rixty
/
SillyTavern
mirror of https://github.com/SillyTavern/SillyTavern.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Move prompt converters to a separate module. Camelcase local variables and add missing JSDocs.

This commit is contained in:
Cohee 2023-12-14 16:00:17 +02:00
parent 348253fd98
commit 2a5340232d
3 changed files with 176 additions and 154 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

225
src/endpoints/backends/chat-completions.js
View File

@ -4,6 +4,7 @@ const fetch = require('node-fetch').default;
const { jsonParser } = require('../../express-common');
const { CHAT_COMPLETION_SOURCES, PALM_SAFETY } = require('../../constants');
const { forwardFetchResponse, getConfigValue, tryParse, uuidv4 } = require('../../util');
const { convertClaudePrompt, convertTextCompletionPrompt } = require('../prompt-converters');
const { readSecret, SECRET_KEYS } = require('../secrets');
const { getTokenizerModel, getSentencepiceTokenizer, getTiktokenTokenizer, sentencepieceTokenizers, TEXT_COMPLETION_MODELS } = require('../tokenizers');
@ -12,89 +13,15 @@ const API_OPENAI = 'https://api.openai.com/v1';
const API_CLAUDE = 'https://api.anthropic.com/v1';
/**
* Convert a prompt from the ChatML objects to the format used by Claude.
* @param {object[]} messages Array of messages
* @param {boolean} addHumanPrefix Add Human prefix
* @param {boolean} addAssistantPostfix Add Assistant postfix
* @param {boolean} withSystemPrompt Build system prompt before "\n\nHuman: "
* @returns {string} Prompt for Claude
* @copyright Prompt Conversion script taken from RisuAI by kwaroran (GPLv3).
*/
function convertClaudePrompt(messages, addHumanPrefix, addAssistantPostfix, withSystemPrompt) {
// Claude doesn't support message names, so we'll just add them to the message content.
for (const message of messages) {
if (message.name && message.role !== 'system') {
message.content = message.name + ': ' + message.content;
delete message.name;
}
}
let systemPrompt = '';
if (withSystemPrompt) {
let lastSystemIdx = -1;
for (let i = 0; i < messages.length - 1; i++) {
const message = messages[i];
if (message.role === 'system' && !message.name) {
systemPrompt += message.content + '\n\n';
} else {
lastSystemIdx = i - 1;
break;
}
}
if (lastSystemIdx >= 0) {
messages.splice(0, lastSystemIdx + 1);
}
}
let requestPrompt = messages.map((v) => {
let prefix = '';
switch (v.role) {
case 'assistant':
prefix = '\n\nAssistant: ';
break;
case 'user':
prefix = '\n\nHuman: ';
break;
case 'system':
// According to the Claude docs, H: and A: should be used for example conversations.
if (v.name === 'example_assistant') {
prefix = '\n\nA: ';
} else if (v.name === 'example_user') {
prefix = '\n\nH: ';
} else {
prefix = '\n\n';
}
break;
}
return prefix + v.content;
}).join('');
if (addHumanPrefix) {
requestPrompt = '\n\nHuman: ' + requestPrompt;
}
if (addAssistantPostfix) {
requestPrompt = requestPrompt + '\n\nAssistant: ';
}
if (withSystemPrompt) {
requestPrompt = systemPrompt + requestPrompt;
}
return requestPrompt;
}
/**
* @param {express.Request} request
* @param {express.Response} response
* Sends a request to Claude API.
* @param {express.Request} request Express request
* @param {express.Response} response Express response
*/
async function sendClaudeRequest(request, response) {
const apiUrl = new URL(request.body.reverse_proxy || API_CLAUDE).toString();
const apiKey = request.body.reverse_proxy ? request.body.proxy_password : readSecret(SECRET_KEYS.CLAUDE);
const api_url = new URL(request.body.reverse_proxy || API_CLAUDE).toString();
const api_key_claude = request.body.reverse_proxy ? request.body.proxy_password : readSecret(SECRET_KEYS.CLAUDE);
if (!api_key_claude) {
if (!apiKey) {
console.log('Claude API key is missing.');
return response.status(400).send({ error: true });
}
@ -121,7 +48,7 @@ async function sendClaudeRequest(request, response) {
stop_sequences.push(...request.body.stop);
}
const generateResponse = await fetch(api_url + '/complete', {
const generateResponse = await fetch(apiUrl + '/complete', {
method: 'POST',
signal: controller.signal,
body: JSON.stringify({
@ -137,7 +64,7 @@ async function sendClaudeRequest(request, response) {
headers: {
'Content-Type': 'application/json',
'anthropic-version': '2023-06-01',
'x-api-key': api_key_claude,
'x-api-key': apiKey,
},
timeout: 0,
});
@ -167,37 +94,21 @@ async function sendClaudeRequest(request, response) {
}
}
function convertChatMLPrompt(messages) {
if (typeof messages === 'string') {
return messages;
}
const messageStrings = [];
messages.forEach(m => {
if (m.role === 'system' && m.name === undefined) {
