Этап 4 / Волна 10: корректировка settlement-кейса — доменная фиксация синтеза, честное покрытие, удержание фокуса / Этап 4 / Волна 11: бизнес-якоря, доменное заземление и устранение утечки дебага

This commit is contained in:
2026-03-28 02:17:19 +03:00
parent 914843a8ba
commit a06e575be4
367 changed files with 432257 additions and 3627 deletions
+28 -5
View File
@@ -11,6 +11,14 @@ function toBooleanFlag(value: string | undefined, defaultValue: boolean): boolea
return !(lowered === "0" || lowered === "false" || lowered === "off" || lowered === "no");
}
function toNumberFlag(value: string | undefined, defaultValue: number): number {
if (!value || value.trim() === "") {
return defaultValue;
}
const parsed = Number(value);
return Number.isFinite(parsed) ? parsed : defaultValue;
}
export const PORT = Number(process.env.PORT ?? 8787);
export const TIMEZONE = process.env.TZ_FALLBACK ?? "Europe/Moscow";
export const DEFAULT_OPENAI_BASE_URL = process.env.OPENAI_BASE_URL ?? "https://api.openai.com/v1";
@@ -45,7 +53,7 @@ export const FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1 = toBooleanFlag(
);
export const FEATURE_ASSISTANT_ANSWER_POLICY_V11 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_ANSWER_POLICY_V11,
false
true
);
export const FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1,
@@ -53,15 +61,15 @@ export const FEATURE_ASSISTANT_ACCOUNTANT_EVAL_V1 = toBooleanFlag(
);
export const FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1,
false
true
);
export const FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1,
false
true
);
export const FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1,
false
true
);
export const FEATURE_ASSISTANT_STAGE2_EVAL_V1 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_STAGE2_EVAL_V1,
@@ -69,12 +77,27 @@ export const FEATURE_ASSISTANT_STAGE2_EVAL_V1 = toBooleanFlag(
);
export const FEATURE_ASSISTANT_LIFECYCLE_RUNTIME_V1 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_LIFECYCLE_RUNTIME_V1,
false
true
);
export const FEATURE_ASSISTANT_LIFECYCLE_ANSWER_V1 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_LIFECYCLE_ANSWER_V1,
true
);
export const FEATURE_ASSISTANT_GRAPH_RUNTIME_V1 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_GRAPH_RUNTIME_V1,
true
);
export const FEATURE_ASSISTANT_MCP_RUNTIME_V1 = toBooleanFlag(
process.env.FEATURE_ASSISTANT_MCP_RUNTIME_V1,
false
);
export const ASSISTANT_MCP_PROXY_URL = (process.env.ASSISTANT_MCP_PROXY_URL ?? "http://127.0.0.1:6003").replace(
/\/+$/,
""
);
export const ASSISTANT_MCP_CHANNEL = process.env.ASSISTANT_MCP_CHANNEL ?? "default";
export const ASSISTANT_MCP_TIMEOUT_MS = toNumberFlag(process.env.ASSISTANT_MCP_TIMEOUT_MS, 1200);
export const ASSISTANT_MCP_LIVE_LIMIT = Math.max(1, Math.trunc(toNumberFlag(process.env.ASSISTANT_MCP_LIVE_LIMIT, 24)));
export const DATA_DIR = process.env.DATA_DIR ?? path.resolve(MODULE_ROOT, "data");
export const TRACES_DIR = path.resolve(DATA_DIR, "traces");
@@ -0,0 +1,265 @@
import {
P0_DEFAULT_ACCEPTANCE_THRESHOLDS,
P0_DEFAULT_QUALITY_GAP_THRESHOLDS,
type P0AcceptanceThresholds,
type P0EvalMetricVector,
type P0QualityGapMetricVector,
type P0QualityGapThresholds
} from "./p0_metric_definitions";
export type P0AcceptanceVerdict = "P0_ACCEPTED" | "P0_ACCEPTED_WITH_LIMITATIONS" | "P0_NOT_ACCEPTED";
export type P0BaselineStabilityVerdict = "P0_BASELINE_STABLE" | "P0_BASELINE_STABLE_WITH_OPEN_QUALITY_GAPS";
interface MetricCheck {
metric: keyof P0EvalMetricVector;
value: number;
threshold: number;
comparator: ">=" | "<=";
passed: boolean;
severity: "blocking" | "quality";
}
interface QualityGapMetricCheck {
metric: keyof P0QualityGapMetricVector;
value: number;
threshold: number;
comparator: ">=" | "<=";
passed: boolean;
}
export interface P0AcceptanceResult {
verdict: P0AcceptanceVerdict;
passed_all: boolean;
blocking_failures: MetricCheck[];
quality_failures: MetricCheck[];
metric_checks: MetricCheck[];
rationale: string[];
}
export interface P0BaselineStabilityResult {
verdict: P0BaselineStabilityVerdict;
