feat: claude integration (#552)

* feat: bare bones implementation of claude client (WIP)

* feat: client implementation of Claude (WIP)

* fix: add claude to store

* feat: bare bones implementation of claude client (WIP)

* switch eventsource

* Try new method of calling claude with anthropic sdk

* (WIP) Finish initial claude client implementation and api

* debugging update

* fix(ClaudeClient.js): fix prompt prefixes for HUMAN_PROMPT and AI_PROMPT
fix(ClaudeClient.js): refactor buildMessages logic for correct handling of messages
refactor(ClaudeClient.js): refactor buildPrompt method to buildMessages for use in BaseClient sendMessage method
refactor(ClaudeClient.js): refactor getCompletion method to sendCompletion for use in BaseClient sendMessage method
refactor(ClaudeClient.js): omit getMessageMapMethod method for future refactoring
refactor(ClaudeClient.js): remove unused sendMessage method to prefer BaseClient message
fix(askClaude.js): error in getIds method was causing a frontend crash, userMessage was not defined
fix(askClaude.js): import abortMessage function from utils module
feat(askClaude.js): add /abort route to handle message abort requests
feat(askClaude.js): create abortControllers map to store abort controllers
feat(askClaude.js): implement abortAsk function to handle message abort logic
feat(askClaude.js): add onStart callback to handle message start logic
feat(HoverButtons.jsx): add 'claude' as a supported endpoint for branching

* fix(ClaudeClient.js): update defaultPrefix and promptPrefix messages

includes 'Remember your instructions' as Claude is trained to recognize labels preceding colons as participants of a conversation

* Change name from claude to anthropic

* add settings to handleSubmit and models to endpoints

* Implement Claude settings

* use svg for anthropic icon

* Implement abort

* Implement reverse proxy

* remove png icons

* replace web browser plugin

* remove default prefix

* fix styling of claude icon

* fix console error from svg properties

* remove single quote requirement from eslintrc

* fix(AnthropicClient.js): fix labels for HUMAN_PROMPT and AI_PROMPT
feat(AnthropicClient.js): add support for custom userLabel and modelLabel options
feat(AnthropicClient.js): add user_id metadata to requestOptions in getCompletion method
feat(anthropic, AnthropicClient.js): add debug logging

* refactor(AnthropicClient.js): change promptSuffix variable declaration from let to const

* fix(EndpointOptionsDialog.jsx): remove unnecessary code that changes endpointName from 'anthropic' to 'Claude'
fix(utils/index.jsx): fix alternateName value for 'anthropic' from 'Claude' to 'Anthropic'

* fix(AnthropicIcon): fix sizing/rendering/name of anthropic icon

* fix(AnthropicClient.js): change maxContextTokens default value to 99999
fix(AnthropicClient.js): change maxResponseTokens default value to 1500
fix(AnthropicClient.js): remove unnecessary code for setting maxContextTokens and maxResponseTokens based on modelOptions
fix(AnthropicClient.js): change max_tokens_to_sample default value to 1500
fix(anthropic.js): pass endpointOption.token to AnthropicClient constructor

* Update .env.example

* fix(AnthropicClient.js): remove exceeding message when it puts us over the token limit
fix(AnthropicClient.js): handle case when the first message exceeds the token limit
fix(AnthropicClient.js): throw error when prompt is too long
fix(AnthropicClient.js): adjust max tokens calculation to use maxOutputTokens
fix(anthropic.js): remove console.log statement in ask route

* feat(server/index): increase incoming json payload allowed size

---------

Co-authored-by: Danny Avila <messagedaniel@protonmail.com>
This commit is contained in:
Dan Orlando 2023-07-13 18:35:15 -07:00 committed by GitHub
parent 981d009508
commit 9e931229e2
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
36 changed files with 1455 additions and 30 deletions

