Merge pull request #24 from ZipperWang/main

添加response API支持，添加流式传输支持
2026-10-03 18:33:14 +08:00 · 2026-08-26 17:29:54 +08:00
parent 7016ca1237 8f9a0c44da
commit b26d4dd1a7
10 changed files with 597 additions and 82 deletions
@@ -24,6 +24,7 @@ interface AIProviderMetadataFile {

 type AIMessagePart = { type: 'text'; text: string } | { type: 'image'; dataUrl: string }
 type AIMessage = { role: string; content: string | AIMessagePart[] }
+type AIChatDeltaHandler = (delta: string) => void
 type AIRequestResult = {
  data: string
  finishReason?: string
@@ -34,6 +35,29 @@ interface OpenAIResponsePayload {
  choices?: Array<{ message?: { content?: string }; finish_reason?: string }>
  usage?: { prompt_tokens?: number; completion_tokens?: number; total_tokens?: number }
 }
+interface OpenAIStreamPayload {
+  error?: { message?: string }
+  choices?: Array<{ delta?: { content?: string }; finish_reason?: string }>
+  usage?: { prompt_tokens?: number; completion_tokens?: number; total_tokens?: number }
+}
+interface OpenAIResponsesPayload {
+  error?: { message?: string }
+  incomplete_details?: { reason?: string }
+  status?: string
+  output_text?: string
+  output?: Array<{
+    type?: string
+    content?: Array<{ type?: string; text?: string }>
+  }>
+  usage?: { input_tokens?: number; output_tokens?: number; total_tokens?: number }
+}
+interface OpenAIResponsesStreamPayload {
+  type?: string
+  delta?: string
+  message?: string
+  error?: { message?: string }
+  response?: OpenAIResponsesPayload
+}
 interface AnthropicResponsePayload {
  error?: { message?: string }
  content?: Array<{ type?: string; text?: string }>
@@ -234,7 +258,8 @@ export class AIProviderService {
  async chat(
    messages: Array<{ role: string; content: string }>,
    options?: AIChatRequestOptions,
-    signal?: AbortSignal
+    signal?: AbortSignal,
+    onDelta?: AIChatDeltaHandler
  ): Promise<{
    success: boolean
    data?: string
@@ -242,7 +267,7 @@ export class AIProviderService {
    error?: string
  }> {
    try {
-      return { success: true, ...(await this.request(messages, options, false, signal)) }
+      return { success: true, ...(await this.request(messages, options, false, signal, onDelta)) }
    } catch (error) {
      if (signal?.aborted) throw error
      return { success: false, error: safeAIError(error) }
@@ -323,12 +348,13 @@ export class AIProviderService {
    messages: AIMessage[],
    options?: AIChatRequestOptions,
    testing = false,
-    signal?: AbortSignal
+    signal?: AbortSignal,
+    onDelta?: AIChatDeltaHandler
  ): Promise<{
    data: string
    usage?: { input?: number; output?: number; total?: number; estimated?: boolean }
  }> {
-    if (options?.apiKey) return this.requestLegacy(messages, options, signal)
+    if (options?.apiKey) return this.requestLegacy(messages, options, signal, onDelta)
    const resolved = this.resolveProvider(options)
    const provider = options?.timeoutMs
      ? {
@@ -336,7 +362,15 @@ export class AIProviderService {
          advanced: { ...resolved.provider.advanced, timeoutMs: options.timeoutMs }
        }
      : resolved.provider
-    return requestProvider(provider, resolved.key, resolved.model, messages, testing, signal)
+    return requestProvider(
+      provider,
+      resolved.key,
+      resolved.model,
+      messages,
+      testing,
+      signal,
+      onDelta
+    )
  }

  private resolveProvider(options?: { providerId?: string; modelId?: string }): {
@@ -359,7 +393,8 @@ export class AIProviderService {
  private async requestLegacy(
    messages: AIMessage[],
    options: AIChatRequestOptions,
-    signal?: AbortSignal
+    signal?: AbortSignal,
+    onDelta?: AIChatDeltaHandler
  ): Promise<AIRequestResult> {
    const provider = deepSeekProvider(options.baseURL, options.model)
    return requestOpenAICompatible(
@@ -368,7 +403,8 @@ export class AIProviderService {
      options.model || provider.defaultModel,
      messages,
      false,
-      signal
+      signal,
+      onDelta
    )
  }

