anything-llm/server/utils/chats/index.js

const { v4: uuidv4 } = require("uuid");
const { WorkspaceChats } = require("../../models/workspaceChats");
const { resetMemory } = require("./commands/reset");
const { getVectorDbClass, getLLMProvider } = require("../helpers");
const { convertToPromptHistory } = require("../helpers/chat/responses");
const { DocumentManager } = require("../DocumentManager");
const { SlashCommandPresets } = require("../../models/slashCommandsPresets");

const VALID_COMMANDS = {
  "/reset": resetMemory,
};

async function grepCommand(message, user = null) {
  const userPresets = await SlashCommandPresets.getUserPresets(user?.id);
  const availableCommands = Object.keys(VALID_COMMANDS);

  // Check if the message starts with any built-in command
  for (let i = 0; i < availableCommands.length; i++) {
    const cmd = availableCommands[i];
    const re = new RegExp(`^(${cmd})`, "i");
    if (re.test(message)) {
      return cmd;
    }
  }

  // Replace all preset commands with their corresponding prompts
  // Allows multiple commands in one message
  let updatedMessage = message;
  for (const preset of userPresets) {
    const regex = new RegExp(preset.command, "g");
    updatedMessage = updatedMessage.replace(regex, preset.prompt);
  }

  return updatedMessage;
}

async function chatWithWorkspace(
  workspace,
  message,
  chatMode = "chat",
  user = null,
  thread = null
) {
  const uuid = uuidv4();
  const updatedMessage = await grepCommand(message, user);

  if (Object.keys(VALID_COMMANDS).includes(updatedMessage)) {
    return await VALID_COMMANDS[updatedMessage](workspace, message, uuid, user);
  }

  const LLMConnector = getLLMProvider({
    provider: workspace?.chatProvider,
    model: workspace?.chatModel,
  });
  const VectorDb = getVectorDbClass();

  const messageLimit = workspace?.openAiHistory || 20;
  const hasVectorizedSpace = await VectorDb.hasNamespace(workspace.slug);
  const embeddingsCount = await VectorDb.namespaceCount(workspace.slug);

  // User is trying to query-mode chat a workspace that has no data in it - so
  // we should exit early as no information can be found under these conditions.
  if ((!hasVectorizedSpace || embeddingsCount === 0) && chatMode === "query") {
    const textResponse =
      workspace?.queryRefusalResponse ??
      "There is no relevant information in this workspace to answer your query.";

    await WorkspaceChats.new({
      workspaceId: workspace.id,
      prompt: message,
      response: {
        text: textResponse,
        sources: [],
        type: chatMode,
      },
      threadId: thread?.id || null,
      include: false,
      user,
    });

    return {
      id: uuid,
      type: "textResponse",
      sources: [],
      close: true,
      error: null,
      textResponse,
    };
  }

  // If we are here we know that we are in a workspace that is:
  // 1. Chatting in "chat" mode and may or may _not_ have embeddings
  // 2. Chatting in "query" mode and has at least 1 embedding
  let contextTexts = [];
  let sources = [];
  let pinnedDocIdentifiers = [];
  const { rawHistory, chatHistory } = await recentChatHistory({
    user,
    workspace,
    thread,
    messageLimit,
    chatMode,
  });

  // See stream.js comment for more information on this implementation.
  await new DocumentManager({
    workspace,
    maxTokens: LLMConnector.promptWindowLimit(),
  })
    .pinnedDocs()
    .then((pinnedDocs) => {
      pinnedDocs.forEach((doc) => {
        const { pageContent, ...metadata } = doc;
        pinnedDocIdentifiers.push(sourceIdentifier(doc));
        contextTexts.push(doc.pageContent);
        sources.push({
          text:
            pageContent.slice(0, 1_000) +
            "...continued on in source document...",
          ...metadata,
        });
      });
    });

  const vectorSearchResults =
    embeddingsCount !== 0
      ? await VectorDb.performSimilaritySearch({
          namespace: workspace.slug,
          input: message,
          LLMConnector,
          similarityThreshold: workspace?.similarityThreshold,
          topN: workspace?.topN,
          filterIdentifiers: pinnedDocIdentifiers,
        })
      : {
          contextTexts: [],
          sources: [],
          message: null,
        };

