totem-llm 0.1.0

package/server/utils/AiProviders/apipie/index.js

@@ -0,0 +1,386 @@
+const { NativeEmbedder } = require("../../EmbeddingEngines/native");
+const { v4: uuidv4 } = require("uuid");
+const {
+  writeResponseChunk,
+  clientAbortedHandler,
+  formatChatHistory,
+} = require("../../helpers/chat/responses");
+const fs = require("fs");
+const path = require("path");
+const { safeJsonParse } = require("../../http");
+const {
+  LLMPerformanceMonitor,
+} = require("../../helpers/chat/LLMPerformanceMonitor");
+
+const cacheFolder = path.resolve(
+  process.env.STORAGE_DIR
+    ? path.resolve(process.env.STORAGE_DIR, "models", "apipie")
+    : path.resolve(__dirname, `../../../storage/models/apipie`)
+);
+
+class ApiPieLLM {
+  constructor(embedder = null, modelPreference = null) {
+    if (!process.env.APIPIE_LLM_API_KEY)
+      throw new Error("No ApiPie LLM API key was set.");
+
+    this.className = "ApiPieLLM";
+    const { OpenAI: OpenAIApi } = require("openai");
+    this.basePath = "https://apipie.ai/v1";
+    this.openai = new OpenAIApi({
+      baseURL: this.basePath,
+      apiKey: process.env.APIPIE_LLM_API_KEY ?? null,
+    });
+    this.model =
+      modelPreference ||
+      process.env.APIPIE_LLM_MODEL_PREF ||
+      "openrouter/mistral-7b-instruct";
+    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;
+
+    if (!fs.existsSync(cacheFolder))
+      fs.mkdirSync(cacheFolder, { recursive: true });
+    this.cacheModelPath = path.resolve(cacheFolder, "models.json");
+    this.cacheAtPath = path.resolve(cacheFolder, ".cached_at");
+  }
+
+  log(text, ...args) {
+    console.log(`\x1b[36m[${this.className}]\x1b[0m ${text}`, ...args);
+  }
+
+  // This checks if the .cached_at file has a timestamp that is more than 1Week (in millis)
+  // from the current date. If it is, then we will refetch the API so that all the models are up
+  // to date.
+  #cacheIsStale() {
+    const MAX_STALE = 6.048e8; // 1 Week in MS
+    if (!fs.existsSync(this.cacheAtPath)) return true;
+    const now = Number(new Date());
+    const timestampMs = Number(fs.readFileSync(this.cacheAtPath));
+    return now - timestampMs > MAX_STALE;
+  }
+
+  // This function fetches the models from the ApiPie API and caches them locally.
+  // We do this because the ApiPie API has a lot of models, and we need to get the proper token context window
+  // for each model and this is a constructor property - so we can really only get it if this cache exists.
+  // We used to have this as a chore, but given there is an API to get the info - this makes little sense.
+  // This might slow down the first request, but we need the proper token context window
+  // for each model and this is a constructor property - so we can really only get it if this cache exists.