baseline_integrity_passed: boolean;
quality_gaps_open: boolean;
blocking_regressions: MetricCheck[];
legacy_quality_failures: MetricCheck[];
quality_gap_failures: QualityGapMetricCheck[];
quality_gap_checks: QualityGapMetricCheck[];
rationale: string[];
}
function buildMetricChecks(input: {
metrics: P0EvalMetricVector;
thresholds: P0AcceptanceThresholds;
}): MetricCheck[] {
const { metrics, thresholds } = input;
return [
{
metric: "problem_first_answer_rate",
value: metrics.problem_first_answer_rate,
threshold: thresholds.problem_first_answer_rate_min,
comparator: ">=",
passed: metrics.problem_first_answer_rate >= thresholds.problem_first_answer_rate_min,
severity: "blocking"
},
{
metric: "entity_leakage_rate",
value: metrics.entity_leakage_rate,
threshold: thresholds.entity_leakage_rate_max,
comparator: "<=",
passed: metrics.entity_leakage_rate <= thresholds.entity_leakage_rate_max,
severity: "blocking"
},
{
metric: "route_correctness_rate",
value: metrics.route_correctness_rate,
threshold: thresholds.route_correctness_rate_min,
comparator: ">=",
passed: metrics.route_correctness_rate >= thresholds.route_correctness_rate_min,
severity: "blocking"
},
{
metric: "domain_purity_rate",
value: metrics.domain_purity_rate,
threshold: thresholds.domain_purity_rate_min,
comparator: ">=",
passed: metrics.domain_purity_rate >= thresholds.domain_purity_rate_min,
severity: "blocking"
},
{
metric: "mechanism_coherence_score",
value: metrics.mechanism_coherence_score,
threshold: thresholds.mechanism_coherence_score_min,
comparator: ">=",
passed: metrics.mechanism_coherence_score >= thresholds.mechanism_coherence_score_min,
severity: "quality"
},
{
metric: "accountant_actionability_score",
value: metrics.accountant_actionability_score,
threshold: thresholds.accountant_actionability_score_min,
comparator: ">=",
passed: metrics.accountant_actionability_score >= thresholds.accountant_actionability_score_min,
severity: "quality"
},
{
metric: "limitation_honesty_rate",
value: metrics.limitation_honesty_rate,
threshold: thresholds.limitation_honesty_rate_min,
comparator: ">=",
passed: metrics.limitation_honesty_rate >= thresholds.limitation_honesty_rate_min,
severity: "quality"
},
{
metric: "top_problem_unit_match_rate",
value: metrics.top_problem_unit_match_rate,
threshold: thresholds.top_problem_unit_match_rate_min,
comparator: ">=",
passed: metrics.top_problem_unit_match_rate >= thresholds.top_problem_unit_match_rate_min,
severity: "quality"
}
];
}
function toFailureLine(check: MetricCheck): string {
return `${check.metric}: ${check.value.toFixed(2)} ${check.comparator} ${check.threshold.toFixed(2)} (failed)`;
}
function toQualityGapFailureLine(check: QualityGapMetricCheck): string {
return `${check.metric}: ${check.value.toFixed(2)} ${check.comparator} ${check.threshold.toFixed(2)} (failed)`;
}
function buildQualityGapChecks(input: {
metrics: P0QualityGapMetricVector;
thresholds: P0QualityGapThresholds;
}): QualityGapMetricCheck[] {
const { metrics, thresholds } = input;
return [
{
metric: "generic_explanation_rate",
value: metrics.generic_explanation_rate,
threshold: thresholds.generic_explanation_rate_max,
comparator: "<=",
passed: metrics.generic_explanation_rate <= thresholds.generic_explanation_rate_max
},
{
metric: "false_confidence_rate",
value: metrics.false_confidence_rate,
threshold: thresholds.false_confidence_rate_max,
comparator: "<=",
passed: metrics.false_confidence_rate <= thresholds.false_confidence_rate_max
},
{
metric: "mechanism_specificity_score",
value: metrics.mechanism_specificity_score,
threshold: thresholds.mechanism_specificity_score_min,
comparator: ">=",
passed: metrics.mechanism_specificity_score >= thresholds.mechanism_specificity_score_min
},
{
metric: "followup_context_retention_score",
value: metrics.followup_context_retention_score,
threshold: thresholds.followup_context_retention_score_min,
comparator: ">=",
passed: metrics.followup_context_retention_score >= thresholds.followup_context_retention_score_min
}
];
}
export function evaluateP0AcceptanceGate(input: {
