fix: decouple rectification narrative from scoring state
This commit is contained in:
Jesse_Chen
@@ -65,6 +65,7 @@ const labeledYearChoicesPattern = /A\s*[.、:：)]?[\s\S]{0,80}(?:19|20)\d{2}\s*
|
||||
const affirmativeAnswerPattern = /^\s*(?:是(?:的)?|对(?:的)?|没错|正确|确认|就是|嗯+|没问题)\s*[。.!！,，]?\s*$/u;
|
||||
const negativeAnswerPattern = /^\s*(?:不是|不对|错了|并不是|否)\s*[。.!！,，]?\s*$/u;
|
||||
const proposedDateQuestionPattern = /(?:19|20)\d{2}\s*年(?:\s*(?:1[0-2]|0?[1-9])\s*月)?(?:\s*(?:3[01]|[12]\d|0?[1-9])\s*(?:日|号))?[\s\S]{0,30}(?:吗|是否|是不是|确认|对不对|正确)/u;
|
||||
const technicalDiscussionRequestPattern = /(?:校时|生时(?:校正|纠正)|出生时间|候选(?:时间|范围|分钟)|收敛(?:结果|进度)?|技术(?:分析|结果)|分盘|\bD\d{1,3}\b|\b(?:UL|A7|A10|KP)\b)/iu;
|
||||
const domainLabels = {
|
||||
career: "事业",
|
||||
education: "学业",
|
||||
@@ -98,6 +99,18 @@ export const rectificationNarrativeOutputSchema = z.object({
|
||||
}).strict().nullable(),
|
||||
}).strict();
|
||||
|
||||
const authoredFollowUpSchema = z.object({
|
||||
kind: z.enum(["new_event", "event_date", "event_detail"]),
|
||||
// Event ids are server-owned. Ordinary conversation prompts deliberately do
|
||||
// not expose the event ledger, so authored outputs may omit the target id.
|
||||
evidenceId: z.string().uuid().nullable().optional(),
|
||||
answerMode: z.enum(["free_text", "yes_no"]).optional(),
|
||||
proposedDate: z.object({
|
||||
value: z.string().regex(/^\d{4}(?:-(?:0[1-9]|1[0-2])(?:-(?:0[1-9]|[12]\d|3[01]))?)?$/),
|
||||
precision: z.enum(["year", "month", "day"]),
|
||||
}).strict().nullable().optional(),
|
||||
}).strict();
|
||||
|
||||
export const rectificationNarrativeAuthoredOutputSchema = z.object({
|
||||
narrative: z.string().trim().min(1).max(12_000),
|
||||
evidenceRequest: z.object({
|
||||
@@ -106,7 +119,7 @@ export const rectificationNarrativeAuthoredOutputSchema = z.object({
|
||||
domains: z.array(domainSchema).min(1).max(4).optional(),
|
||||
datePrecision: z.enum(["month_preferred", "year_accepted"]),
|
||||
prompt: z.string().trim().min(1).max(1_000),
|
||||
followUp: rectificationFollowUpSchema.default({ kind: "new_event", evidenceId: null }),
|
||||
followUp: authoredFollowUpSchema.default({ kind: "new_event", evidenceId: null }),
|
||||
}).strict().nullable(),
|
||||
}).strict();
|
||||
|
||||
@@ -177,11 +190,39 @@ function groundedEvidenceRequest(
|
||||
return domains.length > 0 ? { ...request, domains } : null;
|
||||
}
|
||||
|
||||
function serverOwnedFollowUp(
|
||||
followUp: z.infer<typeof authoredFollowUpSchema>,
|
||||
context: RectificationNarrativeContext,
|
||||
): RectificationFollowUp {
|
||||
if (followUp.kind === "new_event") return { kind: "new_event", evidenceId: null };
|
||||
const active = context.eventLedger?.filter((item) => item.active) ?? [];
|
||||
const allowedIds = new Set(active.map((item) => item.id));
|
||||
const evidenceId = followUp.evidenceId && allowedIds.has(followUp.evidenceId)
|
||||
? followUp.evidenceId
|
||||
: followUp.kind === "event_date"
|
||||
? context.unresolvedEvidence?.at(-1)?.id ?? active.at(-1)?.id
|
||||