@@ -542,6 +578,11 @@ function validateProvider(provider: AIProviderConfig): string | undefined {
    return '默认模型不在模型列表中'
  if (provider.auth.type === 'custom-header' && !provider.auth.headerName?.trim())
    return '请填写自定义认证字段'
+  if (
+    provider.advanced.apiProtocol &&
+    !['chat-completions', 'responses'].includes(provider.advanced.apiProtocol)
+  )
+    return 'OpenAI API 接口配置不正确'
  return undefined
 }

@@ -557,17 +598,26 @@ function buildHeaders(provider: AIProviderSummary, apiKey: string): Record<strin
  return headers
 }

+function hasHeader(headers: Record<string, string>, name: string): boolean {
+  const normalizedName = name.toLowerCase()
+  return Object.keys(headers).some((header) => header.toLowerCase() === normalizedName)
+}
+
 function requestProvider(
  provider: AIProviderSummary,
  apiKey: string,
  model: string,
  messages: AIMessage[],
  testing = false,
-  signal?: AbortSignal
+  signal?: AbortSignal,
+  onDelta?: AIChatDeltaHandler
 ): Promise<AIRequestResult> {
-  return provider.type === 'anthropic-messages'
-    ? requestAnthropic(provider, apiKey, model, messages, testing, signal)
-    : requestOpenAICompatible(provider, apiKey, model, messages, testing, signal)
+  if (provider.type === 'anthropic-messages') {
+    return requestAnthropic(provider, apiKey, model, messages, testing, signal)
+  }
+  return provider.advanced.apiProtocol === 'responses'
+    ? requestOpenAIResponses(provider, apiKey, model, messages, testing, signal, onDelta)
+    : requestOpenAICompatible(provider, apiKey, model, messages, testing, signal, onDelta)
 }

 function toOpenAIMessages(messages: AIMessage[]): Array<{ role: string; content: unknown }> {
@@ -584,6 +634,46 @@ function toOpenAIMessages(messages: AIMessage[]): Array<{ role: string; content:
  }))
 }

+function toOpenAIResponsesRequest(messages: AIMessage[]): {
+  instructions?: string
+  input: Array<{ role: string; content: unknown[] }>
+} {
+  const instructions = messages
+    .filter((message) => message.role === 'system')
+    .map((message) =>
+      typeof message.content === 'string'
+        ? message.content
+        : message.content
+            .filter((part) => part.type === 'text')
+            .map((part) => (part.type === 'text' ? part.text : ''))
+            .join('\n')
+    )
+    .filter(Boolean)
+    .join('\n\n')
+  const input = messages
+    .filter((message) => message.role !== 'system')
+    .map((message) => ({
+      role: message.role === 'assistant' ? 'assistant' : 'user',
+      content:
+        typeof message.content === 'string'
+          ? [
+              {
+                type: message.role === 'assistant' ? 'output_text' : 'input_text',
+                text: message.content
+              }
+            ]
+          : message.content.map((part) =>
+              part.type === 'text'
+                ? {
+                    type: message.role === 'assistant' ? 'output_text' : 'input_text',
+                    text: part.text
+                  }
+                : { type: 'input_image', image_url: part.dataUrl }
+            )
+    }))
+  return { instructions: instructions || undefined, input }
+}
+
 function toAnthropicMessages(messages: AIMessage[]): Array<{ role: string; content: unknown }> {
  return messages
    .filter((message) => message.role !== 'system')
@@ -603,24 +693,27 @@ function toAnthropicMessages(messages: AIMessage[]): Array<{ role: string; conte
    }))
 }

