chtech-anythingllm/server/utils/AiProviders/koboldCPP/index.js

const { NativeEmbedder } = require("../../EmbeddingEngines/native");
const {
  clientAbortedHandler,
  writeResponseChunk,
  formatChatHistory,
} = require("../../helpers/chat/responses");
const {
  LLMPerformanceMonitor,
} = require("../../helpers/chat/LLMPerformanceMonitor");
const { v4: uuidv4 } = require("uuid");

class KoboldCPPLLM {
  constructor(embedder = null, modelPreference = null) {
    const { OpenAI: OpenAIApi } = require("openai");
    if (!process.env.KOBOLD_CPP_BASE_PATH)
      throw new Error(
        "KoboldCPP must have a valid base path to use for the api."
      );

    this.basePath = process.env.KOBOLD_CPP_BASE_PATH;
    this.openai = new OpenAIApi({
      baseURL: this.basePath,
      apiKey: null,
    });
    this.model = modelPreference ?? process.env.KOBOLD_CPP_MODEL_PREF ?? null;
    if (!this.model) throw new Error("KoboldCPP must have a valid model set.");
    this.limits = {
      history: this.promptWindowLimit() * 0.15,
      system: this.promptWindowLimit() * 0.15,
      user: this.promptWindowLimit() * 0.7,
    };

    this.embedder = embedder ?? new NativeEmbedder();
    this.defaultTemp = 0.7;
    this.log(`Inference API: ${this.basePath} Model: ${this.model}`);
  }

  log(text, ...args) {
    console.log(`\x1b[36m[${this.constructor.name}]\x1b[0m ${text}`, ...args);
  }

  #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;
  }

  static promptWindowLimit(_modelName) {
    const limit = process.env.KOBOLD_CPP_MODEL_TOKEN_LIMIT || 4096;
    if (!limit || isNaN(Number(limit)))
      throw new Error("No token context limit was set.");
    return Number(limit);
  }

  // Ensure the user set a value for the token limit
  // and if undefined - assume 4096 window.
  promptWindowLimit() {
    const limit = process.env.KOBOLD_CPP_MODEL_TOKEN_LIMIT || 4096;
    if (!limit || isNaN(Number(limit)))
      throw new Error("No token context limit was set.");
    return Number(limit);
  }

  // Short circuit since we have no idea if the model is valid or not
  // in pre-flight for generic endpoints
  isValidChatCompletionModel(_modelName = "") {
    return true;
  }

  /**
   * Generates appropriate content array for a message + attachments.
   * @param {{userPrompt:string, attachments: import("../../helpers").Attachment[]}}
   * @returns {string|object[]}
   */
  #generateContent({ userPrompt, attachments = [] }) {
    if (!attachments.length) {
      return userPrompt;
    }

    const content = [{ type: "text", text: userPrompt }];
    for (let attachment of attachments) {
      content.push({
        type: "image_url",
        image_url: {
          url: attachment.contentString,
        },
      });
    }
    return content.flat();
  }

  /**
   * Construct the user prompt for this model.
   * @param {{attachments: import("../../helpers").Attachment[]}} param0
   * @returns
   */
  constructPrompt({
    systemPrompt = "",
    contextTexts = [],
    chatHistory = [],
    userPrompt = "",
    attachments = [],
  }) {
    const prompt = {
      role: "system",
      content: `${systemPrompt}${this.#appendContext(contextTexts)}`,
    };
    return [
      prompt,
      ...formatChatHistory(chatHistory, this.#generateContent),
      {
        role: "user",
        content: this.#generateContent({ userPrompt, attachments }),
      },
    ];
  }

  async getChatCompletion(messages = null, { temperature = 0.7 }) {
    const result = await LLMPerformanceMonitor.measureAsyncFunction(
      this.openai.chat.completions
        .create({
          model: this.model,
          messages,
          temperature,
        })
        .catch((e) => {
          throw new Error(e.message);
        })
    );

    if (
      !result.output.hasOwnProperty("choices") ||
      result.output.choices.length === 0
    )
      return null;

    const promptTokens = LLMPerformanceMonitor.countTokens(messages);
    const completionTokens = LLMPerformanceMonitor.countTokens([
      { content: result.output.choices[0].message.content },
    ]);

    return {
      textResponse: result.output.choices[0].message.content,
      metrics: {
        prompt_tokens: promptTokens,
        completion_tokens: completionTokens,
        total_tokens: promptTokens + completionTokens,
        outputTps: completionTokens / result.duration,
        duration: result.duration,
      },
    };
  }

  async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
    const measuredStreamRequest = await LLMPerformanceMonitor.measureStream(
      this.openai.chat.completions.create({
        model: this.model,
        stream: true,
        messages,
        temperature,
      }),
      messages
    );
    return measuredStreamRequest;
  }

  handleStream(response, stream, responseProps) {
    const { uuid = uuidv4(), sources = [] } = responseProps;

    return new Promise(async (resolve) => {
      let fullText = "";
      let usage = {
        prompt_tokens: LLMPerformanceMonitor.countTokens(stream.messages || []),
        completion_tokens: 0,
      };

      const handleAbort = () => {
        usage.completion_tokens = LLMPerformanceMonitor.countTokens([
          { content: fullText },
        ]);
        stream?.endMeasurement(usage);
        clientAbortedHandler(resolve, fullText);
      };
      response.on("close", handleAbort);

      for await (const chunk of stream) {
        const message = chunk?.choices?.[0];
        const token = message?.delta?.content;

        if (token) {
          fullText += token;
          writeResponseChunk(response, {
            uuid,
            sources: [],
            type: "textResponseChunk",
            textResponse: token,
            close: false,
            error: false,
          });
        }

        // KoboldCPP finishes with "length" or "stop"
        if (
          message.finish_reason !== "null" &&
          (message.finish_reason === "length" ||
            message.finish_reason === "stop")
        ) {
          writeResponseChunk(response, {
            uuid,
            sources,
            type: "textResponseChunk",
            textResponse: "",
            close: true,
            error: false,
          });
          response.removeListener("close", handleAbort);
          usage.completion_tokens = LLMPerformanceMonitor.countTokens([
            { content: fullText },
          ]);
          stream?.endMeasurement(usage);
          resolve(fullText);
        }
      }
    });
  }

  // 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 = {
  KoboldCPPLLM,
};