adding prefill parameter to generateRaw() (#4266)
* adding prefill parameter to generateRaw() * additional comments * Substitute params in prefill * Add prefill prompt for /genraw command * Fix param comment * fixing generateRaw param comment * fixing param comment --------- Co-authored-by: qvink <qvink@users.noreply.github.com> Co-authored-by: Cohee <18619528+Cohee1207@users.noreply.github.com>
This commit is contained in:
qvink
+17
-7
@@ -3058,10 +3058,11 @@ class StreamingProcessor {
|
||||
* @param {string} api API to use.
|
||||
* @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
|
||||
* @param {string} [systemPrompt] System prompt to use. Only Instruct mode or OpenAI.
|
||||
* @param {string} [systemPrompt] System prompt to use.
|
||||
* @param {string} [prefill] Prefill for the prompt.
|
||||
* @returns {string | object[]} Prompt ready for use in generation. If using TC, this will be a string. If using CC, this will be an array of chat-style messages.
|
||||
*/
|
||||
export function createRawPrompt(prompt, api, instructOverride, quietToLoud, systemPrompt) {
|
||||
export function createRawPrompt(prompt, api, instructOverride, quietToLoud, systemPrompt, prefill) {
|
||||
const isInstruct = power_user.instruct.enabled && api !== 'openai' && api !== 'novel' && !instructOverride;
|
||||
|
||||
// If the prompt was given as a string, convert to a message-style object assuming user role
|
||||
@@ -3074,6 +3075,9 @@ export function createRawPrompt(prompt, api, instructOverride, quietToLoud, syst
|
||||
if (prompt.length === 0 && !systemPrompt) throw Error('No messages provided');
|
||||
}
|
||||
|
||||
// Substitute the prefill if provided
|
||||
prefill = substituteParams(prefill ?? '');
|
||||
|
||||
// Format each message in the prompt, accounting for the provided roles
|
||||
for (const message of prompt) {
|
||||
let name = '';
|
||||
@@ -3096,12 +3100,17 @@ export function createRawPrompt(prompt, api, instructOverride, quietToLoud, syst
|
||||
prompt.unshift({ role: 'system', content: systemPrompt });
|
||||
}
|
||||
|
||||
// If text completion, convert to text prompt by concatenating all message contents
|
||||
// with Chat Completion, the prefill is an additional assistant message at the end.
|
||||
if (api === 'openai' && prefill) {
|
||||
prompt.push({ role: 'assistant', content: prefill });
|
||||
}
|
||||
|
||||
// if text completion, convert to text prompt by concatenating all message contents and adding the prefill as a promptBias.
|
||||
if (api !== 'openai') {
|
||||
const joiner = isInstruct ? '' : '\n';
|
||||
prompt = prompt.map(message => message.content).join(joiner);
|
||||
prompt = api === 'novel' ? adjustNovelInstructionPrompt(prompt) : prompt;
|
||||
prompt = prompt + (isInstruct ? formatInstructModePrompt(name2, false, '', name1, name2, true, quietToLoud) : '\n'); // add last line
|
||||
prompt = prompt + (isInstruct ? formatInstructModePrompt(name2, false, prefill, name1, name2, true, quietToLoud) : `\n${prefill}`); // add last line
|
||||
}
|
||||
|
||||
return prompt;
|
||||
@@ -3114,12 +3123,13 @@ export function createRawPrompt(prompt, api, instructOverride, quietToLoud, syst
|
||||
* @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
|
||||
* @param {string} [systemPrompt] System prompt to use. Only Instruct mode or OpenAI.
|
||||
* @param {string} [systemPrompt] System prompt to use.
|
||||
* @param {number} [responseLength] Maximum response length. If unset, the global default value is used.
|
||||
* @param {boolean} [trimNames] Whether to allow trimming "{{user}}:" and "{{char}}:" from the response.
|
||||
* @param {string} [prefill] An optional prefill for the prompt.
|
||||
* @returns {Promise<string>} Generated message
|
||||
*/
|
||||
export async function generateRaw(prompt, api, instructOverride, quietToLoud, systemPrompt, responseLength, trimNames = true) {
|
||||
export async function generateRaw(prompt, api, instructOverride, quietToLoud, systemPrompt, responseLength, trimNames = true, prefill = '') {
|
||||
if (!api) {
|
||||
api = main_api;
|
||||
}
|
||||
@@ -3129,7 +3139,7 @@ export async function generateRaw(prompt, api, instructOverride, quietToLoud, sy
|
||||
let eventHook = () => { };
|
||||
|
||||
// construct final prompt from the input. Can either be a string or an array of chat-style messages.