+function modelMaxTokens(provider: AIProviderSummary, model: string): number {
+  return (
+    provider.models.find((item) => item.id === model)?.maxTokens ||
+    provider.advanced.maxTokens ||
+    4096
+  )
+}
+
 async function requestOpenAICompatible(
  provider: AIProviderSummary,
  apiKey: string,
  model: string,
  messages: AIMessage[],
  testing = false,
-  signal?: AbortSignal
+  signal?: AbortSignal,
+  onDelta?: AIChatDeltaHandler
 ): Promise<AIRequestResult> {
  const endpoint = provider.baseUrl.endsWith('/chat/completions')
    ? provider.baseUrl
    : `${provider.baseUrl.replace(/\/+$/, '')}/chat/completions`
-  let modelMaxTokens = 0
-  for (const m of provider.models) {
-    if (m.id === model && m.maxTokens) {
-      modelMaxTokens = m.maxTokens
-    }
-  }
-  const response = await fetchWithTimeout(
+  return fetchWithTimeout(
    endpoint,
    {
      method: 'POST',
@@ -629,30 +722,85 @@ async function requestOpenAICompatible(
        model,
        messages: toOpenAIMessages(messages),
        temperature: provider.advanced.temperature,
-        max_tokens: testing
-          ? 8
-          : modelMaxTokens > 0
-            ? modelMaxTokens
-            : provider.advanced.maxTokens || 4096
+        max_tokens: testing ? 8 : modelMaxTokens(provider, model),
+        ...(provider.advanced.stream ? { stream: true } : {})
      })
    },
    provider.advanced.timeoutMs,
-    signal
+    signal,
+    async (response) => {
+      if (!response.ok) {
+        const payload = await parseJsonResponse<OpenAIResponsePayload>(response)
+        throw new Error(payload.error?.message || `AI 请求失败 (${response.status})`)
+      }
+      if (
+        provider.advanced.stream &&
+        !response.headers.get('content-type')?.toLowerCase().includes('application/json')
+      ) {
+        return parseOpenAIStream(response, onDelta)
+      }
+      const payload = await parseJsonResponse<OpenAIResponsePayload>(response)
+      return {
+        data: String(payload.choices?.[0]?.message?.content || ''),
+        finishReason: String(payload.choices?.[0]?.finish_reason || 'unknown'),
+        usage: toOpenAIUsage(payload.usage)
+      }
+    }
  )
-  const payload = await parseJsonResponse<OpenAIResponsePayload>(response)
-  if (!response.ok) throw new Error(payload.error?.message || `AI 请求失败 (${response.status})`)
-  return {
-    data: String(payload.choices?.[0]?.message?.content || ''),
-    finishReason: String(payload.choices?.[0]?.finish_reason || 'unknown'),
-    usage: payload.usage
-      ? {
-          input: payload.usage.prompt_tokens,
-          output: payload.usage.completion_tokens,
-          total: payload.usage.total_tokens,
-          estimated: false
-        }
-      : undefined
+}
+
+async function requestOpenAIResponses(
+  provider: AIProviderSummary,
+  apiKey: string,
+  model: string,
+  messages: AIMessage[],
+  testing = false,
+  signal?: AbortSignal,
+  onDelta?: AIChatDeltaHandler
+): Promise<AIRequestResult> {
+  const normalizedBaseUrl = provider.baseUrl.replace(/\/+$/, '')
+  const endpoint = normalizedBaseUrl.endsWith('/responses')
+    ? normalizedBaseUrl
+    : normalizedBaseUrl.endsWith('/chat/completions')
+      ? `${normalizedBaseUrl.slice(0, -'/chat/completions'.length)}/responses`
+      : `${normalizedBaseUrl}/responses`
+  const request = toOpenAIResponsesRequest(messages)
+  const headers = buildHeaders(provider, apiKey)
+  if (provider.advanced.stream && !hasHeader(headers, 'accept')) {
+    headers.accept = 'text/event-stream'
  }
+  return fetchWithTimeout(
+    endpoint,
+    {
+      method: 'POST',
+      headers,
+      body: JSON.stringify({
+        model,
+        instructions: request.instructions,
+        input: request.input,
+        temperature: provider.advanced.temperature,
+        max_output_tokens: testing ? 128 : modelMaxTokens(provider, model),
+        store: false,
+        ...(provider.advanced.stream ? { stream: true } : {})
+      })
+    },
+    provider.advanced.timeoutMs,
+    signal,
+    async (response) => {
+      if (!response.ok) {
+        const payload = await parseJsonResponse<OpenAIResponsesPayload>(response)
+        throw new Error(payload.error?.message || `AI 请求失败 (${response.status})`)
+      }
+      if (provider.advanced.stream) return parseOpenAIResponsesStream(response, onDelta)
+      const payload = await parseJsonResponse<OpenAIResponsesPayload>(response)
+      if (payload.error?.message) throw new Error(payload.error.message)
+      return {
+        data: extractOpenAIResponsesText(payload),
+        finishReason: openAIResponsesFinishReason(payload) || 'unknown',
+        usage: toOpenAIResponsesUsage(payload.usage)
+      }
+    }
+  )
 }