+  async #syncModels() {
+    if (fs.existsSync(this.cacheModelPath) && !this.#cacheIsStale())
+      return false;
+
+    this.log("Model cache is not present or stale. Fetching from ApiPie API.");
+    await fetchApiPieModels();
+    return;
+  }
+
+  #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("")
+    );
+  }
+
+  models() {
+    if (!fs.existsSync(this.cacheModelPath)) return {};
+    return safeJsonParse(
+      fs.readFileSync(this.cacheModelPath, { encoding: "utf-8" }),
+      {}
+    );
+  }
+
+  chatModels() {
+    const allModels = this.models();
+    return Object.entries(allModels).reduce(
+      (chatModels, [modelId, modelInfo]) => {
+        // Filter for chat models
+        if (
+          modelInfo.subtype &&
+          (modelInfo.subtype.includes("chat") ||
+            modelInfo.subtype.includes("chatx"))
+        ) {
+          chatModels[modelId] = modelInfo;
+        }
+        return chatModels;
+      },
+      {}
+    );
+  }
+
+  streamingEnabled() {
+    return "streamGetChatCompletion" in this;
+  }
+
+  static promptWindowLimit(modelName) {
+    const cacheModelPath = path.resolve(cacheFolder, "models.json");
+    const availableModels = fs.existsSync(cacheModelPath)
+      ? safeJsonParse(
+          fs.readFileSync(cacheModelPath, { encoding: "utf-8" }),
+          {}
+        )
+      : {};
+    return availableModels[modelName]?.maxLength || 4096;
+  }
+
+  promptWindowLimit() {
+    const availableModels = this.chatModels();
+    return availableModels[this.model]?.maxLength || 4096;
+  }
+
+  async isValidChatCompletionModel(model = "") {
+    await this.#syncModels();
+    const availableModels = this.chatModels();
+    return availableModels.hasOwnProperty(model);
+  }
+
+  /**
+   * 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,
+          detail: "auto",
+        },
+      });
+    }
+    return content.flat();
+  }
+
+  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 }) {
+    if (!(await this.isValidChatCompletionModel(this.model)))
+      throw new Error(
+        `ApiPie chat: ${this.model} is not valid for chat completion!`
+      );
+
+    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;
+
+    return {
+      textResponse: result.output.choices[0].message.content,
+      metrics: {
+        prompt_tokens: result.output.usage?.prompt_tokens || 0,
+        completion_tokens: result.output.usage?.completion_tokens || 0,
+        total_tokens: result.output.usage?.total_tokens || 0,
+        outputTps:
+          (result.output.usage?.completion_tokens || 0) / result.duration,
+        duration: result.duration,
+        model: this.model,
+        provider: this.className,
+        timestamp: new Date(),
+      },
+    };
+  }
+
+  async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
+    if (!(await this.isValidChatCompletionModel(this.model)))
+      throw new Error(
+        `ApiPie chat: ${this.model} is not valid for chat completion!`
+      );
+
+    const measuredStreamRequest = await LLMPerformanceMonitor.measureStream({
+      func: this.openai.chat.completions.create({
+        model: this.model,
+        stream: true,
+        messages,
+        temperature,
+      }),
+      messages,
+      runPromptTokenCalculation: true,
+      modelTag: this.model,
+      provider: this.className,
+    });
+    return measuredStreamRequest;
+  }
+
+  handleStream(response, stream, responseProps) {
+    const { uuid = uuidv4(), sources = [] } = responseProps;
+
+    return new Promise(async (resolve) => {
+      let fullText = "";
+
+      // Establish listener to early-abort a streaming response
+      // in case things go sideways or the user does not like the response.
+      // We preserve the generated text but continue as if chat was completed
+      // to preserve previously generated content.
+      const handleAbort = () => {
+        stream?.endMeasurement({
+          completion_tokens: LLMPerformanceMonitor.countTokens(fullText),
+        });
+        clientAbortedHandler(resolve, fullText);
+      };
+      response.on("close", handleAbort);
+
+      try {
+        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,
+            });
+          }
+
+          if (message === undefined || message.finish_reason !== null) {
+            writeResponseChunk(response, {
+              uuid,
+              sources,
+              type: "textResponseChunk",
+              textResponse: "",
+              close: true,
+              error: false,
+            });
+            response.removeListener("close", handleAbort);
+            stream?.endMeasurement({
+              completion_tokens: LLMPerformanceMonitor.countTokens(fullText),
+            });
+            resolve(fullText);
+          }
+        }
+      } catch (e) {
+        writeResponseChunk(response, {
+          uuid,
