Stage 2 завершён: problem-first ответы и follow-up continuity - ассистент переведён от entity-heavy логики к problem-first ответам с problem-unit слоем, удержанием контекста в follow-up и очисткой пользовательского ответа от сырых технических ссылок.
This commit is contained in:
@@ -4,7 +4,11 @@ import { afterEach, describe, expect, it, vi } from "vitest";
|
||||
const FLAG_KEYS = [
|
||||
"FEATURE_ASSISTANT_INVESTIGATION_STATE_V1",
|
||||
"FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1",
|
||||
"FEATURE_ASSISTANT_CONTRACTS_V11"
|
||||
"FEATURE_ASSISTANT_CONTRACTS_V11",
|
||||
"FEATURE_ASSISTANT_PROBLEM_UNITS_V1",
|
||||
"FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1",
|
||||
"FEATURE_ASSISTANT_ANSWER_POLICY_V11",
|
||||
"FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1"
|
||||
] as const;
|
||||
|
||||
const ORIGINAL_FLAGS: Record<string, string | undefined> = Object.fromEntries(
|
||||
@@ -26,10 +30,18 @@ async function createAppWithFlags(flags: {
|
||||
state: "0" | "1";
|
||||
binding: "0" | "1";
|
||||
contracts?: "0" | "1";
|
||||
problemUnits?: "0" | "1";
|
||||
continuity?: "0" | "1";
|
||||
answerPolicy?: "0" | "1";
|
||||
problemCentric?: "0" | "1";
|
||||
}) {
|
||||
process.env.FEATURE_ASSISTANT_INVESTIGATION_STATE_V1 = flags.state;
|
||||
process.env.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 = flags.binding;
|
||||
process.env.FEATURE_ASSISTANT_CONTRACTS_V11 = flags.contracts ?? "1";
|
||||
process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = flags.problemUnits ?? "0";
|
||||
process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = flags.continuity ?? "0";
|
||||
process.env.FEATURE_ASSISTANT_ANSWER_POLICY_V11 = flags.answerPolicy ?? "0";
|
||||
process.env.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 = flags.problemCentric ?? "0";
|
||||
vi.resetModules();
|
||||
const { createApp } = await import("../src/server");
|
||||
return createApp();
|
||||
@@ -119,4 +131,111 @@ describe.sequential("assistant follow-up state binding", () => {
|
||||
expect(response.body.debug?.investigation_state_snapshot).toBeNull();
|
||||
expect(response.body.debug?.followup_state_usage).toBeUndefined();
|
||||
});
|
||||
|
||||
it("applies problem continuity hints only when continuity flag is ON and follow-up has no strong new anchors", async () => {
|
||||
const app = await createAppWithFlags({
|
||||
state: "1",
|
||||
binding: "1",
|
||||
problemUnits: "1",
|
||||
continuity: "1"
|
||||
});
|
||||
const sessionId = `asst-wave4-problem-continuity-${Date.now()}`;
|
||||
|
||||
const first = await request(app).post("/api/assistant/message").send({
|
||||
session_id: sessionId,
|
||||
useMock: true,
|
||||
promptVersion: "normalizer_v2_0_2",
|
||||
user_message: "Разложи цепочку документов и оплат по контрагентам за 2020-06, где разрыв механизма закрытия."
|
||||
});
|
||||
expect(first.status).toBe(200);
|
||||
expect(first.body.debug?.investigation_state_snapshot?.problem_unit_state).toBeTruthy();
|
||||
|
||||
const second = await request(app).post("/api/assistant/message").send({
|
||||
session_id: sessionId,
|
||||
useMock: true,
|
||||
promptVersion: "normalizer_v2_0_2",
|
||||
user_message: "И по тому же разрыву добавь уточнение."