 async function requestAnthropic(
@@ -680,13 +828,7 @@ async function requestAnthropic(
  const endpoint = provider.baseUrl.endsWith('/messages')
    ? provider.baseUrl
    : `${provider.baseUrl.replace(/\/+$/, '')}/messages`
-  let modelMaxTokens = 0
-  for (const m of provider.models) {
-    if (m.id === model && m.maxTokens) {
-      modelMaxTokens = m.maxTokens
-    }
-  }
-  const response = await fetchWithTimeout(
+  return fetchWithTimeout(
    endpoint,
    {
      method: 'POST',
@@ -696,44 +838,44 @@ async function requestAnthropic(
        system: system || undefined,
        messages: anthropicMessages,
        temperature: provider.advanced.temperature,
-        max_tokens: testing
-          ? 8
-          : modelMaxTokens > 0
-            ? modelMaxTokens
-            : provider.advanced.maxTokens || 4096
+        max_tokens: testing ? 8 : modelMaxTokens(provider, model)
      })
    },
    provider.advanced.timeoutMs,
-    signal
+    signal,
+    async (response) => {
+      const payload = await parseJsonResponse<AnthropicResponsePayload>(response)
+      if (!response.ok)
+        throw new Error(payload.error?.message || `Anthropic 请求失败 (${response.status})`)
+      return {
+        data: Array.isArray(payload.content)
+          ? payload.content
+              .filter((item) => item.type === 'text')
+              .map((item) => item.text || '')
+              .join('\n')
+          : '',
+        finishReason: String(payload.stop_reason || 'unknown'),
+        usage: payload.usage
+          ? {
+              input: payload.usage.input_tokens,
+              output: payload.usage.output_tokens,
+              total:
+                Number(payload.usage.input_tokens || 0) + Number(payload.usage.output_tokens || 0),
+              estimated: false
+            }
+          : undefined
+      }
+    }
  )
-  const payload = await parseJsonResponse<AnthropicResponsePayload>(response)
-  if (!response.ok)
-    throw new Error(payload.error?.message || `Anthropic 请求失败 (${response.status})`)
-  return {
-    data: Array.isArray(payload.content)
-      ? payload.content
-          .filter((item) => item.type === 'text')
-          .map((item) => item.text || '')
-          .join('\n')
-      : '',
-    finishReason: String(payload.stop_reason || 'unknown'),
-    usage: payload.usage
-      ? {
-          input: payload.usage.input_tokens,
-          output: payload.usage.output_tokens,
-          total: Number(payload.usage.input_tokens || 0) + Number(payload.usage.output_tokens || 0),
-          estimated: false
-        }
-      : undefined
-  }
 }

-async function fetchWithTimeout(
+async function fetchWithTimeout<T>(
  url: string,
  init: RequestInit,
  timeoutMs: number,
-  signal?: AbortSignal
-): Promise<Response> {
+  signal: AbortSignal | undefined,
+  consume: (response: Response) => Promise<T>
+): Promise<T> {
  const controller = new AbortController()
  let timedOut = false
  const abortFromCaller = (): void =>
@@ -748,7 +890,8 @@ async function fetchWithTimeout(
    Math.max(1_000, timeoutMs || 120_000)
  )
  try {
-    return await fetch(url, { ...init, signal: controller.signal })
+    const response = await fetch(url, { ...init, signal: controller.signal })
+    return await consume(response)
  } catch (error) {
    if (signal?.aborted) throw new DOMException('AI request cancelled', 'AbortError')
    if (timedOut) throw new DOMException('AI request timed out', 'TimeoutError')
@@ -759,6 +902,218 @@ async function fetchWithTimeout(
  }
 }

