mirror of
https://github.com/Mintplex-Labs/anything-llm.git
synced 2024-08-19 20:00:11 +02:00
* refactor stream/chat/embed-stram to be a single execution logic path so that it is easier to maintain and build upon * no thread in sync chat since only api uses it adjust import locations
211 lines
5.5 KiB
JavaScript
211 lines
5.5 KiB
JavaScript
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 VALID_COMMANDS = {
|
|
"/reset": resetMemory,
|
|
};
|
|
|
|
function grepCommand(message) {
|
|
const availableCommands = Object.keys(VALID_COMMANDS);
|
|
|
|
for (let i = 0; i < availableCommands.length; i++) {
|
|
const cmd = availableCommands[i];
|
|
const re = new RegExp(`^(${cmd})`, "i");
|
|
if (re.test(message)) {
|
|
return cmd;
|
|
}
|
|
}
|
|
|
|
return null;
|
|
}
|
|
|
|
async function chatWithWorkspace(
|
|
workspace,
|
|
message,
|
|
chatMode = "chat",
|
|
user = null,
|
|
thread = null
|
|
) {
|
|
const uuid = uuidv4();
|
|
const command = grepCommand(message);
|
|
|
|
if (!!command && Object.keys(VALID_COMMANDS).includes(command)) {
|
|
return await VALID_COMMANDS[command](workspace, message, uuid, user);
|
|
}
|
|
|
|
const LLMConnector = getLLMProvider(workspace?.chatModel);
|
|
const VectorDb = getVectorDbClass();
|
|
const { safe, reasons = [] } = await LLMConnector.isSafe(message);
|
|
if (!safe) {
|
|
return {
|
|
id: uuid,
|
|
type: "abort",
|
|
textResponse: null,
|
|
sources: [],
|
|
close: true,
|
|
error: `This message was moderated and will not be allowed. Violations for ${reasons.join(
|
|
", "
|
|
)} found.`,
|
|
};
|
|
}
|
|
|
|
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") {
|
|
return {
|
|
id: uuid,
|
|
type: "textResponse",
|
|
sources: [],
|
|
close: true,
|
|
error: null,
|
|
textResponse:
|
|
"There is no relevant information in this workspace to answer your query.",
|
|
};
|
|
}
|
|
|
|
// 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
|
|
const { rawHistory, chatHistory } = await recentChatHistory({
|
|
user,
|
|
workspace,
|
|
thread,
|
|
messageLimit,
|
|
chatMode,
|
|
});
|
|
|
|
const {
|
|
contextTexts = [],
|
|
sources = [],
|
|
message: error,
|
|
} = embeddingsCount !== 0 // if there no embeddings don't bother searching.
|
|
? await VectorDb.performSimilaritySearch({
|
|
namespace: workspace.slug,
|
|
input: message,
|
|
LLMConnector,
|
|
similarityThreshold: workspace?.similarityThreshold,
|
|
topN: workspace?.topN,
|
|
})
|
|
: {
|
|
contextTexts: [],
|
|
sources: [],
|
|
message: null,
|
|
};
|
|
|
|
// Failed similarity search if it was run at all and failed.
|
|
if (!!error) {
|
|
return {
|
|
id: uuid,
|
|
type: "abort",
|
|
textResponse: null,
|
|
sources: [],
|
|
close: true,
|
|
error,
|
|
};
|
|
}
|
|
|
|
// If in query mode and no sources are found, do not
|
|
// let the LLM try to hallucinate a response or use general knowledge and exit early
|
|
if (chatMode === "query" && sources.length === 0) {
|
|
return {
|
|
id: uuid,
|
|
type: "textResponse",
|
|
sources: [],
|
|
close: true,
|
|
error: null,
|
|
textResponse:
|
|
"There is no relevant information in this workspace to answer your query.",
|
|
};
|
|
}
|
|
|
|
// 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: message,
|
|
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,
|
|
chatMode = null,
|
|
}) {
|
|
if (chatMode === "query") return { rawHistory: [], chatHistory: [] };
|
|
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."
|
|
);
|
|
}
|
|
|
|
module.exports = {
|
|
recentChatHistory,
|
|
chatWithWorkspace,
|
|
chatPrompt,
|
|
grepCommand,
|
|
VALID_COMMANDS,
|
|
};
|