: active.at(-1)?.id;
|
||||
if (!evidenceId) return { kind: "new_event", evidenceId: null };
|
||||
if (followUp.kind === "event_date" && followUp.answerMode === "yes_no" && followUp.proposedDate) {
|
||||
return rectificationFollowUpSchema.parse({
|
||||
kind: followUp.kind,
|
||||
evidenceId,
|
||||
answerMode: "yes_no",
|
||||
proposedDate: followUp.proposedDate,
|
||||
});
|
||||
}
|
||||
return rectificationFollowUpSchema.parse({ kind: followUp.kind, evidenceId });
|
||||
}
|
||||
|
||||
function completeAuthoredOutput(
|
||||
output: z.infer<typeof rectificationNarrativeAuthoredOutputSchema>,
|
||||
packet: RectificationTechnicalPacket,
|
||||
context: RectificationNarrativeContext,
|
||||
): RectificationNarrativeModelOutput {
|
||||
const evidenceRequest = groundedEvidenceRequest(output.evidenceRequest, packet);
|
||||
const groundedRequest = groundedEvidenceRequest(output.evidenceRequest, packet);
|
||||
const evidenceRequest = groundedRequest
|
||||
? { ...groundedRequest, followUp: serverOwnedFollowUp(groundedRequest.followUp, context) }
|
||||
: null;
|
||||
return {
|
||||
...output,
|
||||
evidenceRequest,
|
||||
@@ -200,6 +241,7 @@ function completeAuthoredOutput(
|
||||
function parseModelOutput(
|
||||
text: string,
|
||||
packet: RectificationTechnicalPacket,
|
||||
context: RectificationNarrativeContext,
|
||||
): RectificationNarrativeModelOutput {
|
||||
const normalized = text.trim().replace(/^```(?:json)?\s*/i, "").replace(/\s*```$/, "");
|
||||
const start = normalized.indexOf("{");
|
||||
@@ -208,15 +250,18 @@ function parseModelOutput(
|
||||
const parsed: unknown = JSON.parse(normalized.slice(start, end + 1));
|
||||
const legacy = rectificationNarrativeOutputSchema.safeParse(parsed);
|
||||
if (legacy.success) {
|
||||
const groundedRequest = groundedEvidenceRequest(legacy.data.evidenceRequest, packet);
|
||||
return {
|
||||
...legacy.data,
|
||||
// The next conversational topic is authored by the model, but the
|
||||
// scoring-domain allowlist remains server-owned. If no grounded routing
|
||||
// domain exists, keep the prose and omit only the optional follow-up state.
|
||||
evidenceRequest: groundedEvidenceRequest(legacy.data.evidenceRequest, packet),
|
||||
evidenceRequest: groundedRequest
|
||||
? { ...groundedRequest, followUp: serverOwnedFollowUp(groundedRequest.followUp, context) }
|
||||
: null,
|
||||
};
|
||||
}
|
||||
return completeAuthoredOutput(rectificationNarrativeAuthoredOutputSchema.parse(parsed), packet);
|
||||
return completeAuthoredOutput(rectificationNarrativeAuthoredOutputSchema.parse(parsed), packet, context);
|
||||
}
|
||||
|
||||
function narrativeTimes(value: string): string[] {
|
||||
@@ -296,6 +341,13 @@ export function validateNarrativeAgainstPacket(
|
||||
}
|
||||
}
|
||||
if (output.evidenceRequest) {
|
||||
if (phase === "intermediate" && !/[?？]/u.test(output.narrative)) {
|
||||
issues.push("evidence request is not visibly asked in narrative");
|
||||
}
|
||||
if (phase === "intermediate"
|
||||