+async function parseOpenAIStream(
+  response: Response,
+  onDelta?: AIChatDeltaHandler
+): Promise<AIRequestResult> {
+  let data = ''
+  let finishReason = 'unknown'
+  let usage: AIRequestResult['usage']
+  const stream = await readSSE(response, (eventData) => {
+    let payload: OpenAIStreamPayload
+    try {
+      payload = JSON.parse(eventData) as OpenAIStreamPayload
+    } catch {
+      throw new Error('模型服务返回了无法解析的流式数据')
+    }
+    if (payload.error?.message) throw new Error(payload.error.message)
+    if (payload.usage) usage = toOpenAIUsage(payload.usage)
+    const choice = payload.choices?.[0]
+    if (choice?.finish_reason) finishReason = choice.finish_reason
+    const delta = choice?.delta?.content
+    if (typeof delta === 'string' && delta) {
+      data += delta
+      onDelta?.(delta)
+    }
+  })
+
+  if (!stream.sawData) {
+    let payload: OpenAIResponsePayload
+    try {
+      payload = JSON.parse(stream.rawBody) as OpenAIResponsePayload
+    } catch {
+      throw new Error('模型服务返回了无法解析的流式数据')
+    }
+    if (payload.error?.message) throw new Error(payload.error.message)
+    const content = String(payload.choices?.[0]?.message?.content || '')
+    if (content) onDelta?.(content)
+    return {
+      data: content,
+      finishReason: String(payload.choices?.[0]?.finish_reason || 'unknown'),
+      usage: toOpenAIUsage(payload.usage)
+    }
+  }
+  return { data, finishReason, usage }
+}
+
+async function parseOpenAIResponsesStream(
+  response: Response,
+  onDelta?: AIChatDeltaHandler
+): Promise<AIRequestResult> {
+  let data = ''
+  let completedResponse: OpenAIResponsesPayload | undefined
+  let usage: AIRequestResult['usage']
+  const stream = await readSSE(response, (eventData) => {
+    let payload: OpenAIResponsesStreamPayload
+    try {
+      payload = JSON.parse(eventData) as OpenAIResponsesStreamPayload
+    } catch {
+      throw new Error('模型服务返回了无法解析的 Responses 流式数据')
+    }
+    const errorMessage = payload.error?.message || payload.response?.error?.message
+    if (errorMessage) throw new Error(errorMessage)
+    if (payload.response?.usage) usage = toOpenAIResponsesUsage(payload.response.usage)
+    if (payload.type === 'response.output_text.delta' && typeof payload.delta === 'string') {
+      data += payload.delta
+      onDelta?.(payload.delta)
+    }
+    if (payload.type === 'response.completed') completedResponse = payload.response
+    if (payload.type === 'response.failed' || payload.type === 'response.incomplete') {
+      const reason = payload.response?.incomplete_details?.reason
+      throw new Error(reason ? `模型响应未完成：${reason}` : payload.message || '模型响应未完成')
+    }
+  })
+
+  if (!stream.sawData) {
+    let payload: OpenAIResponsesPayload
+    try {
+      payload = JSON.parse(stream.rawBody) as OpenAIResponsesPayload
+    } catch {
+      throw new Error('模型服务返回了无法解析的 Responses 流式数据')
+    }
+    if (payload.error?.message) throw new Error(payload.error.message)
+    const content = extractOpenAIResponsesText(payload)
+    if (content) onDelta?.(content)
+    return {
+      data: content,
+      finishReason: openAIResponsesFinishReason(payload) || 'unknown',
+      usage: toOpenAIResponsesUsage(payload.usage)
+    }
+  }
+
+  if (!data && completedResponse) {
+    data = extractOpenAIResponsesText(completedResponse)
+    if (data) onDelta?.(data)
+  }
+  return {
+    data,
+    finishReason: completedResponse
+      ? openAIResponsesFinishReason(completedResponse) || 'unknown'
+      : 'unknown',
+    usage
+  }
+}
+
+async function readSSE(
+  response: Response,
+  onEvent: (eventData: string) => void
+): Promise<{ sawData: boolean; rawBody: string }> {
+  if (!response.body) throw new Error('模型服务未返回可读取的流式响应')
+  const reader = response.body.getReader()
+  const decoder = new TextDecoder()
+  let buffer = ''
+  let rawBody = ''
+  let sawData = false
+  let finished = false
+
+  const consumeEvent = (event: string): boolean => {
+    const eventData = event
+      .split(/\r\n|\r|\n/)
+      .filter((line) => line.startsWith('data:'))
+      .map((line) => {
+        const value = line.slice(5)
+        return value.startsWith(' ') ? value.slice(1) : value
+      })
+      .join('\n')
+    if (!eventData) return false
+    sawData = true
+    rawBody = ''
+    if (eventData.trim() === '[DONE]') return true
+    onEvent(eventData)
+    return false
+  }
+
+  const consumeBuffer = (): void => {
+    while (!finished) {
+      const boundary = /(?:\r\n|\r|\n){2}/.exec(buffer)
+      if (!boundary) return
+      const event = buffer.slice(0, boundary.index)
+      buffer = buffer.slice(boundary.index + boundary[0].length)
+      finished = consumeEvent(event)
+    }
+  }
+
+  try {
+    while (!finished) {
+      const chunk = await reader.read()
+      if (chunk.done) {
+        const tail = decoder.decode()
+        buffer += tail
+        if (!sawData) rawBody += tail
+        consumeBuffer()
+        break
+      }
+      const text = decoder.decode(chunk.value, { stream: true })
+      buffer += text
+      if (!sawData) rawBody += text
+      consumeBuffer()
+    }
+    if (!finished && buffer.trim()) finished = consumeEvent(buffer)
+    if (finished) await reader.cancel().catch(() => undefined)
+  } catch (error) {
+    await reader.cancel(error).catch(() => undefined)
+    throw error
+  } finally {
+    reader.releaseLock()
+  }
+
+  return { sawData, rawBody }
+}
+
+function extractOpenAIResponsesText(payload: OpenAIResponsesPayload): string {
+  if (typeof payload.output_text === 'string') return payload.output_text
+  return (payload.output || [])
+    .flatMap((item) => item.content || [])
+    .filter((item) => item.type === 'output_text')
+    .map((item) => item.text || '')
+    .join('')
+}
+
+function openAIResponsesFinishReason(payload: OpenAIResponsesPayload): string | undefined {
+  const reason = payload.incomplete_details?.reason
+  if (reason === 'max_output_tokens') return 'max_tokens'
+  if (reason) return reason
+  if (payload.status === 'completed') return 'stop'
+  if (payload.status === 'failed') return 'error'
+  return undefined
+}
+
+function toOpenAIResponsesUsage(
+  usage: OpenAIResponsesPayload['usage']
+): AIRequestResult['usage'] | undefined {
+  return usage
+    ? {
+        input: usage.input_tokens,
+        output: usage.output_tokens,
+        total: usage.total_tokens,
+        estimated: false
+      }
+    : undefined
+}
+
+function toOpenAIUsage(
+  usage: OpenAIResponsePayload['usage']
+): AIRequestResult['usage'] | undefined {
+  return usage
+    ? {
+        input: usage.prompt_tokens,
+        output: usage.completion_tokens,
+        total: usage.total_tokens,
+        estimated: false
+      }
+    : undefined
+}
+
 async function parseJsonResponse<T>(response: Response): Promise<T> {
  const body = await response.text()
  try {
@@ -145,6 +145,34 @@ const isAbortError = (error: unknown): boolean =>
  (error instanceof DOMException && error.name === 'AbortError') ||
  (error instanceof Error && error.name === 'AbortError')

