rixty/SillyTavern
1
0
Fork 0
mirror of https://github.com/SillyTavern/SillyTavern.git synced 2025-06-05 21:59:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'upstream/staging' into improve-bulk-edit-and-fixes

Browse Source
This commit is contained in:
Wolfsblvt
2024-03-30 19:03:03 +01:00
parent b747bdf89b 9f201a78de
commit 1ff40f0af4
70 changed files with 1341 additions and 361 deletions
Download Patch File Download Diff File Expand all files Collapse all files

132
public/script.js
View File

@@ -2867,20 +2867,22 @@ class StreamingProcessor {
* @param {string} prompt Prompt to generate a message from
* @param {string} api API to use. Main API is used if not specified.
* @param {boolean} instructOverride true to override instruct mode, false to use the default value
* @param {boolean} quietToLoud true to generate a message in system mode, false to generate a message in character mode
* @returns {Promise<string>} Generated message
*/
export async function generateRaw(prompt, api, instructOverride) {
export async function generateRaw(prompt, api, instructOverride, quietToLoud) {
if (!api) {
api = main_api;
}
const abortController = new AbortController();
const isInstruct = power_user.instruct.enabled && main_api !== 'openai' && main_api !== 'novel' && !instructOverride;
const isQuiet = true;
prompt = substituteParams(prompt);
prompt = api == 'novel' ? adjustNovelInstructionPrompt(prompt) : prompt;
prompt = isInstruct ? formatInstructModeChat(name1, prompt, false, true, '', name1, name2, false) : prompt;
prompt = isInstruct ? (prompt + formatInstructModePrompt(name2, false, '', name1, name2)) : (prompt + '\n');
prompt = isInstruct ? (prompt + formatInstructModePrompt(name2, false, '', name1, name2, isQuiet, quietToLoud)) : (prompt + '\n');
let generateData = {};
@@ -3150,10 +3152,6 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
mesExamples = '';
}
const mesExamplesRaw = mesExamples;
if (mesExamples && isInstruct) {
mesExamples = formatInstructModeExamples(mesExamples, name1, name2);
}
/**
* Adds a block heading to the examples string.
* @param {string} examplesStr
@@ -3161,13 +3159,17 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
*/
function addBlockHeading(examplesStr) {
const exampleSeparator = power_user.context.example_separator ? `${substituteParams(power_user.context.example_separator)}\n` : '';
const blockHeading = main_api === 'openai' ? '<START>\n' : exampleSeparator;
const blockHeading = main_api === 'openai' ? '<START>\n' : (exampleSeparator || (isInstruct ? '<START>\n' : ''));
return examplesStr.split(/<START>/gi).slice(1).map(block => `${blockHeading}${block.trim()}\n`);
}
let mesExamplesArray = addBlockHeading(mesExamples);
let mesExamplesRawArray = addBlockHeading(mesExamplesRaw);
if (mesExamplesArray && isInstruct) {
mesExamplesArray = formatInstructModeExamples(mesExamplesArray, name1, name2);
}
// First message in fresh 1-on-1 chat reacts to user/character settings changes
if (chat.length) {
chat[0].mes = substituteParams(chat[0].mes);
@@ -3281,6 +3283,8 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
let chat2 = [];
let continue_mag = '';
const userMessageIndices = [];
for (let i = coreChat.length - 1, j = 0; i >= 0; i--, j++) {
if (main_api == 'openai') {
chat2[i] = coreChat[j].mes;
@@ -3308,6 +3312,22 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
chat2[i] = chat2[i].slice(0, chat2[i].lastIndexOf(coreChat[j].mes) + coreChat[j].mes.length);
continue_mag = coreChat[j].mes;
}
if (coreChat[j].is_user) {
userMessageIndices.push(i);
}
}
let addUserAlignment = isInstruct && power_user.instruct.user_alignment_message;
let userAlignmentMessage = '';
if (addUserAlignment) {
const alignmentMessage = {
name: name1,
mes: power_user.instruct.user_alignment_message,
is_user: true,
};
userAlignmentMessage = formatMessageHistoryItem(alignmentMessage, isInstruct, false);
}
// Add persona description to prompt
@@ -3366,6 +3386,7 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
allAnchors,
quiet_prompt,
cyclePrompt,
userAlignmentMessage,
].join('').replace(/\r/gm, '');
return getTokenCount(encodeString, power_user.token_padding);
}
@@ -3382,18 +3403,24 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
}
// Collect enough messages to fill the context
let arrMes = [];
let arrMes = new Array(chat2.length);
let tokenCount = getMessagesTokenCount();
for (let item of chat2) {
// not needed for OAI prompting
if (main_api == 'openai') {
break;
let lastAddedIndex = -1;
// Pre-allocate all injections first.
// If it doesn't fit - user shot himself in the foot
for (const index of injectedIndices) {
const item = chat2[index];