metrics: P0EvalMetricVector;
thresholds?: Partial<P0AcceptanceThresholds>;
}): P0AcceptanceResult {
const thresholds: P0AcceptanceThresholds = {
...P0_DEFAULT_ACCEPTANCE_THRESHOLDS,
...(input.thresholds ?? {})
};
const checks = buildMetricChecks({
metrics: input.metrics,
thresholds
});
const blockingFailures = checks.filter((item) => item.severity === "blocking" && !item.passed);
const qualityFailures = checks.filter((item) => item.severity === "quality" && !item.passed);
let verdict: P0AcceptanceVerdict = "P0_NOT_ACCEPTED";
if (blockingFailures.length === 0 && qualityFailures.length === 0) {
verdict = "P0_ACCEPTED";
} else if (blockingFailures.length === 0) {
verdict = "P0_ACCEPTED_WITH_LIMITATIONS";
}
const rationale: string[] = [];
if (blockingFailures.length > 0) {
rationale.push("Blocking metrics failed:");
rationale.push(...blockingFailures.map((item) => `- ${toFailureLine(item)}`));
}
if (qualityFailures.length > 0) {
rationale.push("Quality metrics failed:");
rationale.push(...qualityFailures.map((item) => `- ${toFailureLine(item)}`));
}
if (rationale.length === 0) {
rationale.push("All acceptance thresholds are satisfied.");
}
return {
verdict,
passed_all: checks.every((item) => item.passed),
blocking_failures: blockingFailures,
quality_failures: qualityFailures,
metric_checks: checks,
rationale
};
}
export function evaluateP0BaselineStabilityGate(input: {
metrics: P0EvalMetricVector;
qualityGapMetrics: P0QualityGapMetricVector;
acceptanceThresholds?: Partial<P0AcceptanceThresholds>;
qualityGapThresholds?: Partial<P0QualityGapThresholds>;
}): P0BaselineStabilityResult {
const acceptance = evaluateP0AcceptanceGate({
metrics: input.metrics,
thresholds: input.acceptanceThresholds
});
const qualityGapThresholds: P0QualityGapThresholds = {
...P0_DEFAULT_QUALITY_GAP_THRESHOLDS,
...(input.qualityGapThresholds ?? {})
};
const qualityGapChecks = buildQualityGapChecks({
metrics: input.qualityGapMetrics,
thresholds: qualityGapThresholds
});
const qualityGapFailures = qualityGapChecks.filter((item) => !item.passed);
const baselineIntegrityPassed = acceptance.blocking_failures.length === 0;
const qualityGapsOpen = acceptance.quality_failures.length > 0 || qualityGapFailures.length > 0;
const verdict: P0BaselineStabilityVerdict =
baselineIntegrityPassed && !qualityGapsOpen ? "P0_BASELINE_STABLE" : "P0_BASELINE_STABLE_WITH_OPEN_QUALITY_GAPS";
const rationale: string[] = [];
if (!baselineIntegrityPassed) {
rationale.push("Blocking baseline regressions detected:");
rationale.push(...acceptance.blocking_failures.map((item) => `- ${toFailureLine(item)}`));
}
if (acceptance.quality_failures.length > 0) {
rationale.push("Legacy quality thresholds still below full acceptance:");
rationale.push(...acceptance.quality_failures.map((item) => `- ${toFailureLine(item)}`));
}
if (qualityGapFailures.length > 0) {
rationale.push("Wave 9 quality-gap thresholds still open:");
rationale.push(...qualityGapFailures.map((item) => `- ${toQualityGapFailureLine(item)}`));
}
if (rationale.length === 0) {
rationale.push("Baseline is stable and Wave 9 quality-gap thresholds are satisfied.");
}
return {
verdict,
baseline_integrity_passed: baselineIntegrityPassed,
quality_gaps_open: qualityGapsOpen,
blocking_regressions: acceptance.blocking_failures,
legacy_quality_failures: acceptance.quality_failures,
quality_gap_failures: qualityGapFailures,
quality_gap_checks: qualityGapChecks,
rationale
};
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,425 @@
import type { ProblemUnitType } from "../types/stage2ProblemUnits";
export type P0DomainId = "settlements_60_62" | "vat_document_register_book" | "month_close_costs_20_44";
export type P0QueryClass =
| "symptom_first"
| "lifecycle_first"
| "causal_query"
| "mixed_ambiguity"
| "followup_investigation"
| "noisy_input"
| "translit_noisy"
| "multi_intent";
export type P0RouteExpected = "hybrid_store_plus_live" | "store_feature_risk" | "store_canonical" | "batch_refresh_then_store" | "live_mcp_drilldown" | "no_route";
export interface P0ExpectedAnnotations {
