Merge pull request #3484 from SillyTavern/reasoning-parsing-streaming Implement reasoning parsing during streaming, based on provided prefix/suffix

d15d49b295b0f84c8ed1a480043f605e71e768f9

Cohee <18619528+Cohee1207@users.noreply.github.com>

Signed

Parents: 580856064e05faeb7375eb2b95a63648aea0d028 87afb1633bd80bb4c14bb58ecb3dd8782c0ea7b5

public/script.js+2 -1 Viewed

             // Update reasoning
             await this.reasoningHandler.process(messageId, mesChanged);
 +            processedText = chat[messageId]['mes'];
             // Token count update.
             const tokenCountText = this.reasoningHandler.reasoning + processedText;

@@ -3373,7 +3374,7 @@ class StreamingProcessor {
3373	3374	this.messageLogprobs.push(...(Array.isArray(logprobs) ? logprobs : [logprobs]));
3374	3375	}
3375	3376	// Get the updated reasoning string into the handler
3376	3377	this.reasoningHandler.updateReasoning(this.messageId, state?.reasoning ?? '');
3377	3378	await eventSource.emit(event_types.STREAM_TOKEN_RECEIVED, text);
3378	3379	await sw.tick(async () => await this.onProgressStreaming(this.messageId, this.continueMessage + text));
3379	3380	}

public/scripts/reasoning.js+114 -22 Viewed

 } from '../lib.js';
 import { chat, closeMessageEditor, event_types, eventSource, main_api, messageFormatting, saveChatConditional, saveSettingsDebounced, substituteParams, updateMessageBlock } from '../script.js';
 import { getRegexedString, regex_placement } from './extensions/regex/engine.js';
 import { getCurrentLocale, t, translate } from './i18n.js';
 import { MacrosParser } from './macros.js';
 import { chat_completion_sources, getChatCompletionModel, oai_settings } from './openai.js';
 import { Popup } from './popup.js';

 import { enumTypes, SlashCommandEnumValue } from './slash-commands/SlashCommandEnumValue.js';
 import { SlashCommandParser } from './slash-commands/SlashCommandParser.js';
 import { textgen_types, textgenerationwebui_settings } from './textgen-settings.js';
 import { copyText, escapeRegex, isFalseBoolean, setDatasetProperty, trimSpaces } from './utils.js';
++
 +/**
 + * Enum representing the type of the reasoning for a message (where it came from)
 + * @enum {string}
 + * @readonly
 + */
 +export const ReasoningType = {
 +    Model: 'model',
 +    Parsed: 'parsed',
 +    Manual: 'manual',
 +    Edited: 'edited',
 +};
 /**
  * Gets a message from a jQuery element.

  * This class is used inside the {@link StreamingProcessor} to manage reasoning states and UI updates.
  */
 export class ReasoningHandler {
 +    /** @type {boolean} True if the model supports reasoning, but hides the reasoning output */
     #isHiddenReasoningModel;
 +    /** @type {boolean} True if the handler is currently handling a manual parse of reasoning blocks */
 +    #isParsingReasoning = false;
 +    /** @type {number?} When reasoning is being parsed manually, and the reasoning has ended, this will be the index at which the actual messages starts */
 +    #parsingReasoningMesStartIndex = null;
     /**
      * @param {Date?} [timeStarted=null] - When the generation started

     constructor(timeStarted = null) {
         /** @type {ReasoningState} The current state of the reasoning process */
         this.state = ReasoningState.None;
 +        /** @type {ReasoningType?} The type of the reasoning (where it came from) */
 +        this.type = null;
         /** @type {string} The reasoning output */
         this.reasoning = '';
         /** @type {Date} When the reasoning started */

         /** @type {Date} Initial starting time of the generation */
         this.initialTime = timeStarted ?? new Date();
 -        /** @type {boolean} True if the model supports reasoning, but hides the reasoning output */
         this.#isHiddenReasoningModel = isHiddenReasoningModel();
         // Cached DOM elements for reasoning

             this.state = ReasoningState.Hidden;
+        }
 +        this.type = extra?.reasoning_type;
         this.reasoning = extra?.reasoning ?? '';
         if (this.state !== ReasoningState.None) {

         // Make sure reset correctly clears all relevant states
         if (reset) {
             this.state = this.#isHiddenReasoningModel ? ReasoningState.Thinking : ReasoningState.None;
 +            this.type = null;
             this.reasoning = '';
             this.initialTime = new Date();
             this.startTime = null;

