fix(bom-costing): bound GPT-5 reasoning time
This commit is contained in:
parent
608da97538
commit
b3fd61dadb
@ -97,6 +97,7 @@ BOM_COSTING_MARKET_TIMEOUT_MS=60000
|
||||
BOM_COSTING_AI_TIMEOUT_MS=300000
|
||||
BOM_COSTING_AI_MAX_IMAGES=20
|
||||
BOM_COSTING_AI_MAX_TOKENS=24000
|
||||
BOM_COSTING_AI_REASONING_EFFORT=low
|
||||
BOM_COSTING_MODELING_PROVIDER=
|
||||
|
||||
# 素材库白底图回传
|
||||
|
||||
@ -228,7 +228,7 @@ describe('BOM 多模态分析器', () => {
|
||||
name: '多模态模型',
|
||||
baseUrl: 'https://ai.example.com/v1',
|
||||
apiKey: 'test-ai-key',
|
||||
model: 'vision-model',
|
||||
model: 'gpt-5.6-luna',
|
||||
extra: null,
|
||||
isActive: true,
|
||||
sortOrder: 0,
|
||||
@ -710,6 +710,9 @@ describe('BOM 多模态分析器', () => {
|
||||
message.content.some((content) => content.type === 'image_url'),
|
||||
),
|
||||
).toBe(false)
|
||||
expect(requestBody).toMatchObject({
|
||||
reasoning_effort: 'low',
|
||||
})
|
||||
expect(output.trace).toMatchObject({
|
||||
imageCount: 4,
|
||||
factBatchCount: 0,
|
||||
|
||||
@ -28,6 +28,13 @@ export interface BomCostingMultimodalCallResult {
|
||||
|
||||
export type BomCostingMultimodalOperation = 'IMAGE_FACTS' | 'SYNTHESIS'
|
||||
|
||||
const BOM_COSTING_REASONING_EFFORTS = new Set([
|
||||
'minimal',
|
||||
'low',
|
||||
'medium',
|
||||
'high',
|
||||
])
|
||||
|
||||
export function readBomCostingBoundedInteger(
|
||||
value: string | undefined,
|
||||
fallback: number,
|
||||
@ -44,6 +51,20 @@ function getAiServiceType() {
|
||||
return process.env.BOM_COSTING_AI_SERVICE_TYPE?.trim() || 'prompt-optimizer'
|
||||
}
|
||||
|
||||
function getBomCostingReasoningEffort(model: string) {
|
||||
const configured = process.env.BOM_COSTING_AI_REASONING_EFFORT
|
||||
?.trim()
|
||||
.toLowerCase()
|
||||
if (configured && BOM_COSTING_REASONING_EFFORTS.has(configured)) {
|
||||
return configured
|
||||
}
|
||||
|
||||
// The configured GPT-5-compatible gateway terminates long, non-streaming
|
||||
// reasoning around two minutes. BOM synthesis is structured extraction, not
|
||||
// an open-ended reasoning task, so keep the default response bounded.
|
||||
return /^gpt-5(?:[.-]|$)/i.test(model) ? 'low' : undefined
|
||||
}
|
||||
|
||||
export async function getBomCostingMultimodalProviderConfig():
|
||||
Promise<BomCostingMultimodalProviderConfig | null> {
|
||||
const active = await providerConfigService.getActive(getAiServiceType())
|
||||
@ -280,6 +301,7 @@ export async function callBomCostingMultimodalModel(input: {
|
||||
timeoutMs: number
|
||||
}): Promise<BomCostingMultimodalCallResult> {
|
||||
const endpoint = `${input.config.baseUrl.replace(/\/+$/, '')}/chat/completions`
|
||||
const reasoningEffort = getBomCostingReasoningEffort(input.config.model)
|
||||
const startedAt = Date.now()
|
||||
let response: Response
|
||||
try {
|
||||
@ -299,6 +321,7 @@ export async function callBomCostingMultimodalModel(input: {
|
||||
temperature: 0.1,
|
||||
max_tokens: input.maxTokens,
|
||||
stream: true,
|
||||
...(reasoningEffort ? { reasoning_effort: reasoningEffort } : {}),
|
||||
}),
|
||||
cache: 'no-store',
|
||||
signal: AbortSignal.timeout(input.timeoutMs),
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user