&& !normalizedQuestion(output.narrative).includes(normalizedQuestion(output.evidenceRequest.prompt))) {
|
||||
issues.push("evidence request prompt is not present in narrative");
|
||||
}
|
||||
for (const domain of output.evidenceRequest.domains) {
|
||||
if (!allowedDomains.has(domain)) issues.push(`evidence domain ${domain} is not packet-grounded`);
|
||||
}
|
||||
@@ -518,6 +570,24 @@ function promptFor(
|
||||
retryIssues: boundedReceiptIssues(retryIssues),
|
||||
});
|
||||
}
|
||||
if (phase === "intermediate" && !technicalDiscussionRequestPattern.test(context.latestUserText ?? "")) {
|
||||
return JSON.stringify({
|
||||
task: "像正常的人一样接住用户刚才的叙述并自由交谈。不要把自己写成记录员、问卷或校时流程播报器。用户可以继续叙述，也可以自然转到别的话题。",
|
||||
phase,
|
||||
conversation: {
|
||||
recentConversation: context.recentConversation?.slice(-40),
|
||||
latestUserText: context.latestUserText,
|
||||
previousAssistantNarrative: context.previousAssistantNarrative,
|
||||
},
|
||||
rules: [
|
||||
"优先回应用户话里的真实内容、感受、选择或转折，不要复述成档案摘要。",
|
||||
"除非用户主动询问校时进度或技术结果，不要说已记录、先记为、对校时有价值、参与候选时间核对、不会因单一事件确认分钟，也不要主动谈候选范围、分盘、评分、收敛或内部处理。",
|
||||
"不必每轮提问。需要提问时只问自然推进对话真正需要的问题，不要机械补年月、结果或转折。",
|
||||
"只返回 narrative 和 evidenceRequest。没有在 narrative 中逐字提出一个用户可见的明确问题时，evidenceRequest 必须为 null；若提出问题，evidenceRequest.prompt 必须与 narrative 中的问题文字完全一致。followUp 只表达 kind、answerMode 和 proposedDate，不要生成 evidenceId，目标事件由服务器绑定。",
|
||||
],
|
||||
retryIssues: boundedReceiptIssues(retryIssues),
|
||||
});
|
||||
}
|
||||
return JSON.stringify({
|
||||
task: "write_grounded_rectification_narrative",
|
||||
phase,
|
||||
@@ -568,11 +638,12 @@ function promptFor(
|
||||
|
||||
function fallbackNarrative(packet: RectificationTechnicalPacket, phase: RectificationNarrativePhase): string {
|
||||
const candidate = packet.candidate;
|
||||
if (phase === "intermediate") {
|
||||
return "我听到了。你可以顺着这段经历继续说，也可以自然讲下一件想到的事。";
|
||||
}
|
||||
const phaseLine = phase === "final" && candidate.status === "ready_for_confirmation"
|
||||
? "当前证据已形成候选总结，但仍有残余不确定性；只有明确确认后才会替换当前排盘时间。"
|
||||
: phase === "final"
|
||||
? "当前证据只能支持候选范围，系统验证尚未闭环；本次不会替换当前排盘时间，也不再强制追问更多人生事件。"
|
||||
: `我收到了这段叙述。你可以继续讲这段经历，也可以按自己的节奏说下一件想到的事。`;
|
||||
: "当前证据只能支持候选范围，系统验证尚未闭环；本次不会替换当前排盘时间，也不再强制追问更多人生事件。";
|
||||
return [
|
||||
`当前仍在核对 ${candidate.range.startTime}–${candidate.range.endTime} 的候选范围，还不能把其中某一分钟当作确定出生时间。`,
|
||||
phaseLine,
|
||||
@@ -740,7 +811,12 @@ export async function generateRectificationNarrative(input: {
|
||||
issues,
|
||||
), { signal, attempt });
|
||||
const modelId = modelIdSchema.parse(generated.modelId ?? defaultModelId);
|
||||
const output = parseModelOutput(generated.text, input.packet);
|
||||
const parsedOutput = parseModelOutput(generated.text, input.packet, input.context ?? {});
|
||||
const output = input.phase === "intermediate"
|
||||
&& parsedOutput.evidenceRequest
|
||||
&& !/[?？]/u.test(parsedOutput.narrative)
|
||||
? { ...parsedOutput, evidenceRequest: null }
|
||||
: parsedOutput;
|
||||
const validation = validateNarrativeAgainstPacket(
|
||||
output,
|
||||
input.packet,
|
||||
|
||||
Reference in New Issue
Block a user