  // Failed similarity search if it was run at all and failed.
  if (!!vectorSearchResults.message) {
    return {
      id: uuid,
      type: "abort",
      textResponse: null,
      sources: [],
      close: true,
      error: vectorSearchResults.message,
    };
  }

  const { fillSourceWindow } = require("../helpers/chat");
  const filledSources = fillSourceWindow({
    nDocs: workspace?.topN || 4,
    searchResults: vectorSearchResults.sources,
    history: rawHistory,
    filterIdentifiers: pinnedDocIdentifiers,
  });

  // Why does contextTexts get all the info, but sources only get current search?
  // This is to give the ability of the LLM to "comprehend" a contextual response without
  // populating the Citations under a response with documents the user "thinks" are irrelevant
  // due to how we manage backfilling of the context to keep chats with the LLM more correct in responses.
  // If a past citation was used to answer the question - that is visible in the history so it logically makes sense
  // and does not appear to the user that a new response used information that is otherwise irrelevant for a given prompt.
  // TLDR; reduces GitHub issues for "LLM citing document that has no answer in it" while keep answers highly accurate.
  contextTexts = [...contextTexts, ...filledSources.contextTexts];
  sources = [...sources, ...vectorSearchResults.sources];

  // If in query mode and no context chunks are found from search, backfill, or pins -  do not
  // let the LLM try to hallucinate a response or use general knowledge and exit early
  if (chatMode === "query" && contextTexts.length === 0) {
    const textResponse =
      workspace?.queryRefusalResponse ??
      "There is no relevant information in this workspace to answer your query.";

    await WorkspaceChats.new({
      workspaceId: workspace.id,
      prompt: message,
      response: {
        text: textResponse,
        sources: [],
        type: chatMode,
      },
      threadId: thread?.id || null,
      include: false,
      user,
    });

    return {
      id: uuid,
      type: "textResponse",
      sources: [],
      close: true,
      error: null,
      textResponse,
    };
  }

  // Compress & Assemble message to ensure prompt passes token limit with room for response
  // and build system messages based on inputs and history.
  const messages = await LLMConnector.compressMessages(
    {
      systemPrompt: chatPrompt(workspace),
      userPrompt: updatedMessage,
      contextTexts,
      chatHistory,
    },
    rawHistory
  );

  // Send the text completion.
  const textResponse = await LLMConnector.getChatCompletion(messages, {
    temperature: workspace?.openAiTemp ?? LLMConnector.defaultTemp,
  });

  if (!textResponse) {
    return {
      id: uuid,
      type: "abort",
      textResponse: null,
      sources: [],
      close: true,
      error: "No text completion could be completed with this input.",
    };
  }

  const { chat } = await WorkspaceChats.new({
    workspaceId: workspace.id,
    prompt: message,
    response: { text: textResponse, sources, type: chatMode },
    threadId: thread?.id || null,
    user,
  });
  return {
    id: uuid,
    type: "textResponse",
    close: true,
    error: null,
    chatId: chat.id,
    textResponse,
    sources,
  };
}

async function recentChatHistory({
  user = null,
  workspace,
  thread = null,
  messageLimit = 20,
}) {
  const rawHistory = (
    await WorkspaceChats.where(
      {
        workspaceId: workspace.id,
        user_id: user?.id || null,
        thread_id: thread?.id || null,
        include: true,
      },
      messageLimit,
      { id: "desc" }
    )
  ).reverse();
  return { rawHistory, chatHistory: convertToPromptHistory(rawHistory) };
}

function chatPrompt(workspace) {
  return (
    workspace?.openAiPrompt ??
    "Given the following conversation, relevant context, and a follow up question, reply with an answer to the current question the user is asking. Return only your response to the question given the above information following the users instructions as needed."
  );
}

// We use this util function to deduplicate sources from similarity searching
// if the document is already pinned.
// Eg: You pin a csv, if we RAG + full-text that you will get the same data
// points both in the full-text and possibly from RAG - result in bad results
// even if the LLM was not even going to hallucinate.
function sourceIdentifier(sourceDocument) {
  if (!sourceDocument?.title || !sourceDocument?.published) return uuidv4();
  return `title:${sourceDocument.title}-timestamp:${sourceDocument.published}`;
}

module.exports = {
  sourceIdentifier,
  recentChatHistory,
  chatWithWorkspace,
  chatPrompt,
  grepCommand,
  VALID_COMMANDS,
};