messageStrings.push('System: ' + m.content);
}
else if (m.role === 'system' && m.name !== undefined) {
messageStrings.push(m.name + ': ' + m.content);
}
else {
messageStrings.push(m.role + ': ' + m.content);
}
});
return messageStrings.join('\n') + '\nassistant:';
}
/**
* Sends a request to Scale Spellbook API.
* @param {import("express").Request} request Express request
* @param {import("express").Response} response Express response
*/
async function sendScaleRequest(request, response) {
const apiUrl = new URL(request.body.api_url_scale).toString();
const apiKey = readSecret(SECRET_KEYS.SCALE);
const api_url = new URL(request.body.api_url_scale).toString();
const api_key_scale = readSecret(SECRET_KEYS.SCALE);
if (!api_key_scale) {
if (!apiKey) {
console.log('Scale API key is missing.');
return response.status(400).send({ error: true });
}
const requestPrompt = convertChatMLPrompt(request.body.messages);
const requestPrompt = convertTextCompletionPrompt(request.body.messages);
console.log('Scale request:', requestPrompt);
try {
@ -207,12 +118,12 @@ async function sendScaleRequest(request, response) {
controller.abort();
});
const generateResponse = await fetch(api_url, {
const generateResponse = await fetch(apiUrl, {
method: 'POST',
body: JSON.stringify({ input: { input: requestPrompt } }),
headers: {
'Content-Type': 'application/json',
'Authorization': `Basic ${api_key_scale}`,
'Authorization': `Basic ${apiKey}`,
},
timeout: 0,
});
@ -236,8 +147,9 @@ async function sendScaleRequest(request, response) {
}
/**
* @param {express.Request} request
* @param {express.Response} response
* Sends a request to Google AI API.
* @param {express.Request} request Express request
* @param {express.Response} response Express response
*/
async function sendPalmRequest(request, response) {
const api_key_palm = readSecret(SECRET_KEYS.PALM);
@ -312,6 +224,11 @@ async function sendPalmRequest(request, response) {
}
}
/**
* Sends a request to Google AI API.
* @param {express.Request} request Express request
* @param {express.Response} response Express response
*/
async function sendAI21Request(request, response) {
if (!request.body) return response.sendStatus(400);
const controller = new AbortController();
@ -533,24 +450,24 @@ router.post('/bias', jsonParser, async function (request, response) {
});
router.post('/generate', jsonParser, function (request, response_generate_openai) {
if (!request.body) return response_generate_openai.status(400).send({ error: true });
router.post('/generate', jsonParser, function (request, response) {
if (!request.body) return response.status(400).send({ error: true });
switch (request.body.chat_completion_source) {
case CHAT_COMPLETION_SOURCES.CLAUDE: return sendClaudeRequest(request, response_generate_openai);
case CHAT_COMPLETION_SOURCES.SCALE: return sendScaleRequest(request, response_generate_openai);
case CHAT_COMPLETION_SOURCES.AI21: return sendAI21Request(request, response_generate_openai);
case CHAT_COMPLETION_SOURCES.PALM: return sendPalmRequest(request, response_generate_openai);
case CHAT_COMPLETION_SOURCES.CLAUDE: return sendClaudeRequest(request, response);
case CHAT_COMPLETION_SOURCES.SCALE: return sendScaleRequest(request, response);
case CHAT_COMPLETION_SOURCES.AI21: return sendAI21Request(request, response);
case CHAT_COMPLETION_SOURCES.PALM: return sendPalmRequest(request, response);
}
let api_url;
let api_key_openai;
let apiUrl;
let apiKey;
let headers;
let bodyParams;
if (request.body.chat_completion_source !== CHAT_COMPLETION_SOURCES.OPENROUTER) {
api_url = new URL(request.body.reverse_proxy || API_OPENAI).toString();
api_key_openai = request.body.reverse_proxy ? request.body.proxy_password : readSecret(SECRET_KEYS.OPENAI);
apiUrl = new URL(request.body.reverse_proxy || API_OPENAI).toString();
apiKey = request.body.reverse_proxy ? request.body.proxy_password : readSecret(SECRET_KEYS.OPENAI);
headers = {};
bodyParams = {};
@ -558,8 +475,8 @@ router.post('/generate', jsonParser, function (request, response_generate_openai
bodyParams['user'] = uuidv4();
}
} else {
api_url = 'https://openrouter.ai/api/v1';
api_key_openai = readSecret(SECRET_KEYS.OPENROUTER);
apiUrl = 'https://openrouter.ai/api/v1';
apiKey = readSecret(SECRET_KEYS.OPENROUTER);
// OpenRouter needs to pass the referer: https://openrouter.ai/docs
headers = { 'HTTP-Referer': request.headers.referer };
bodyParams = { 'transforms': ['middle-out'] };
@ -569,9 +486,9 @@ router.post('/generate', jsonParser, function (request, response_generate_openai
}
}
if (!api_key_openai && !request.body.reverse_proxy) {
if (!apiKey && !request.body.reverse_proxy) {