+const createDeltaBatcher = (
+  publish: (delta: string) => void,
+  intervalMs = 40
+): { push: (delta: string) => void; close: (flushPending: boolean) => void } => {
+  let pending = ''
+  let timer: ReturnType<typeof setTimeout> | undefined
+  const flush = (): void => {
+    timer = undefined
+    if (!pending) return
+    const delta = pending
+    pending = ''
+    publish(delta)
+  }
+  return {
+    push: (delta: string): void => {
+      if (!delta) return
+      pending += delta
+      if (!timer) timer = setTimeout(flush, intervalMs)
+    },
+    close: (flushPending: boolean): void => {
+      if (timer) clearTimeout(timer)
+      timer = undefined
+      if (flushPending) flush()
+      else pending = ''
+    }
+  }
+}
+
 /**
 * Main-process search orchestrator. It owns the only transition from raw
 * candidates to Final Evidence; the AI and Renderer never receive a wider
@@ -791,6 +819,23 @@ export class AiSearchPipelineService {
        timings: snapshotTimings()
      })
      const aiGenerationStartedAt = Date.now()
+      const answerDeltaBatcher = createDeltaBatcher((answerDelta) =>
+        emit({
+          stage: 'ai_generating',
+          status: 'running',
+          message: '正在生成带来源的回答',
+          answerDelta,
+          plan,
+          modelName: aiConfig.modelName,
+          stats: {
+            matchedMessages: evidenceBuild.candidateCount,
+            evidenceCount: evidence.length,
+            contextEvidenceCount: evidence.length,
+            tokenEstimate
+          },
+          timings: snapshotTimings()
+        })
+      )
      const answer = await this.chatForSearchRequest(
        request.requestId,
        aiConfig.providerId,
@@ -803,8 +848,9 @@ export class AiSearchPipelineService {
          },
          { role: 'user', content: prompt }
        ],
-        signal
-      )
+        signal,
+        answerDeltaBatcher.push
+      ).finally(() => answerDeltaBatcher.close(!signal.aborted))
      signal.throwIfAborted()
      timings.aiGenerationMs = Date.now() - aiGenerationStartedAt
      if (!answer.success || !answer.data) {
@@ -1009,13 +1055,14 @@ export class AiSearchPipelineService {
    providerId: string | undefined,
    modelId: string,
    messages: Array<{ role: string; content: string }>,
-    signal: AbortSignal
+    signal: AbortSignal,
+    onDelta?: (delta: string) => void
  ): ReturnType<AIProviderService['chat']> {
    if (!this.canUseAiForRequest(requestId, providerId)) {
      return { success: false, error: '当前搜索请求未授权向该 AI 服务发送内容' }
    }
    signal.throwIfAborted()
-    return this.aiProvider.chat(messages, { providerId, modelId }, signal)
+    return this.aiProvider.chat(messages, { providerId, modelId }, signal, onDelta)
  }

