diff --git a/CHANGELOG.md b/CHANGELOG.md index b05a5af..e73a0e3 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,3 +1,8 @@ +# 2026-06-20 +## 体验改进 + - 现在支持调整在额外模型解析过程中，提供的消息楼层数。默认为 2。 + - 增加了应答格式 `格式化输出(v4兼容)`, 用于 官deepseek-v4 等渠道。（注：如果同时使用其他包含额外解析的插件，可能会有兼容性问题） + - # 2026-04-25 ## 体验改进 - 现在额外模型解析支持指定当前环境下的已有预设，代表使用那个预设的提示词结构，api/预设脚本/正则 等依然使用当前预设/手动指定的。 diff --git a/src/function/update/invoke_extra_model.ts b/src/function/update/invoke_extra_model.ts index 771cf9f..9b805a3 100644 --- a/src/function/update/invoke_extra_model.ts +++ b/src/function/update/invoke_extra_model.ts @@ -21,15 +21,143 @@ import { useDataStore } from '@/store'; import { normalizeBaseURL } from '@/util'; import { literalYamlify, uuidv4 } from '@util/common'; import { compare } from 'compare-versions'; +import YAML from 'yaml'; //测试用，为了使首次请求必失败 let debug_extra_request_counter = 0; +const V4_COMPATIBLE_FORMATTED_OUTPUT = '格式化输出(v4兼容)'; +const JSON_OBJECT_CUSTOM_INCLUDE_BODY = Object.freeze({ + response_format: { + type: 'json_object', + }, +}); +const DISABLED_THINKING_CUSTOM_INCLUDE_BODY = Object.freeze({ + thinking: { + type: 'disabled', + }, +}); + function generateRandomHeader(): string { return _.times(4, () => uuidv4().slice(0, 8)).join('\n'); } -function setExtraAnalysisStates() { +function isV4CompatibleFormattedOutput(): boolean { + return useDataStore().settings.额外模型解析配置.应答格式 === V4_COMPATIBLE_FORMATTED_OUTPUT; +} + +function assertV4CompatibleFormattedOutputUsable() { + const store = useDataStore(); + if ( + store.settings.额外模型解析配置.应答格式 === V4_COMPATIBLE_FORMATTED_OUTPUT && + store.settings.额外模型解析配置.模型来源 === '与插头相同' + ) { + throw new Error( + '[MVU额外模型解析]格式化输出(v4兼容)需要额外模型来源为自定义，不能与插头相同。' + ); + } +} + +function isPlainObject(value: unknown): value is Record { + return typeof value === 'object' && value !== null && !Array.isArray(value); +} + +function parseCustomIncludeBody(body: unknown): Record { + if (typeof body !== 'string' || body.trim() === '') { + return {}; + } + + const parsed = YAML.parse(body); + if (isPlainObject(parsed)) { + return parsed; + } + if (Array.isArray(parsed)) { + return Object.assign({}, ...parsed.filter(isPlainObject)); + } + throw new Error('[MVU额外模型解析]custom_include_body 不是 YAML object，无法合并配置。'); +} + +function buildJsonObjectCustomIncludeBody(original_body: unknown): string { + const store = useDataStore(); + return YAML.stringify({ + ...parseCustomIncludeBody(original_body), + ...JSON_OBJECT_CUSTOM_INCLUDE_BODY, + ...(store.settings.额外模型解析配置.关闭thinking + ? DISABLED_THINKING_CUSTOM_INCLUDE_BODY + : {}), + }).trimEnd(); +} + +async function saveSillyTavernSettings() { + const save_settings = + typeof builtin === 'undefined' ? undefined : builtin.saveSettings.bind(builtin); + if (typeof save_settings !== 'function') { + throw new Error('[MVU额外模型解析]无法获取 SillyTavern saveSettings，不能临时更新配置。'); + } + await save_settings(); +} + +let temporary_json_object_response_format_state: { + had_original_body: boolean; + original_body: unknown; +} | null = null; + +async function setTemporaryJsonObjectResponseFormat() { + if (!isV4CompatibleFormattedOutput()) { + temporary_json_object_response_format_state = null; + return; + } + + assertV4CompatibleFormattedOutputUsable(); + const oai_settings = SillyTavern.chatCompletionSettings; + if (!isPlainObject(oai_settings)) { + throw new Error('[MVU额外模型解析]无法获取 SillyTavern OpenAI 设置。'); + } + + const had_original_body = Object.prototype.hasOwnProperty.call( + oai_settings, + 'custom_include_body' + ); + const original_body = oai_settings.custom_include_body; + oai_settings.custom_include_body = buildJsonObjectCustomIncludeBody(original_body); + try { + await saveSillyTavernSettings(); + temporary_json_object_response_format_state = { + had_original_body, + original_body, + }; + } catch (error) { + if (had_original_body) { + oai_settings.custom_include_body = original_body; + } else { + delete oai_settings.custom_include_body; + } + temporary_json_object_response_format_state = null; + throw error; + } +} + +async function restoreTemporaryJsonObjectResponseFormat() { + if (!temporary_json_object_response_format_state) { + return; + } + + const oai_settings = SillyTavern.chatCompletionSettings; + if (!isPlainObject(oai_settings)) { + throw new Error('[MVU额外模型解析]无法获取 SillyTavern OpenAI 设置，不能恢复配置。'); + } + + const { had_original_body, original_body } = temporary_json_object_response_format_state; + temporary_json_object_response_format_state = null; + if (had_original_body) { + oai_settings.custom_include_body = original_body; + } else { + delete oai_settings.custom_include_body; + } + await saveSillyTavernSettings(); +} + +async function setExtraAnalysisStates() { const store = useDataStore(); if (store.runtimes.is_during_extra_analysis === true) { @@ -41,15 +169,25 @@ function setExtraAnalysisStates() { //因为这个操作是幂等的，所以无所谓。 