if (typeof item !== 'string') {
continue;
}
tokenCount += getTokenCount(item.replace(/\r/gm, ''));
chatString = item + chatString;
if (tokenCount < this_max_context) {
arrMes[arrMes.length] = item;
arrMes[index] = item;
lastAddedIndex = Math.max(lastAddedIndex, index);
} else {
break;
}
@@ -3402,8 +3429,62 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
await delay(1);
}
for (let i = 0; i < chat2.length; i++) {
// not needed for OAI prompting
if (main_api == 'openai') {
break;
}
// Skip already injected messages
if (arrMes[i] !== undefined) {
continue;
}
const item = chat2[i];
if (typeof item !== 'string') {
continue;
}
tokenCount += getTokenCount(item.replace(/\r/gm, ''));
chatString = item + chatString;
if (tokenCount < this_max_context) {
arrMes[i] = item;
lastAddedIndex = Math.max(lastAddedIndex, i);
} else {
break;
}
// Prevent UI thread lock on tokenization
await delay(1);
}
// Add user alignment message if last message is not a user message
const stoppedAtUser = userMessageIndices.includes(lastAddedIndex);
if (addUserAlignment && !stoppedAtUser) {
tokenCount += getTokenCount(userAlignmentMessage.replace(/\r/gm, ''));
chatString = userAlignmentMessage + chatString;
arrMes.push(userAlignmentMessage);
injectedIndices.push(arrMes.length - 1);
}
// Unsparse the array. Adjust injected indices
const newArrMes = [];
const newInjectedIndices = [];
for (let i = 0; i < arrMes.length; i++) {
if (arrMes[i] !== undefined) {
newArrMes.push(arrMes[i]);
if (injectedIndices.includes(i)) {
newInjectedIndices.push(newArrMes.length - 1);
}
}
}
arrMes = newArrMes;
injectedIndices = newInjectedIndices;
if (main_api !== 'openai') {
setInContextMessages(arrMes.length, type);
setInContextMessages(arrMes.length - injectedIndices.length, type);
}
// Estimate how many unpinned example messages fit in the context
@@ -3446,15 +3527,19 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
console.debug('generating prompt');
chatString = '';
arrMes = arrMes.reverse();
arrMes.forEach(function (item, i, arr) {// For added anchors and others
arrMes.forEach(function (item, i, arr) {
// OAI doesn't need all of this
if (main_api === 'openai') {
return;
}
// Cohee: I'm not even sure what this is for anymore
// Cohee: This removes a newline from the end of the last message in the context
// Last prompt line will add a newline if it's not a continuation
// In instruct mode it only removes it if wrap is enabled and it's not a quiet generation
if (i === arrMes.length - 1 && type !== 'continue') {
item = item.replace(/\n?$/, '');
if (!isInstruct || (power_user.instruct.wrap && type !== 'quiet')) {
item = item.replace(/\n?$/, '');
}
}
mesSend[mesSend.length] = { message: item, extensionPrompts: [] };
@@ -3493,7 +3578,7 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
//TODO: respect output_sequence vs last_output_sequence settings
//TODO: decide how to prompt this to clarify who is talking 'Narrator', 'System', etc.
if (isInstruct) {
lastMesString += '\n' + quietAppend; // + power_user.instruct.output_sequence + '\n';
lastMesString += quietAppend; // + power_user.instruct.output_sequence + '\n';
} else {
lastMesString += quietAppend;
}
@@ -3514,7 +3599,8 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
// Get instruct mode line
if (isInstruct && !isContinue) {
const name = (quiet_prompt && !quietToLoud) ? (quietName ?? 'System') : (isImpersonate ? name1 : name2);
lastMesString += formatInstructModePrompt(name, isImpersonate, promptBias, name1, name2);
const isQuiet = quiet_prompt && type == 'quiet';
lastMesString += formatInstructModePrompt(name, isImpersonate, promptBias, name1, name2, isQuiet, quietToLoud);
}
// Get non-instruct impersonation line
@@ -3681,7 +3767,7 @@ async function Generate(type, { automatic_trigger, force_name2, quiet_prompt, qu
};
finalMesSend.forEach((item, i) => {
item.injected = Array.isArray(injectedIndices) && injectedIndices.includes(i);
item.injected = injectedIndices.includes(finalMesSend.length - i - 1);
});
let data = {
@@ -4049,10 +4135,6 @@ function doChatInject(messages, isContinue) {
}
}
for (let i = 0; i < injectedIndices.length; i++) {
injectedIndices[i] = messages.length - injectedIndices[i] - 1;
}
messages.reverse();
return injectedIndices;
}
@@ -8309,7 +8391,7 @@ function addDebugFunctions() {
registerDebugFunction('generationTest', 'Send a generation request', 'Generates text using the currently selected API.', async () => {
const text = prompt('Input text:', 'Hello');
toastr.info('Working on it...');
const message = await generateRaw(text, null, '');
const message = await generateRaw(text, null, false, false);
alert(message);
});