  private clearPendingAuthorization(requestId: string): void {
@@ -144,7 +144,9 @@ export function AISearchWorkspace({
    startSearch,
    cancelSearch,
    resetSearchRun
-  } = useAiSearchRun()
+  } = useAiSearchRun({
+    onAnswerDelta: (delta) => setAnswer((current) => current + delta)
+  })

  const resetSearchResult = (): void => {
    const reset = createSearchResultResetState()
@@ -520,6 +522,19 @@ export function AISearchWorkspace({
            </div>
          </section>
        </div>
+        {answer && (
+          <section
+            className="ai-search-summary-block ai-search-streaming-answer"
+            aria-label="正在生成的回答"
+            aria-live="polite"
+          >
+            <div className="ai-search-section-heading">
+              <span />
+              回答生成中
+            </div>
+            <div className="ai-search-answer">{renderMarkdown(answer)}</div>
+          </section>
+        )}
      </div>
    )
  }
@@ -24,7 +24,11 @@ export type AiSearchRunOutcome =
 const errorMessage = (error: unknown): string =>
  error instanceof Error ? error.message : '读取聊天记录失败'

-export function useAiSearchRun(): {
+export function useAiSearchRun({
+  onAnswerDelta
+}: {
+  onAnswerDelta?: (delta: string) => void
+} = {}): {
  status: AiSearchRunStatus
  requestId: string
  result: AiSearchPipelineResult | null
@@ -44,6 +48,11 @@ export function useAiSearchRun(): {
  const [progress, setProgress] = useState<SearchProgressByStage>({})
  const [agentTrace, setAgentTrace] = useState<AiSearchAgentRun['trace']>([])
  const requestIdRef = useRef('')
+  const onAnswerDeltaRef = useRef(onAnswerDelta)
+
+  useEffect(() => {
+    onAnswerDeltaRef.current = onAnswerDelta
+  }, [onAnswerDelta])

  const createRequestId = (): string => globalThis.crypto?.randomUUID?.() || `search-${Date.now()}`