+          sources,
+          type: "abort",
+          textResponse: null,
+          close: true,
+          error: e.message,
+        });
+        response.removeListener("close", handleAbort);
+        stream?.endMeasurement({
+          completion_tokens: LLMPerformanceMonitor.countTokens(fullText),
+        });
+        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);
+  }
+}
+
+async function fetchApiPieModels(providedApiKey = null) {
+  const apiKey = providedApiKey || process.env.APIPIE_LLM_API_KEY || null;
+  return await fetch(`https://apipie.ai/v1/models`, {
+    method: "GET",
+    headers: {
+      "Content-Type": "application/json",
+      ...(apiKey ? { Authorization: `Bearer ${apiKey}` } : {}),
+    },
+  })
+    .then((res) => res.json())
+    .then(({ data = [] }) => {
+      const models = {};
+      data.forEach((model) => {
+        models[`${model.provider}/${model.model}`] = {
+          id: `${model.provider}/${model.model}`,
+          name: `${model.provider}/${model.model}`,
+          organization: model.provider,
+          subtype: model.subtype,
+          maxLength: model.max_tokens,
+        };
+      });
+
+      // Cache all response information
+      if (!fs.existsSync(cacheFolder))
+        fs.mkdirSync(cacheFolder, { recursive: true });
+      fs.writeFileSync(
+        path.resolve(cacheFolder, "models.json"),
+        JSON.stringify(models),
+        {
+          encoding: "utf-8",
+        }
+      );
+      fs.writeFileSync(
+        path.resolve(cacheFolder, ".cached_at"),
+        String(Number(new Date())),
+        {
+          encoding: "utf-8",
+        }
+      );
+
+      return models;
+    })
+    .catch((e) => {
+      console.error(e);
+      return {};
+    });
+}
+
+module.exports = {
+  ApiPieLLM,
+  fetchApiPieModels,
+};

package/server/utils/AiProviders/azureOpenAi/index.js

@@ -0,0 +1,232 @@
+const { NativeEmbedder } = require("../../EmbeddingEngines/native");
+const {
+  formatChatHistory,
+  handleDefaultStreamResponseV2,
+} = require("../../helpers/chat/responses");
+const {
+  LLMPerformanceMonitor,
+} = require("../../helpers/chat/LLMPerformanceMonitor");
+
+class AzureOpenAiLLM {
+  constructor(embedder = null, modelPreference = null) {
+    const { OpenAI } = require("openai");
+    if (!process.env.AZURE_OPENAI_ENDPOINT)
+      throw new Error("No Azure API endpoint was set.");
+    if (!process.env.AZURE_OPENAI_KEY)
+      throw new Error("No Azure API key was set.");
+
+    this.className = "AzureOpenAiLLM";
+    this.openai = new OpenAI({
+      apiKey: process.env.AZURE_OPENAI_KEY,
+      baseURL: AzureOpenAiLLM.formatBaseUrl(process.env.AZURE_OPENAI_ENDPOINT),
+    });
+    this.model =
+      modelPreference ||
+      process.env.AZURE_OPENAI_MODEL_PREF ||
+      process.env.OPEN_MODEL_PREF;
+    /* 
+      Note: Azure OpenAI deployments do not expose model metadata that would allow us to
+      programmatically detect whether the deployment uses a reasoning model (o1, o1-mini, o3-mini, etc.).
+      As a result, we rely on the user to explicitly set AZURE_OPENAI_MODEL_TYPE="reasoning"
+      when using reasoning models, as incorrect configuration might result in chat errors.
+    */
+    this.isOTypeModel =
+      process.env.AZURE_OPENAI_MODEL_TYPE === "reasoning" || false;
+    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(
+      `Initialized. Model "${this.model}" @ ${this.promptWindowLimit()} tokens.\nAPI-Version: ${this.apiVersion}.\nModel Type: ${this.isOTypeModel ? "reasoning" : "default"}`
+    );
+  }
+
+  /**
+   * Formats the Azure OpenAI endpoint URL to the correct format.
+   * @param {string} azureOpenAiEndpoint - The Azure OpenAI endpoint URL.
+   * @returns {string} The formatted URL.
+   */
+  static formatBaseUrl(azureOpenAiEndpoint) {
+    try {
+      const url = new URL(azureOpenAiEndpoint);
+      url.pathname = "/openai/v1";
+      url.protocol = "https";
+      url.search = "";
+      url.hash = "";
+      return url.href;
+    } catch {
+      throw new Error(
+        `"${azureOpenAiEndpoint}" is not a valid URL. Check your settings for the Azure OpenAI provider and set a valid endpoint URL.`
+      );
+    }
+  }
+
+  #log(text, ...args) {
+    console.log(`\x1b[32m[AzureOpenAi]\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 ${i}]\n\n`;
+        })
+        .join("")
+    );
+  }
+
+  streamingEnabled() {
+    return "streamGetChatCompletion" in this;
+  }
+
+  static promptWindowLimit(_modelName) {
+    return !!process.env.AZURE_OPENAI_TOKEN_LIMIT
+      ? Number(process.env.AZURE_OPENAI_TOKEN_LIMIT)
+      : 4096;
+  }
+
+  // Sure the user selected a proper value for the token limit
+  // could be any of these https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/models#gpt-4-models
+  // and if undefined - assume it is the lowest end.