|
||||
prompt = createRawPrompt(prompt, api, instructOverride, quietToLoud, systemPrompt);
|
||||
prompt = createRawPrompt(prompt, api, instructOverride, quietToLoud, systemPrompt, prefill);
|
||||
|
||||
try {
|
||||
if (responseLengthCustomized) {
|
||||
|
||||
@@ -1672,17 +1672,18 @@ export function initDefaultSlashCommands() {
|
||||
returns: 'generated text',
|
||||
namedArgumentList: [
|
||||
new SlashCommandNamedArgument(
|
||||
'lock', 'lock user input during generation', [ARGUMENT_TYPE.BOOLEAN], false, false, null, commonEnumProviders.boolean('onOff')(),
|
||||
'lock', 'lock user input during generation', [ARGUMENT_TYPE.BOOLEAN], false, false, 'off', commonEnumProviders.boolean('onOff')(),
|
||||
),
|
||||
new SlashCommandNamedArgument(
|
||||
'instruct', 'use instruct mode', [ARGUMENT_TYPE.BOOLEAN], false, false, 'on', commonEnumProviders.boolean('onOff')(),
|
||||
),
|
||||
new SlashCommandNamedArgument(
|
||||
'stop', 'one-time custom stop strings', [ARGUMENT_TYPE.LIST], false,
|
||||
'stop', 'one-time custom stop strings', [ARGUMENT_TYPE.LIST], false, false, '[]',
|
||||
),
|
||||
SlashCommandNamedArgument.fromProps({
|
||||
name: 'as',
|
||||
description: 'role of the output prompt',
|
||||
defaultValue: 'system',
|
||||
typeList: [ARGUMENT_TYPE.STRING],
|
||||
enumList: [
|
||||
new SlashCommandEnumValue('system', null, enumTypes.enum, enumIcons.assistant),
|
||||
@@ -1690,10 +1691,16 @@ export function initDefaultSlashCommands() {
|
||||
],
|
||||
}),
|
||||
new SlashCommandNamedArgument(
|
||||
'system', 'system prompt at the start', [ARGUMENT_TYPE.STRING], false,
|
||||
'system', 'system prompt at the start', [ARGUMENT_TYPE.STRING, ARGUMENT_TYPE.VARIABLE_NAME], false,
|
||||
),
|
||||
new SlashCommandNamedArgument(
|
||||
'length', 'API response length in tokens', [ARGUMENT_TYPE.NUMBER], false,
|
||||
'prefill', 'prefill prompt at the end', [ARGUMENT_TYPE.STRING, ARGUMENT_TYPE.VARIABLE_NAME], false,
|
||||
),
|
||||
new SlashCommandNamedArgument(
|
||||
'length', 'API response length in tokens', [ARGUMENT_TYPE.NUMBER, ARGUMENT_TYPE.VARIABLE_NAME], false,
|
||||
),
|
||||
new SlashCommandNamedArgument(
|
||||
'trim', 'trim {{user}} and {{char}} prefixes from the output', [ARGUMENT_TYPE.BOOLEAN], false, false, 'on', commonEnumProviders.boolean('onOff')(),
|
||||
),
|
||||
],
|
||||
unnamedArgumentList: [
|
||||
@@ -3628,7 +3635,9 @@ async function generateRawCallback(args, value) {
|
||||
const as = args?.as || 'system';
|
||||
const quietToLoud = as === 'char';
|
||||
const systemPrompt = resolveVariable(args?.system) || '';
|
||||
const prefillPrompt = resolveVariable(args?.prefill) || '';
|
||||
const length = Number(resolveVariable(args?.length) ?? 0) || 0;
|
||||
const trimNames = !isFalseBoolean(args?.trim);
|
||||
|
||||
try {
|
||||
if (lock) {
|
||||
@@ -3636,7 +3645,7 @@ async function generateRawCallback(args, value) {
|
||||
}
|
||||
|
||||
setEphemeralStopStrings(resolveVariable(args?.stop));
|
||||
const result = await generateRaw(value, '', isFalseBoolean(args?.instruct), quietToLoud, systemPrompt, length);
|
||||
const result = await generateRaw(value, '', isFalseBoolean(args?.instruct), quietToLoud, systemPrompt, length, trimNames, prefillPrompt);
|
||||
return result;
|
||||
} catch (err) {
|
||||
console.error('Error on /genraw generation', err);
|
||||
|
||||
Reference in New Issue
Block a user