@@ -54,6 +63,7 @@ export function useAiSearchRun(): {
    const unsubscribe = window.api.onAiSearchProgress((event: AiSearchProgressEvent) => {
      if (!isCurrentRequest(event.requestId)) return
      setProgress((current) => ({ ...current, [event.stage]: event }))
+      if (event.answerDelta) onAnswerDeltaRef.current?.(event.answerDelta)
      if (event.agentTrace) {
        setAgentTrace((current) =>
          current.some((item) => item.sequence === event.agentTrace?.sequence)
@@ -332,6 +332,49 @@ export function AIProviderEditor({
              }
            />
          </label>
+          {provider.type !== 'anthropic-messages' && (
+            <>
+              <label className="wide">
+                OpenAI API 接口
+                <Select
+                  value={provider.advanced.apiProtocol || 'chat-completions'}
+                  onValueChange={(value) =>
+                    patch({
+                      advanced: {
+                        ...provider.advanced,
+                        apiProtocol: value as AIProviderConfig['advanced']['apiProtocol']
+                      }
+                    })
+                  }
+                >
+                  <SelectTrigger aria-label="OpenAI API 接口">
+                    <SelectValue />
+                  </SelectTrigger>
+                  <SelectContent>
+                    <SelectItem value="chat-completions">Chat Completions（兼容模式）</SelectItem>
+                    <SelectItem value="responses">Responses API</SelectItem>
+                  </SelectContent>
+                </Select>
+                <small>Codex 或仅支持 /responses 的中转服务请选择 Responses API。</small>
+              </label>
+              <label className="wide ai-provider-stream-toggle" htmlFor="ai-provider-stream">
+                <span className="ai-provider-stream-toggle-control">
+                  <Checkbox
+                    id="ai-provider-stream"
+                    aria-label="启用流式响应"
+                    checked={Boolean(provider.advanced.stream)}
+                    onCheckedChange={(checked) =>
+                      patch({
+                        advanced: { ...provider.advanced, stream: checked === true }
+                      })
+                    }
+                  />
+                  启用流式响应（stream: true）
+                </span>
+                <small>问问微信会边生成边显示，其他功能仍会等待完整结果。</small>
+              </label>
+            </>
+          )}
          <label className="wide">
            额外 Headers（JSON）
            <Textarea
@@ -92,6 +92,13 @@ export function createProviderFromPreset(presetId = 'deepseek'): AIProviderConfi
      }
    ],
    defaultModel: preset.model,
-    advanced: { timeoutMs: 120000, temperature: 0.7, maxTokens: 4096, extraHeaders: {} }
+    advanced: {
+      timeoutMs: 120000,
+      temperature: 0.7,
+      maxTokens: 4096,
+      stream: false,
+      apiProtocol: 'chat-completions',
+      extraHeaders: {}
+    }
  }
 }
@@ -623,6 +623,27 @@
  white-space: pre-wrap;
 }

+.ai-search-streaming-answer {
+  width: 100%;
+}
+
+.ai-search-streaming-answer .ai-search-answer::after {
+  content: '';
+  display: inline-block;
+  width: 2px;
+  height: 1em;
+  margin-left: 3px;
+  background: var(--wxex-ai);
+  vertical-align: -0.12em;
+  animation: ai-search-stream-cursor 0.8s steps(2, start) infinite;
+}
+
+@keyframes ai-search-stream-cursor {
+  50% {
+    opacity: 0;
+  }
+}
+
 .ai-search-answer-evidence,
 .ai-search-trace {
  display: flex;
@@ -792,6 +813,9 @@
  .ai-search-result,
  .ai-search-summary-block,
  .ai-search-answer > *,
+  .ai-search-streaming-answer .ai-search-answer::after,
+  .ai-search-evidence-card,
+  .ai-search-history-popover,
  .ai-search-evidence-focus-flash,
  .ai-search-pipeline-step.active .ai-search-pipeline-mark,
  .ai-search-sync-progress-track span {
@@ -1553,6 +1553,15 @@
  min-height: 90px;
  resize: vertical;
 }
+.ai-provider-stream-toggle-control {
+  display: flex;
+  align-items: center;
+  gap: 7px;
+}
+.ai-provider-stream-toggle small {
+  color: #66706b;
+  font-weight: 400;
+}
 .ai-provider-preview pre {
  overflow: auto;
  max-height: 280px;
@@ -7,6 +7,8 @@ export type AIProviderType =

 export type AIAuthType = 'bearer' | 'x-api-key' | 'custom-header' | 'none'

+export type AIOpenAIProtocol = 'chat-completions' | 'responses'
+
 export interface AIProviderAuth {
  type: AIAuthType
  headerName?: string
@@ -34,6 +36,8 @@ export interface AIProviderAdvancedSettings {
  timeoutMs: number
  temperature?: number
  maxTokens?: number
+  stream?: boolean
+  apiProtocol?: AIOpenAIProtocol
  extraHeaders: Record<string, string>
 }

@@ -73,6 +73,7 @@ export interface AiSearchProgressEvent {
  stage: AiSearchProgressStage
  status: AiSearchProgressStatus
  message: string
+  answerDelta?: string
  plan?: AiSearchPlan
  stats?: {
    knowledgeMessageCount?: number