      * Updates the reasoning text/string for a message.
+     *
      * @param {number} messageId - The ID of the message to update
      * @param {string?} [reasoning=null] - The reasoning text to update - If null or empty, uses the current reasoning
      * @param {Object} [options={}] - Optional arguments
      * @param {boolean} [options.persist=false] - Whether to persist the reasoning to the message object
 +     * @param {boolean} [options.allowReset=false] - Whether to allow empty reasoning provided to reset the reasoning, instead of just taking the existing one
      * @returns {boolean} - Returns true if the reasoning was changed, otherwise false
      */
     updateReasoning(messageId, reasoning = null, { persist = false, allowReset = false } = {}) {
         if (messageId == -1 || !chat[messageId]) {
             return false;
+        }
         reasoning = allowReset ? reasoning ?? this.reasoning : reasoning || this.reasoning;
         reasoning = power_user.trim_spaces ? reasoning.trimtrimSpaces() : reasoning);
         // Ensure the chat extra exists
         if (!chat[messageId].extra) {

         const reasoningChanged = extra.reasoning !== reasoning;
         this.reasoning = getRegexedString(reasoning ?? '', regex_placement.REASONING);
 +        this.type = (this.#isParsingReasoning || this.#parsingReasoningMesStartIndex) ? ReasoningType.Parsed : ReasoningType.Model;
++
         if (persist) {
             // Build and save the reasoning data to message extras
             extra.reasoning = this.reasoning;
             extra.reasoning_duration = this.getDuration();
 +            extra.reasoning_type = (this.#isParsingReasoning || this.#parsingReasoningMesStartIndex) ? ReasoningType.Parsed : ReasoningType.Model;
+        }
         return reasoningChanged;

      * @returns {Promise<void>}
      */
     async process(messageId, mesChanged) {
 -        if (!this.reasoning && !this.#isHiddenReasoningModel) return;
 +        mesChanged = this.#autoParseReasoningFromMessage(messageId, mesChanged);
++
 +        if (!this.reasoning && !this.#isHiddenReasoningModel)
 +            return;
         // Ensure reasoning string is updated and regexes are applied correctly
         const reasoningChanged = this.updateReasoning(messageId, null, { persist: true });

+        }
+    }
 +    #autoParseReasoningFromMessage(messageId, mesChanged) {
 +        if (!power_user.reasoning.auto_parse)
 +            return;
 +        if (!power_user.reasoning.prefix || !power_user.reasoning.suffix)
 +            return mesChanged;
++
 +        /** @type {{ mes: string, [key: string]: any}} */
 +        const message = chat[messageId];
 +        if (!message) return mesChanged;
++
 +        // If we are done with reasoning parse, we just split the message correctly so the reasoning doesn't show up inside of it.
 +        if (this.#parsingReasoningMesStartIndex) {
 +            message.mes = trimSpaces(message.mes.slice(this.#parsingReasoningMesStartIndex));
 +            return mesChanged;
 +        }
++
 +        if (this.state === ReasoningState.None) {
 +            // If streamed message starts with the opening, cut it out and put all inside reasoning
 +            if (message.mes.startsWith(power_user.reasoning.prefix) && message.mes.length > power_user.reasoning.prefix.length) {
 +                this.#isParsingReasoning = true;
++
 +                // Manually set starting state here, as we might already have received the ending suffix
 +                this.state = ReasoningState.Thinking;
 +                this.startTime = this.initialTime;
 +            }
 +        }
++
 +        if (!this.#isParsingReasoning)
 +            return mesChanged;
++
 +        // If we are in manual parsing mode, all currently streaming mes tokens will go the the reasoning block
 +        const originalMes = message.mes;
 +        this.reasoning = originalMes.slice(power_user.reasoning.prefix.length);
 +        message.mes = '';
++
 +        // If the reasoning contains the ending suffix, we cut that off and continue as message streaming
 +        if (this.reasoning.includes(power_user.reasoning.suffix)) {
 +            this.reasoning = this.reasoning.slice(0, this.reasoning.indexOf(power_user.reasoning.suffix));
 +            this.#parsingReasoningMesStartIndex = originalMes.indexOf(power_user.reasoning.suffix) + power_user.reasoning.suffix.length;
 +            message.mes = trimSpaces(originalMes.slice(this.#parsingReasoningMesStartIndex));
 +            this.#isParsingReasoning = false;
 +        }
++
 +        // Only return the original mesChanged value if we haven't cut off the complete message
 +        return message.mes.length ? mesChanged : false;
 +    }
++
     /**
      * Completes the reasoning process for a message.
+     *