console.log('OpenAI API key is missing.');
return response_generate_openai.status(400).send({ error: true });
return response.status(400).send({ error: true });
}
// Add custom stop sequences
@ -580,10 +497,10 @@ router.post('/generate', jsonParser, function (request, response_generate_openai
}
const isTextCompletion = Boolean(request.body.model && TEXT_COMPLETION_MODELS.includes(request.body.model)) || typeof request.body.messages === 'string';
const textPrompt = isTextCompletion ? convertChatMLPrompt(request.body.messages) : '';
const textPrompt = isTextCompletion ? convertTextCompletionPrompt(request.body.messages) : '';
const endpointUrl = isTextCompletion && request.body.chat_completion_source !== CHAT_COMPLETION_SOURCES.OPENROUTER ?
`${api_url}/completions` :
`${api_url}/chat/completions`;
`${apiUrl}/completions` :
`${apiUrl}/chat/completions`;
const controller = new AbortController();
request.socket.removeAllListeners('close');
@ -596,7 +513,7 @@ router.post('/generate', jsonParser, function (request, response_generate_openai
method: 'post',
headers: {
'Content-Type': 'application/json',
'Authorization': 'Bearer ' + api_key_openai,
'Authorization': 'Bearer ' + apiKey,
...headers,
},
body: JSON.stringify({
@ -621,52 +538,55 @@ router.post('/generate', jsonParser, function (request, response_generate_openai
console.log(JSON.parse(String(config.body)));
makeRequest(config, response_generate_openai, request);
makeRequest(config, response, request);
/**
*
* @param {*} config
* @param {express.Response} response_generate_openai
* @param {express.Request} request
* @param {Number} retries
* @param {Number} timeout
* Makes a fetch request to the OpenAI API endpoint.
* @param {import('node-fetch').RequestInit} config Fetch config
* @param {express.Response} response Express response
* @param {express.Request} request Express request
* @param {Number} retries Number of retries left
* @param {Number} timeout Request timeout in ms
*/
async function makeRequest(config, response_generate_openai, request, retries = 5, timeout = 5000) {
async function makeRequest(config, response, request, retries = 5, timeout = 5000) {
try {
const fetchResponse = await fetch(endpointUrl, config);
if (request.body.stream) {
console.log('Streaming request in progress');
forwardFetchResponse(fetchResponse, response_generate_openai);
forwardFetchResponse(fetchResponse, response);
return;
}
if (fetchResponse.ok) {
let json = await fetchResponse.json();
response_generate_openai.send(json);
response.send(json);
console.log(json);
console.log(json?.choices[0]?.message);
} else if (fetchResponse.status === 429 && retries > 0) {
console.log(`Out of quota, retrying in ${Math.round(timeout / 1000)}s`);
setTimeout(() => {
timeout *= 2;
makeRequest(config, response_generate_openai, request, retries - 1, timeout);
makeRequest(config, response, request, retries - 1, timeout);
}, timeout);
} else {
await handleErrorResponse(fetchResponse);
}
} catch (error) {
console.log('Generation failed', error);
if (!response_generate_openai.headersSent) {
response_generate_openai.send({ error: true });
if (!response.headersSent) {
response.send({ error: true });
} else {
response_generate_openai.end();
response.end();
}
}
}
async function handleErrorResponse(response) {
const responseText = await response.text();
/**
* @param {import("node-fetch").Response} errorResponse
*/
async function handleErrorResponse(errorResponse) {
const responseText = await errorResponse.text();
const errorData = tryParse(responseText);
const statusMessages = {
@ -680,21 +600,20 @@ router.post('/generate', jsonParser, function (request, response_generate_openai
502: 'Bad gateway',
};
const message = errorData?.error?.message || statusMessages[response.status] || 'Unknown error occurred';
const quota_error = response.status === 429 && errorData?.error?.type === 'insufficient_quota';
const message = errorData?.error?.message || statusMessages[errorResponse.status] || 'Unknown error occurred';
const quota_error = errorResponse.status === 429 && errorData?.error?.type === 'insufficient_quota';
console.log(message);
if (!response_generate_openai.headersSent) {
response_generate_openai.send({ error: { message }, quota_error: quota_error });
} else if (!response_generate_openai.writableEnded) {
response_generate_openai.write(response);
if (!response.headersSent) {
response.send({ error: { message }, quota_error: quota_error });
} else if (!response.writableEnded) {
response.write(errorResponse);
} else {
response_generate_openai.end();
response.end();
}
}
});
module.exports = {
router,
convertClaudePrompt,
};