View file

@ -0,0 +1,319 @@
const Keyv = require('keyv');
// const { Agent, ProxyAgent } = require('undici');
const BaseClient = require('./BaseClient');
const {
encoding_for_model: encodingForModel,
get_encoding: getEncoding
} = require('@dqbd/tiktoken');
const Anthropic = require('@anthropic-ai/sdk');
const HUMAN_PROMPT = '\n\nHuman:';
const AI_PROMPT = '\n\nAssistant:';
const tokenizersCache = {};
class AnthropicClient extends BaseClient {
constructor(apiKey, options = {}, cacheOptions = {}) {
super(apiKey, options, cacheOptions)
cacheOptions.namespace = cacheOptions.namespace || 'anthropic';
this.conversationsCache = new Keyv(cacheOptions);
this.apiKey = apiKey || process.env.ANTHROPIC_API_KEY;
this.sender = 'Anthropic';
this.userLabel = HUMAN_PROMPT;
this.assistantLabel = AI_PROMPT;
this.setOptions(options);
}
setOptions(options) {
if (this.options && !this.options.replaceOptions) {
// nested options aren't spread properly, so we need to do this manually
this.options.modelOptions = {
...this.options.modelOptions,
...options.modelOptions
};
delete options.modelOptions;
// now we can merge options
this.options = {
...this.options,
...options,
};
} else {
this.options = options;
}
const modelOptions = this.options.modelOptions || {};
this.modelOptions = {
...modelOptions,
// set some good defaults (check for undefined in some cases because they may be 0)
model: modelOptions.model || 'claude-1',
temperature: typeof modelOptions.temperature === 'undefined' ? 0.7 : modelOptions.temperature, // 0 - 1, 0.7 is recommended
topP: typeof modelOptions.topP === 'undefined' ? 0.7 : modelOptions.topP, // 0 - 1, default: 0.7
topK: typeof modelOptions.topK === 'undefined' ? 40 : modelOptions.topK, // 1-40, default: 40
stop: modelOptions.stop // no stop method for now
};
this.maxContextTokens = this.options.maxContextTokens || 99999;
this.maxResponseTokens = this.modelOptions.maxOutputTokens || 1500;
this.maxPromptTokens =
this.options.maxPromptTokens || this.maxContextTokens - this.maxResponseTokens;
if (this.maxPromptTokens + this.maxResponseTokens > this.maxContextTokens) {
throw new Error(
`maxPromptTokens + maxOutputTokens (${this.maxPromptTokens} + ${this.maxResponseTokens} = ${
this.maxPromptTokens + this.maxResponseTokens
}) must be less than or equal to maxContextTokens (${this.maxContextTokens})`
);
}
this.startToken = '||>';
this.endToken = '';
this.gptEncoder = this.constructor.getTokenizer('cl100k_base');
if (!this.modelOptions.stop) {
const stopTokens = [this.startToken];
if (this.endToken && this.endToken !== this.startToken) {
stopTokens.push(this.endToken);
}
stopTokens.push(`${this.userLabel}`);
stopTokens.push('<|diff_marker|>');
this.modelOptions.stop = stopTokens;
}
return this;
}
getClient() {
if(this.options.reverseProxyUrl) {
return new Anthropic({
apiKey: this.apiKey,
baseURL: this.options.reverseProxyUrl
});
}
else {
return new Anthropic({
apiKey: this.apiKey,
});
}
};
async buildMessages(messages, parentMessageId) {
const orderedMessages = this.constructor.getMessagesForConversation(messages, parentMessageId);
if (this.options.debug) {
console.debug('AnthropicClient: orderedMessages', orderedMessages, parentMessageId);
}
const formattedMessages = orderedMessages.map((message) => ({
author: message.isCreatedByUser ? this.userLabel : this.assistantLabel,
content: message?.content ?? message.text
}));
let identityPrefix = '';
if (this.options.userLabel) {
identityPrefix = `\nHuman's name: ${this.options.userLabel}`;
}
if (this.options.modelLabel) {
identityPrefix = `${identityPrefix}\nYou are ${this.options.modelLabel}`;
}
let promptPrefix = (this.options.promptPrefix || '').trim();
if (promptPrefix) {
// If the prompt prefix doesn't end with the end token, add it.
if (!promptPrefix.endsWith(`${this.endToken}`)) {
promptPrefix = `${promptPrefix.trim()}${this.endToken}\n\n`;
}
promptPrefix = `\nContext:\n${promptPrefix}`;
}
if (identityPrefix) {
promptPrefix = `${identityPrefix}${promptPrefix}`;
}
const promptSuffix = `${promptPrefix}${this.assistantLabel}\n`; // Prompt AI to respond.
let currentTokenCount = this.getTokenCount(promptSuffix);
let promptBody = '';
const maxTokenCount = this.maxPromptTokens;
const context = [];
// Iterate backwards through the messages, adding them to the prompt until we reach the max token count.
// Do this within a recursive async function so that it doesn't block the event loop for too long.
// Also, remove the next message when the message that puts us over the token limit is created by the user.
