mirror of
https://github.com/hicccc77/WeFlow.git
synced 2026-10-25 17:36:32 +08:00
Yi Zhan
be0e564031
OpenAI Responses rejects temperature for gpt-6-* the same way it does for gpt-5 and the o-series. Share one omitsTemperature() helper so agent chat, title, and memory synthesis all skip it. Signed-off-by: Yi Zhan <stevesough@gmail.com>
88 lines
3.8 KiB
TypeScript
88 lines
3.8 KiB
TypeScript
import { streamText } from 'ai'
|
||
import { createAgentModel } from './agentModelAdapter'
|
||
import { omitsTemperature } from './agentSamplingOptions'
|
||
import { createAgentTextRedactionSession } from './agentTextRedaction'
|
||
import type { AgentModelConfig } from './agentService'
|
||
|
||
export type AgentTitleResult = {
|
||
title: string
|
||
generated: boolean
|
||
error?: string
|
||
}
|
||
|
||
function fallbackTitle(value: string): string {
|
||
return value
|
||
.replace(/@\S+\[[^\]]+\]/g, '')
|
||
.replace(/[？?。！!，,、：:\s]+/g, ' ')
|
||
.trim()
|
||
.slice(0, 18) || '新对话'
|
||
}
|
||
|
||
function titleSamplingOptions(config: AgentModelConfig): { temperature?: number } {
|
||
const modelId = String(config.model || '').trim().toLocaleLowerCase()
|
||
const modernClaudeModel = /(?:^|\/)claude-(?:opus-4-(?:[7-9]|\d{2,})|(?:sonnet|opus|haiku|fable|mythos)-(?:[5-9]|\d{2,})(?:-|$))/.test(modelId)
|
||
return omitsTemperature(config) || modernClaudeModel ? {} : { temperature: 0.2 }
|
||
}
|
||
|
||
function titleErrorMessage(error: unknown): string {
|
||
if (error instanceof Error) return error.message || error.name
|
||
return String(error || '标题模型调用失败')
|
||
}
|
||
|
||
export function normalizeGeneratedAgentTitle(value: unknown): string | null {
|
||
const envelope = /^\s*<title>([^<>\r\n]+)<\/title>\s*$/iu.exec(String(value || ''))
|
||
if (!envelope) return null
|
||
const title = envelope[1]
|
||
.replace(/^[“”"'《》【】\s]+|[“”"'《》【】。！？!?\s]+$/g, '')
|
||
.replace(/\s+/g, ' ')
|
||
.trim()
|
||
if (!title || title.length < 2 || title.length > 24) return null
|
||
return title
|
||
}
|
||
|
||
/**
|
||
* 使用当前会话的同源模型生成简短中文标题；调用前先脱敏，失败时返回确定性的本地兜底标题。
|
||
* 标题失败不会影响主对话运行。
|
||
*/
|
||
export async function generateAgentTitle(
|
||
conversationText: string,
|
||
modelConfig?: AgentModelConfig,
|
||
): Promise<AgentTitleResult> {
|
||
const textRedaction = createAgentTextRedactionSession()
|
||
const safeConversationText = textRedaction.redactText(String(conversationText || ''))
|
||
const firstUserText = safeConversationText
|
||
.replace(/^(?:用户|User)\s*[:：]\s*/i, '')
|
||
.split(/\n(?:助手|Assistant)\s*[:：]/i)[0]
|
||
const fallback = fallbackTitle(firstUserText)
|
||
const hasApiKey = Boolean(modelConfig?.apiKey) || modelConfig?.provider === 'ollama'
|
||
if (!hasApiKey || !modelConfig?.model || !modelConfig.baseURL) {
|
||
return { title: fallback, generated: false, error: '标题请求缺少可用的同源模型配置' }
|
||
}
|
||
|
||
try {
|
||
// 使用与主智能体运行相同的流式传输。一些自定义 OpenAI 兼容网关会终止非流式请求，
|
||
// 但其流式聊天补全端点仍能正常工作。
|
||
const result = streamText({
|
||
model: createAgentModel(modelConfig),
|
||
system: '你负责给 AI 助手的历史会话命名。输出必须严格符合 <title>标题正文</title>，且标签之外不得有任何内容。标题正文使用 8 到 16 个汉字，简洁、具体。',
|
||
prompt: `概括下面会话的核心主题并命名：\n\n${safeConversationText.slice(0, 1800)}`,
|
||
maxOutputTokens: 64,
|
||
...titleSamplingOptions(modelConfig),
|
||
reasoning: 'none',
|
||
})
|
||
let streamedText = ''
|
||
for await (const delta of result.textStream) streamedText += delta
|
||
const generated = normalizeGeneratedAgentTitle(streamedText)
|
||
if (!generated) {
|
||
const detail = `标题模型未返回正文（finishReason: ${String(await result.finishReason || 'unknown')}）`
|
||
console.warn(`[Agent title] ${modelConfig.provider}/${modelConfig.model}: ${detail}`)
|
||
return { title: fallback, generated: false, error: detail }
|
||
}
|
||
return { title: textRedaction.redactText(generated), generated: true }
|
||
} catch (error) {
|
||
const detail = titleErrorMessage(error)
|
||
console.warn(`[Agent title] ${modelConfig.provider}/${modelConfig.model}: ${detail}`)
|
||
return { title: fallback, generated: false, error: detail }
|
||
}
|
||
}
|