input })
      .then((res) => {
        if (!res.hasOwnProperty("results"))
          throw new Error("OpenAI moderation: No results!");
        if (res.results.length === 0)
          throw new Error("OpenAI moderation: No results length!");
        return res.results[0];
      })
      .catch((error) => {
        throw new Error(
          `OpenAI::CreateModeration failed with: ${error.message}`
        );
      });

    if (!flagged) return { safe: true, reasons: [] };
    const reasons = Object.keys(categories)
      .map((category) => {
        const value = categories[category];
        if (value === true) {
          return category.replace("/", " or ");
        } else {
          return null;
        }
      })
      .filter((reason) => !!reason);

    return { safe: false, reasons };
  }

  async getChatCompletion(messages = null, { temperature = 0.7 }) {
    if (!(await this.isValidChatCompletionModel(this.model)))
      throw new Error(
        `OpenAI chat: ${this.model} is not valid for chat completion!`
      );

    const result = await this.openai.chat.completions
      .create({
        model: this.model,
        messages,
        temperature,
      })
      .catch((e) => {
        throw new Error(e.response.data.error.message);
      });

    if (!result.hasOwnProperty("choices") || result.choices.length === 0)
      return null;
    return result.choices[0].message.content;
  }

  async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
    if (!(await this.isValidChatCompletionModel(this.model)))
      throw new Error(
        `OpenAI chat: ${this.model} is not valid for chat completion!`
      );

    const streamRequest = await this.openai.chat.completions.create({
      model: this.model,
      stream: true,
      messages,
      temperature,
    });
    return streamRequest;
  }

  handleStream(response, stream, responseProps) {
    return handleDefaultStreamResponseV2(response, stream, responseProps);
  }

  // Simple wrapper for dynamic embedder & normalize interface for all LLM implementations
  async embedTextInput(textInput) {
    return await this.embedder.embedTextInput(textInput);
  }
  async embedChunks(textChunks = []) {
    return await this.embedder.embedChunks(textChunks);
  }

  async compressMessages(promptArgs = {}, rawHistory = []) {
    const { messageArrayCompressor } = require("../../helpers/chat");
    const messageArray = this.constructPrompt(promptArgs);
    return await messageArrayCompressor(this, messageArray, rawHistory);
  }
}

module.exports = {
  OpenAiLLM,
};