103
src/endpoints/prompt-converters.js Normal file
View File

@ -0,0 +1,103 @@
/**
* Convert a prompt from the ChatML objects to the format used by Claude.
* @param {object[]} messages Array of messages
* @param {boolean} addHumanPrefix Add Human prefix
* @param {boolean} addAssistantPostfix Add Assistant postfix
* @param {boolean} withSystemPrompt Build system prompt before "\n\nHuman: "
* @returns {string} Prompt for Claude
* @copyright Prompt Conversion script taken from RisuAI by kwaroran (GPLv3).
*/
function convertClaudePrompt(messages, addHumanPrefix, addAssistantPostfix, withSystemPrompt) {
// Claude doesn't support message names, so we'll just add them to the message content.
for (const message of messages) {
if (message.name && message.role !== 'system') {
message.content = message.name + ': ' + message.content;
delete message.name;
}
}
let systemPrompt = '';
if (withSystemPrompt) {
let lastSystemIdx = -1;
for (let i = 0; i < messages.length - 1; i++) {
const message = messages[i];
if (message.role === 'system' && !message.name) {
systemPrompt += message.content + '\n\n';
} else {
lastSystemIdx = i - 1;
break;
}
}
if (lastSystemIdx >= 0) {
messages.splice(0, lastSystemIdx + 1);
}
}
let requestPrompt = messages.map((v) => {
let prefix = '';
switch (v.role) {
case 'assistant':
prefix = '\n\nAssistant: ';
break;
case 'user':
prefix = '\n\nHuman: ';
break;
case 'system':
// According to the Claude docs, H: and A: should be used for example conversations.
if (v.name === 'example_assistant') {
prefix = '\n\nA: ';
} else if (v.name === 'example_user') {
prefix = '\n\nH: ';
} else {
prefix = '\n\n';
}
break;
}
return prefix + v.content;
}).join('');
if (addHumanPrefix) {
requestPrompt = '\n\nHuman: ' + requestPrompt;
}
if (addAssistantPostfix) {
requestPrompt = requestPrompt + '\n\nAssistant: ';
}
if (withSystemPrompt) {
requestPrompt = systemPrompt + requestPrompt;
}
return requestPrompt;
}
/**
* Convert a prompt from the ChatML objects to the format used by Text Completion API.
* @param {object[]} messages Array of messages
* @returns {string} Prompt for Text Completion API
*/
function convertTextCompletionPrompt(messages) {
if (typeof messages === 'string') {
return messages;
}
const messageStrings = [];
messages.forEach(m => {
if (m.role === 'system' && m.name === undefined) {
messageStrings.push('System: ' + m.content);
}
else if (m.role === 'system' && m.name !== undefined) {
messageStrings.push(m.name + ': ' + m.content);
}
else {
messageStrings.push(m.role + ': ' + m.content);
}
});
return messageStrings.join('\n') + '\nassistant:';
}
module.exports = {
convertClaudePrompt,
convertTextCompletionPrompt,
};

2
src/endpoints/tokenizers.js
View File

@ -4,7 +4,7 @@ const express = require('express');
const { SentencePieceProcessor } = require('@agnai/sentencepiece-js');
const tiktoken = require('@dqbd/tiktoken');
const { Tokenizer } = require('@agnai/web-tokenizers');
const { convertClaudePrompt } = require('./backends/chat-completions');
const { convertClaudePrompt } = require('./prompt-converters');
const { readSecret, SECRET_KEYS } = require('./secrets');
const { TEXTGEN_TYPES } = require('../constants');
const { jsonParser } = require('../express-common');