         // Update states to the relevant DOM elements
         setDatasetProperty(this.messageDom, 'reasoningState', this.state !== ReasoningState.None ? this.state : null);
         setDatasetProperty(this.messageReasoningDetailsDom, 'state', this.state);
 +        setDatasetProperty(this.messageReasoningDetailsDom, 'type', this.type);
         // Update the reasoning message
         const reasoning = power_user.trim_spaces ? this.reasoning.trimtrimSpaces() : this.reasoning);
         const displayReasoning = messageFormatting(reasoning, '', false, false, messageId, {}, true);
         this.messageReasoningContentDom.innerHTML = displayReasoning;

         const element = this.messageReasoningHeaderDom;
         const duration = this.getDuration();
         let data = null;
 +        let title = '';
         if (duration) {
             const durationStrseconds = moment.duration(duration).locale(getCurrentLocale()).humanizeasSeconds({ s: 50, ss: 3 });
 -            const secondsStr = moment.duration(duration).asSeconds();
 -            const span = document.createElement('span');
 +            const durationStr = moment.duration(duration).locale(getCurrentLocale()).humanize({ s: 50, ss: 3 });
             spanelement.titletextContent = t`Thought for ${secondsStrdurationStr} seconds`;
             span.textContentdata = durationStrString(seconds);
+-
 +            title = `${seconds} seconds`;
 -            element.textContent = t`Thought for `;
 -            element.appendChild(span);
 -            data = String(secondsStr);
         } else if ([ReasoningState.Done, ReasoningState.Hidden].includes(this.state)) {
             element.textContent = t`Thought for some time`;
             data = 'unknown';

             data = null;
+        }
 +        if (this.type !== ReasoningType.Model) {
 +            title += ` [${translate(this.type)}]`;
 +            title = title.trim();
 +        }
 +        element.title = title;
++
         setDatasetProperty(this.messageReasoningDetailsDom, 'duration', data);
         setDatasetProperty(element, 'duration', data);
+    }

         callback: async (args, value) => {
             const messageId = !isNaN(Number(args.at)) ? Number(args.at) : chat.length - 1;
             const message = chat[messageId];
             if (!message?.extra) {
                 return '';
+            }
 +            // Make sure the message has an extra object
 +            if (!message.extra || typeof message.extra !== 'object') {
 +                message.extra = {};
 +            }
             message.extra.reasoning = String(value ?? '');
 +            message.extra.reasoning_type = ReasoningType.Manual;
             await saveChatConditional();
             closeMessageEditor('reasoning');

         const textarea = messageBlock.find('.reasoning_edit_textarea');
         const reasoning = getRegexedString(String(textarea.val()), regex_placement.REASONING, { isEdit: true });
         message.extra.reasoning = reasoning;
 +        message.extra.reasoning_type = message.extra.reasoning_type ? ReasoningType.Edited : ReasoningType.Manual;
         await saveChatConditional();
         updateMessageBlock(messageId, message);
         textarea.remove();

             return;
+        }
         message.extra.reasoning = '';
 +        delete message.extra.reasoning_type;
 +        delete message.extra.reasoning_duration;
         await saveChatConditional();
         updateMessageBlock(messageId, message);
         const textarea = messageBlock.find('.reasoning_edit_textarea');

             return '';
         });
         if (didReplace && power_user.trim_spaces) {
             reasoning = reasoning.trimtrimSpaces(reasoning);
             content = content.trimtrimSpaces(content);
+        }
         return { reasoning, content };

             return null;
+        }
 +        if (message.extra?.reasoning) {
 +            console.debug('[Reasoning] Message already has reasoning', idx);
 +            return null;
 +        }
++
         const parsedReasoning = parseReasoningFromString(message.mes);
         // No reasoning block found

         // If reasoning was found, add it to the message
         if (parsedReasoning.reasoning) {
             message.extra.reasoning = getRegexedString(parsedReasoning.reasoning, regex_placement.REASONING);
 +            message.extra.reasoning_type = ReasoningType.Parsed;
+        }
         // Update the message text if it was changed

public/scripts/utils.js+14 -1 Viewed

 import { getContext } from './extensions.js';
 import { characters, getRequestHeaders, this_chid } from '../script.js';
 import { isMobile } from './RossAscends-mods.js';
 import { collapseNewlines, power_user } from './power-user.js';
 import { debounce_timeout } from './constants.js';
 import { Popup, POPUP_RESULT, POPUP_TYPE } from './popup.js';
 import { SlashCommandClosure } from './slash-commands/SlashCommandClosure.js';

+}
 /**
 + * Trims leading and trailing whitespace from the input string based on a configuration setting.
 + * @param {string} input - The string to be trimmed
 + * @returns {string} The trimmed string if trimming is enabled; otherwise, returns the original string
 + */
++
 +export function trimSpaces(input) {
 +    if (!input || typeof input !== 'string') {
 +        return input;
 +    }
 +    return power_user.trim_spaces ? input.trim() : input;
 +}
++
 +/**
  * Trims a string to the end of a nearest sentence.
  * @param {string} input The string to trim.
  * @returns {string} The trimmed string.