// Otherwise, remove only the exceeding message. This is due to Anthropic's strict payload rule to start with "Human:".
const nextMessage = {
remove: false,
tokenCount: 0,
messageString: '',
};
const buildPromptBody = async () => {
if (currentTokenCount < maxTokenCount && formattedMessages.length > 0) {
const message = formattedMessages.pop();
const isCreatedByUser = message.author === this.userLabel;
const messageString = `${message.author}\n${message.content}${this.endToken}\n`;
let newPromptBody = `${messageString}${promptBody}`;
context.unshift(message);
const tokenCountForMessage = this.getTokenCount(messageString);
const newTokenCount = currentTokenCount + tokenCountForMessage;
if (!isCreatedByUser) {
nextMessage.messageString = messageString;
nextMessage.tokenCount = tokenCountForMessage;
}
if (newTokenCount > maxTokenCount) {
if (!promptBody) {
// This is the first message, so we can't add it. Just throw an error.
throw new Error(`Prompt is too long. Max token count is ${maxTokenCount}, but prompt is ${newTokenCount} tokens long.`);
}
// Otherwise, ths message would put us over the token limit, so don't add it.
// if created by user, remove next message, otherwise remove only this message
if (isCreatedByUser) {
nextMessage.remove = true;
}
return false;
}
promptBody = newPromptBody;
currentTokenCount = newTokenCount;
// wait for next tick to avoid blocking the event loop
await new Promise(resolve => setImmediate(resolve));
return buildPromptBody();
}
return true;
};
await buildPromptBody();
if (nextMessage.remove) {
promptBody = promptBody.replace(nextMessage.messageString, '');
currentTokenCount -= nextMessage.tokenCount;
context.shift();
}
const prompt = `${promptBody}${promptSuffix}`;
// Add 2 tokens for metadata after all messages have been counted.
currentTokenCount += 2;
// Use up to `this.maxContextTokens` tokens (prompt + response), but try to leave `this.maxTokens` tokens for the response.
this.modelOptions.maxOutputTokens = Math.min(this.maxContextTokens - currentTokenCount, this.maxResponseTokens);
return { prompt, context };
}
getCompletion() {
console.log('AnthropicClient doesn\'t use getCompletion (all handled in sendCompletion)');
}
// TODO: implement abortController usage
async sendCompletion(payload, { onProgress, abortController }) {
if (!abortController) {
abortController = new AbortController();
}
const { signal } = abortController;
const modelOptions = { ...this.modelOptions };
if (typeof onProgress === 'function') {
modelOptions.stream = true;
}
const { debug } = this.options;
if (debug) {
console.debug();
console.debug(modelOptions);
console.debug();
}
const client = this.getClient();
const metadata = {
user_id: this.user,
};
let text = '';
const requestOptions = {
prompt: payload,
model: this.modelOptions.model,
stream: this.modelOptions.stream || true,
max_tokens_to_sample: this.modelOptions.maxOutputTokens || 1500,
metadata,
...modelOptions
};
if (this.options.debug) {
console.log('AnthropicClient: requestOptions');
console.dir(requestOptions, { depth: null });
}
const response = await client.completions.create(requestOptions);
signal.addEventListener('abort', () => {
if (this.options.debug) {
console.log('AnthropicClient: message aborted!');
}
response.controller.abort();
});
for await (const completion of response) {
if (this.options.debug) {
// Uncomment to debug message stream
// console.debug(completion);
}
text += completion.completion;
onProgress(completion.completion);
}
signal.removeEventListener('abort', () => {
if (this.options.debug) {
console.log('AnthropicClient: message aborted!');
}
response.controller.abort();
});
return text.trim();
}
// I commented this out because I will need to refactor this for the BaseClient/all clients
// getMessageMapMethod() {
// return ((message) => ({
// author: message.isCreatedByUser ? this.userLabel : this.assistantLabel,
// content: message?.content ?? message.text
// })).bind(this);
// }
getSaveOptions() {
return {
...this.modelOptions
};
}
getBuildMessagesOptions() {
if (this.options.debug) {
console.log('AnthropicClient doesn\'t use getBuildMessagesOptions');
}
}
static getTokenizer(encoding, isModelName = false, extendSpecialTokens = {}) {
if (tokenizersCache[encoding]) {
return tokenizersCache[encoding];
}
let tokenizer;
if (isModelName) {
tokenizer = encodingForModel(encoding, extendSpecialTokens);
} else {
tokenizer = getEncoding(encoding, extendSpecialTokens);
}
tokenizersCache[encoding] = tokenizer;
return tokenizer;
}
getTokenCount(text) {
return this.gptEncoder.encode(text, 'all').length;
}
}
module.exports = AnthropicClient;