|
||||
});
|
||||
|
||||
expect(second.status).toBe(200);
|
||||
expect(second.body.debug?.followup_state_usage?.applied).toBe(true);
|
||||
expect(second.body.debug?.followup_state_usage?.context_patch?.problem_continuity_available).toBe(true);
|
||||
expect(second.body.debug?.followup_state_usage?.context_patch?.problem_continuity_applied).toBe(true);
|
||||
expect(second.body.debug?.followup_state_usage?.context_patch?.strong_new_anchor_detected).toBe(false);
|
||||
});
|
||||
|
||||
it("does not apply follow-up continuity when user gives strong new anchors", async () => {
|
||||
const app = await createAppWithFlags({
|
||||
state: "1",
|
||||
binding: "1",
|
||||
problemUnits: "1",
|
||||
continuity: "1"
|
||||
});
|
||||
const sessionId = `asst-wave4-strong-anchor-${Date.now()}`;
|
||||
|
||||
const first = await request(app).post("/api/assistant/message").send({
|
||||
session_id: sessionId,
|
||||
useMock: true,
|
||||
promptVersion: "normalizer_v2_0_2",
|
||||
user_message: "Разбери хвосты по счету 60 за 2020-06 и покажи проблемные цепочки."
|
||||
});
|
||||
expect(first.status).toBe(200);
|
||||
expect(first.body.debug?.investigation_state_snapshot?.turn_index).toBe(1);
|
||||
|
||||
const second = await request(app).post("/api/assistant/message").send({
|
||||
session_id: sessionId,
|
||||
useMock: true,
|
||||
promptVersion: "normalizer_v2_0_2",
|
||||
user_message: "И отдельно по счету 97 за 2020-07."
|
||||
});
|
||||
|
||||
expect(second.status).toBe(200);
|
||||
expect(second.body.debug?.followup_state_usage).toBeUndefined();
|
||||
expect(second.body.debug?.investigation_state_snapshot?.turn_index).toBe(2);
|
||||
});
|
||||
|
||||
it("keeps UTF-8 follow-up period refinement in-scope with soft continuity hints", async () => {
|
||||
const app = await createAppWithFlags({
|
||||
state: "1",
|
||||
binding: "1",
|
||||
problemUnits: "1",
|
||||
continuity: "1",
|
||||
answerPolicy: "1",
|
||||
problemCentric: "1"
|
||||
});
|
||||
|
||||
const first = await request(app).post("/api/assistant/message").send({
|
||||
useMock: true,
|
||||
promptVersion: "normalizer_v2_0_2",
|
||||
user_message: "Посмотри, пожалуйста, где по поставщикам сейчас хвосты уже похожи именно на проблему, а не просто на шум."
|
||||
});
|
||||
|
||||
expect(first.status).toBe(200);
|
||||
expect(first.body.reply_type).not.toBe("out_of_scope");
|
||||
|
||||
const second = await request(app).post("/api/assistant/message").send({
|
||||
session_id: first.body.session_id,
|
||||
useMock: true,
|
||||
promptVersion: "normalizer_v2_0_2",
|
||||
user_message: "А если только за июнь 2020 смотреть, по кому это сильнее всего видно и что из этого реально может мешать закрытию?"
|
||||
});
|
||||
|
||||
expect(second.status).toBe(200);
|
||||
expect(second.body.reply_type).not.toBe("out_of_scope");
|
||||
expect(second.body.debug?.followup_state_usage?.applied).toBe(true);
|
||||
expect(second.body.debug?.followup_state_usage?.context_patch?.problem_continuity_applied).toBe(true);
|
||||
expect(second.body.debug?.followup_state_usage?.context_patch?.strong_new_anchor_detected).toBe(false);
|
||||
expect(
|
||||
(second.body.debug?.routes ?? []).some((item: { route?: string }) => item.route && item.route !== "no_route")
|
||||
).toBe(true);
|
||||
|
||||
const third = await request(app).post("/api/assistant/message").send({
|
||||
useMock: true,
|
||||
promptVersion: "normalizer_v2_0_2",
|
||||
user_message: "Проверь, пожалуйста, по 60-му счёту за июнь 2020, где есть самый явный проблемный участок по расчётам с поставщиками."
|
||||
});
|
||||
|
||||
expect(third.status).toBe(200);
|
||||
expect(third.body.reply_type).not.toBe("out_of_scope");
|
||||
});
|
||||
});
|
||||
|
||||
Reference in New Issue
Block a user