+  promptWindowLimit() {
+    return !!process.env.AZURE_OPENAI_TOKEN_LIMIT
+      ? Number(process.env.AZURE_OPENAI_TOKEN_LIMIT)
+      : 4096;
+  }
+
+  isValidChatCompletionModel(_modelName = "") {
+    // The Azure user names their "models" as deployments and they can be any name
+    // so we rely on the user to put in the correct deployment as only they would
+    // know it.
+    return true;
+  }
+
+  /**
+   * Generates appropriate content array for a message + attachments.
+   * @param {{userPrompt:string, attachments: import("../../helpers").Attachment[]}}
+   * @returns {string|object[]}
+   */
+  #generateContent({ userPrompt, attachments = [] }) {
+    if (!attachments.length) {
+      return userPrompt;
+    }
+
+    const content = [{ type: "text", text: userPrompt }];
+    for (let attachment of attachments) {
+      content.push({
+        type: "image_url",
+        image_url: {
+          url: attachment.contentString,
+        },
+      });
+    }
+    return content.flat();
+  }
+
+  constructPrompt({
+    systemPrompt = "",
+    contextTexts = [],
+    chatHistory = [],
+    userPrompt = "",
+    attachments = [], // This is the specific attachment for only this prompt
+  }) {
+    const prompt = {
+      role: this.isOTypeModel ? "user" : "system",
+      content: `${systemPrompt}${this.#appendContext(contextTexts)}`,
+    };
+    return [
+      prompt,
+      ...formatChatHistory(chatHistory, this.#generateContent),
+      {
+        role: "user",
+        content: this.#generateContent({ userPrompt, attachments }),
+      },
+    ];
+  }
+
+  async getChatCompletion(messages = [], { temperature = 0.7 }) {
+    if (!this.model)
+      throw new Error(
+        "No AZURE_OPENAI_MODEL_PREF ENV defined. This must the name of a deployment on your Azure account for an LLM chat model like GPT-3.5."
+      );
+
+    const result = await LLMPerformanceMonitor.measureAsyncFunction(
+      this.openai.chat.completions.create({
+        messages,
+        model: this.model,
+        ...(this.isOTypeModel ? {} : { temperature }),
+      })
+    );
+
+    if (
+      !result.output.hasOwnProperty("choices") ||
+      result.output.choices.length === 0
+    )
+      return null;
+
+    return {
+      textResponse: result.output.choices[0].message.content,
+      metrics: {
+        prompt_tokens: result.output.usage.prompt_tokens || 0,
+        completion_tokens: result.output.usage.completion_tokens || 0,
+        total_tokens: result.output.usage.total_tokens || 0,
+        outputTps: result.output.usage.completion_tokens / result.duration,
+        duration: result.duration,
+        model: this.model,
+        provider: this.className,
+        timestamp: new Date(),
+      },
+    };
+  }
+
+  async streamGetChatCompletion(messages = [], { temperature = 0.7 }) {
+    if (!this.model)
+      throw new Error(
+        "No AZURE_OPENAI_MODEL_PREF ENV defined. This must the name of a deployment on your Azure account for an LLM chat model like GPT-3.5."
+      );
+
+    const measuredStreamRequest = await LLMPerformanceMonitor.measureStream({
+      func: await this.openai.chat.completions.create({
+        messages,
+        model: this.model,
+        ...(this.isOTypeModel ? {} : { temperature }),
+        n: 1,
+        stream: true,
+      }),
+      messages,
+      runPromptTokenCalculation: true,
+      modelTag: this.model,
+      provider: this.className,
+    });
+
+    return measuredStreamRequest;
+  }
+
+  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 = {
+  AzureOpenAiLLM,
+};