View file

@ -370,7 +370,6 @@ class BaseClient {
}
async sendMessage(message, opts = {}) {
console.log('BaseClient: sendMessage', message, opts);
const {
user,
conversationId,
@ -379,6 +378,7 @@ class BaseClient {
userMessage,
} = await this.handleStartMethods(message, opts);
this.user = user;
// It's not necessary to push to currentMessages
// depending on subclass implementation of handling messages
this.currentMessages.push(userMessage);

View file

@ -3,6 +3,7 @@ const OpenAIClient = require('./OpenAIClient');
const PluginsClient = require('./PluginsClient');
const GoogleClient = require('./GoogleClient');
const TextStream = require('./TextStream');
const AnthropicClient = require('./AnthropicClient');
const toolUtils = require('./tools/util');
module.exports = {
@ -11,5 +12,6 @@ module.exports = {
PluginsClient,
GoogleClient,
TextStream,
AnthropicClient,
...toolUtils
};

View file

@ -34,7 +34,7 @@
"name": "Browser",
"pluginKey": "browser",
"description": "Scrape and summarize webpage data",
"icon": "/assets/web-browser.png",
"icon": "/assets/web-browser.svg",
"authConfig": [
{
"authField": "OPENAI_API_KEY",

View file

@ -1,5 +1,5 @@
const conversationPreset = {
// endpoint: [azureOpenAI, openAI, bingAI, chatGPTBrowser]
// endpoint: [azureOpenAI, openAI, bingAI, anthropic, chatGPTBrowser]
endpoint: {
type: String,
default: null,
@ -151,7 +151,7 @@ const agentOptions = {
default: null
}
};
module.exports = {
conversationPreset,
agentOptions

View file

@ -20,7 +20,9 @@
},
"homepage": "https://github.com/danny-avila/LibreChat#readme",
"dependencies": {
"@anthropic-ai/sdk": "^0.5.0",
"@dqbd/tiktoken": "^1.0.2",
"@fortaine/fetch-event-source": "^3.0.6",
"@keyv/mongo": "^2.1.8",
"@waylaidwanderer/chatgpt-api": "^1.37.0",
"axios": "^1.3.4",