store.runtimes.is_during_extra_analysis = true; + try { + await setTemporaryJsonObjectResponseFormat(); + } catch (error) { + store.runtimes.is_during_extra_analysis = false; + throw error; + } } -function unsetExtraAnalysisStates() { +async function unsetExtraAnalysisStates() { const store = useDataStore(); SillyTavern.unregisterMacro('lastUserMessage'); clearExtraModelRequestOverrides(); - store.runtimes.is_during_extra_analysis = false; store.runtimes.is_function_call_enabled = false; + try { + await restoreTemporaryJsonObjectResponseFormat(); + } finally { + store.runtimes.is_during_extra_analysis = false; + } } let is_analysis_in_progress = false; @@ -78,28 +216,36 @@ export async function invokeExtraModelWithStrategy(): Promise { is_manual_canceled: boolean; }> => { let is_manual_canceled = false; + let did_set_extra_analysis_states = false; try { - setExtraAnalysisStates(); + await setExtraAnalysisStates(); + did_set_extra_analysis_states = true; return { result: await recordedInvoke(), is_manual_canceled: false }; } catch (e) { /** 已经记录, 忽略 */ if (e === 'Clicked stop button') is_manual_canceled = true; } finally { - unsetExtraAnalysisStates(); + if (did_set_extra_analysis_states) { + await unsetExtraAnalysisStates(); + } } return { result: null, is_manual_canceled: is_manual_canceled }; }; const concurrentInvoke = async (times: number) => { const uuids = _.times(times, uuidv4); + let did_set_extra_analysis_states = false; try { - setExtraAnalysisStates(); + await setExtraAnalysisStates(); + did_set_extra_analysis_states = true; //在函数调用的模式下，允许接受 **任意** 有效的函数结果，因此被允许被覆盖。 return await Promise.any(uuids.map(recordedInvoke)); } catch (e) { /** 已经记录, 忽略 */ } finally { uuids.forEach(stopGenerationById); - unsetExtraAnalysisStates(); + if (did_set_extra_analysis_states) { + await unsetExtraAnalysisStates(); + } } return null; }; @@ -152,6 +298,8 @@ export async function invokeExtraModelWithStrategy(): Promise { ); } return concurrentInvoke(store.settings.额外模型解析配置.请求次数 - 1); + default: + return null; } } finally { is_analysis_in_progress = false; @@ -162,11 +310,15 @@ export async function invokeExtraModelWithStrategy(): Promise { * @brief 调用额外模型解析，可能会抛出异常。 */ export async function generateExtraModel(): Promise { + let did_set_extra_analysis_states = false; try { - setExtraAnalysisStates(); + await setExtraAnalysisStates(); + did_set_extra_analysis_states = true; return await invokeExtraModel(); } finally { - unsetExtraAnalysisStates(); + if (did_set_extra_analysis_states) { + await unsetExtraAnalysisStates(); + } } } @@ -245,7 +397,7 @@ function normalizeGenerateResultByResponseFormat( result: string | GenerateToolCallResult, response_format: string ): string { - if (response_format === '格式化输出') { + if (response_format === '格式化输出' || response_format === V4_COMPATIBLE_FORMATTED_OUTPUT) { const formatted = extractFromFormattedOutput(result); if (formatted) { return formatted; @@ -257,6 +409,9 @@ function normalizeGenerateResultByResponseFormat( async function requestReply(generation_id?: string, batch_id?: string): Promise { const store = useDataStore(); const response_format = store.settings.额外模型解析配置.应答格式; + const is_v4_compatible_formatted_output = response_format === V4_COMPATIBLE_FORMATTED_OUTPUT; + + assertV4CompatibleFormattedOutputUsable(); const config: GenerateRawConfig = { user_input: '遵循指令', @@ -280,6 +435,9 @@ async function requestReply(generation_id?: string, batch_id?: string): Promise< top_p: unset_if_equal(store.settings.额外模型解析配置.top_p, 1), top_k: unset_if_equal(store.settings.额外模型解析配置.top_k, 0), }; + if (is_v4_compatible_formatted_output) { + config.custom_api.source = 'custom'; + } } let task = decoded_extra_model_task; @@ -292,6 +450,10 @@ async function requestReply(generation_id?: string, batch_id?: string): Promise< task += '\n You are in formatted-output mode. Do not output tags, markdown, or prose. Return only a JSON object matching the provided json_schema: {"analysis":"...","json_patch":[...]}. Put MVU JsonPatch dialect operations in `json_patch`.'; config.json_schema = MVU_JSON_PATCH_RESPONSE_SCHEMA; + } else if (is_v4_compatible_formatted_output) { + task += + '\n You are in formatted-output mode. Do not output tags, markdown, or prose. Return only a JSON object: {"analysis":"...","json_patch":[...]}. Put MVU JsonPatch dialect operations in `json_patch`. Return exactly one JSON object that conforms to this schema:' + + JSON.stringify(MVU_JSON_PATCH_RESPONSE_SCHEMA.value); } //因为部分预设会用到 {{lastUserMessage}}，因此进行修正。 diff --git a/src/panel/update/Prompt.vue b/src/panel/update/Prompt.vue index 12bded0..6b52f01 100644 --- a/src/panel/update/Prompt.vue +++ b/src/panel/update/Prompt.vue @@ -29,6 +29,18 @@ /> + + + + 关闭 + + +