View file

@ -24,8 +24,8 @@ config.validate(); // Validate the config
const app = express();
app.use(errorController);
app.use(express.json());
app.use(express.urlencoded({ extended: true }));
app.use(express.json({ limit: '3mb' }));
app.use(express.urlencoded({ extended: true, limit: '3mb' }));
app.use(express.static(path.join(projectPath, 'dist')));
app.use(express.static(path.join(projectPath, 'public')));

View file

@ -0,0 +1,186 @@
const express = require('express');
const router = express.Router();
const crypto = require('crypto');
const { titleConvo, AnthropicClient } = require('../../../app');
const requireJwtAuth = require('../../../middleware/requireJwtAuth');
const { abortMessage } = require('../../../utils');
const { saveMessage, getConvoTitle, saveConvo, getConvo } = require('../../../models');
const { handleError, sendMessage, createOnProgress } = require('./handlers');
const abortControllers = new Map();
router.post('/abort', requireJwtAuth, async (req, res) => {
return await abortMessage(req, res, abortControllers);
});
router.post('/', requireJwtAuth, async (req, res) => {
const { endpoint, text, parentMessageId, conversationId: oldConversationId } = req.body;
if (text.length === 0) return handleError(res, { text: 'Prompt empty or too short' });
if (endpoint !== 'anthropic') return handleError(res, { text: 'Illegal request' });
const endpointOption = {
promptPrefix: req.body?.promptPrefix ?? null,
modelLabel: req.body?.modelLabel ?? null,
token: req.body?.token ?? null,
modelOptions: {
model: req.body?.model ?? 'claude-1',
temperature: req.body?.temperature ?? 0.7,
maxOutputTokens: req.body?.maxOutputTokens ?? 1024,
topP: req.body?.topP ?? 0.7,
topK: req.body?.topK ?? 40
}
};
const conversationId = oldConversationId || crypto.randomUUID();
return await ask({
text,
endpointOption,
conversationId,
parentMessageId,
req,
res
});
});
const ask = async ({ text, endpointOption, parentMessageId = null, conversationId, req, res }) => {
res.writeHead(200, {
Connection: 'keep-alive',
'Content-Type': 'text/event-stream',
'Cache-Control': 'no-cache, no-transform',
'Access-Control-Allow-Origin': '*',
'X-Accel-Buffering': 'no'
});
let userMessage;
let userMessageId;
let responseMessageId;
let lastSavedTimestamp = 0;
const { overrideParentMessageId = null } = req.body;
try {
const getIds = (data) => {
userMessage = data.userMessage;
userMessageId = data.userMessage.messageId;
responseMessageId = data.responseMessageId;
if (!conversationId) {
conversationId = data.conversationId;
}
};
const { onProgress: progressCallback, getPartialText } = createOnProgress({
onProgress: ({ text: partialText }) => {
const currentTimestamp = Date.now();
if (currentTimestamp - lastSavedTimestamp > 500) {
lastSavedTimestamp = currentTimestamp;
saveMessage({
messageId: responseMessageId,
sender: 'Anthropic',
conversationId,
parentMessageId: overrideParentMessageId || userMessageId,
text: partialText,
unfinished: true,
cancelled: false,
error: false
});
}
}
});
const abortController = new AbortController();
abortController.abortAsk = async function () {
this.abort();
const responseMessage = {
messageId: responseMessageId,
sender: 'Anthropic',
conversationId,
parentMessageId: overrideParentMessageId || userMessageId,
text: getPartialText(),
model: endpointOption.modelOptions.model,
unfinished: false,
cancelled: true,
error: false,
};
saveMessage(responseMessage);
return {
title: await getConvoTitle(req.user.id, conversationId),
final: true,
conversation: await getConvo(req.user.id, conversationId),
requestMessage: userMessage,
responseMessage: responseMessage
};
};
const onStart = (userMessage) => {
sendMessage(res, { message: userMessage, created: true });
abortControllers.set(userMessage.conversationId, { abortController, ...endpointOption });
}
const client = new AnthropicClient(endpointOption.token);
let response = await client.sendMessage(text, {
getIds,
debug: false,
user: req.user.id,
conversationId,
parentMessageId,
overrideParentMessageId,
...endpointOption,
onProgress: progressCallback.call(null, {
res,
text,
parentMessageId: overrideParentMessageId || userMessageId
}),
onStart,
abortController
});
if (overrideParentMessageId) {
response.parentMessageId = overrideParentMessageId;
}
await saveConvo(req.user.id, {
...endpointOption,
...endpointOption.modelOptions,
conversationId,
endpoint: 'anthropic'
});
await saveMessage(response);
sendMessage(res, {
title: await getConvoTitle(req.user.id, conversationId),
final: true,
conversation: await getConvo(req.user.id, conversationId),
requestMessage: userMessage,
responseMessage: response
});
res.end();
if (parentMessageId == '00000000-0000-0000-0000-000000000000') {
const title = await titleConvo({ text, response });
await saveConvo(req.user.id, {
conversationId,
title
});
}
} catch (error) {
console.error(error);
const errorMessage = {
messageId: responseMessageId,
sender: 'Anthropic',
conversationId,
parentMessageId,
unfinished: false,
cancelled: false,
error: true,
text: error.message
};
await saveMessage(errorMessage);
handleError(res, errorMessage);
}
};
module.exports = router;

View file

@ -7,6 +7,7 @@ const google = require('./google');
const askBingAI = require('./askBingAI');
const gptPlugins = require('./gptPlugins');
const askChatGPTBrowser = require('./askChatGPTBrowser');
const anthropic = require('./anthropic');
// router.use('/azureOpenAI', askAzureOpenAI);
router.use(['/azureOpenAI', '/openAI'], openAI);
@ -14,5 +15,6 @@ router.use('/google', google);
router.use('/bingAI', askBingAI);
router.use('/chatGPTBrowser', askChatGPTBrowser);
router.use('/gptPlugins', gptPlugins);
router.use('/anthropic', anthropic);
module.exports = router;

View file

@ -16,6 +16,12 @@ const getChatGPTBrowserModels = () => {
return models;
};
const getAnthropicModels = () => {
let models = ['claude-1', 'claude-1-100k', 'claude-instant-1', 'claude-instant-1-100k', 'claude-2'];
if (process.env.ANTHROPIC_MODELS) models = String(process.env.ANTHROPIC_MODELS).split(',');
return models;
};
const getPluginModels = () => {
let models = ['gpt-4', 'gpt-4-0613', 'gpt-3.5-turbo', 'gpt-3.5-turbo-16k', 'gpt-3.5-turbo-0613', 'gpt-3.5-turbo-0301'];
@ -55,7 +61,7 @@ router.get('/', async function (req, res) {
? { availableModels: getOpenAIModels(), userProvide: openAIApiKey === 'user_provided' }
: false;
const azureOpenAI = azureOpenAIApiKey
? { availableModels: getOpenAIModels({ azure: true}), userProvide: azureOpenAIApiKey === 'user_provided' }
? { availableModels: getOpenAIModels({ azure: true }), userProvide: azureOpenAIApiKey === 'user_provided' }
: false;
const gptPlugins = openAIApiKey || azureOpenAIApiKey
? { availableModels: getPluginModels(), availableTools, availableAgents: ['classic', 'functions'], userProvide: userProvidedOpenAI }
@ -69,8 +75,14 @@ router.get('/', async function (req, res) {
availableModels: getChatGPTBrowserModels()
}
: false;
const anthropic = process.env.ANTHROPIC_API_KEY
? {
userProvide: process.env.ANTHROPIC_API_KEY == 'user_provided',
availableModels: getAnthropicModels()
}
: false;
res.send(JSON.stringify({ azureOpenAI, openAI, google, bingAI, chatGPTBrowser, gptPlugins }));
res.send(JSON.stringify({ azureOpenAI, openAI, google, bingAI, chatGPTBrowser, gptPlugins, anthropic }));
});
module.exports = { router, getOpenAIModels, getChatGPTBrowserModels };