domain_expected: P0DomainId;
route_expected: P0RouteExpected | P0RouteExpected[];
problem_unit_type_expected: ProblemUnitType;
mechanism_class_expected: string;
evidence_expectation: string;
must_show_limitation: boolean;
must_not_leak_internal: boolean;
first_check_action_expected: string;
}
export interface P0EvalCase {
case_id: string;
domain: P0DomainId;
user_query_raw: string;
query_class: P0QueryClass;
expected_problem_unit_type: ProblemUnitType;
expected_mechanism_summary: string;
expected_first_check: string;
must_include_limitations: boolean;
forbidden_leakage_tokens: string[];
followup_seed_query?: string;
followup_expected_context_tokens?: string[];
notes: string;
expected_annotations: P0ExpectedAnnotations;
}
export interface P0EvalCorpusFile {
suite_id: string;
suite_version: string;
schema_version: string;
scenario_count: number;
case_ids: string[];
cases: P0EvalCase[];
}
export interface P0EvalMetricVector {
problem_first_answer_rate: number;
mechanism_coherence_score: number;
entity_leakage_rate: number;
accountant_actionability_score: number;
route_correctness_rate: number;
domain_purity_rate: number;
limitation_honesty_rate: number;
top_problem_unit_match_rate: number;
}
export type P0MetricName = keyof P0EvalMetricVector;
export interface P0QualityGapMetricVector {
generic_explanation_rate: number;
false_confidence_rate: number;
mechanism_specificity_score: number;
followup_context_retention_score: number;
}
export type P0QualityGapMetricName = keyof P0QualityGapMetricVector;
export interface P0MetricDefinition {
description: string;
unit: "rate" | "score";
direction: "higher_is_better" | "lower_is_better";
}
export interface P0QualityGapMetricDefinition {
description: string;
unit: "rate" | "score";
direction: "higher_is_better" | "lower_is_better";
}
export const P0_METRIC_DEFINITIONS: Record<P0MetricName, P0MetricDefinition> = {
problem_first_answer_rate: {
description: "Share of cases where answer is problem-first and avoids entity-dump framing.",
unit: "rate",
direction: "higher_is_better"
},
mechanism_coherence_score: {
description: "Average 0..5 score for mechanism clarity and alignment with expected defect class.",
unit: "score",
direction: "higher_is_better"
},
entity_leakage_rate: {
description: "Share of cases with internal/technical leakage in user-facing answer.",
unit: "rate",
direction: "lower_is_better"
},
accountant_actionability_score: {
description: "Average 0..5 score for first-check concreteness and practical accounting actionability.",
unit: "score",
direction: "higher_is_better"
},
route_correctness_rate: {
description: "Share of cases routed to expected path class for query type.",
unit: "rate",
direction: "higher_is_better"
},
domain_purity_rate: {
description: "Share of cases where top-3 retrieval stays inside expected P0 domain boundary.",
unit: "rate",
direction: "higher_is_better"
},
limitation_honesty_rate: {
description: "Share of cases where mandatory limitations are explicit when evidence is incomplete.",
unit: "rate",
direction: "higher_is_better"
},
top_problem_unit_match_rate: {
description: "Share of cases where top problem unit type matches expected annotation.",
unit: "rate",
direction: "higher_is_better"
}
};
export const P0_QUALITY_GAP_METRIC_DEFINITIONS: Record<P0QualityGapMetricName, P0QualityGapMetricDefinition> = {
generic_explanation_rate: {
description: "Share of cases where direct answer stays generic and mechanism/actionability specificity is insufficient.",
unit: "rate",
direction: "lower_is_better"
},
false_confidence_rate: {
description: "Share of cases with overconfident framing not supported by evidence/limitation state.",
unit: "rate",
direction: "lower_is_better"
},
mechanism_specificity_score: {
description: "Average 0..5 score of mechanism-level specificity in accountant-facing answer.",
unit: "score",
direction: "higher_is_better"
},
followup_context_retention_score: {
description: "Rate of correct context retention in follow-up investigation cases.",
unit: "rate",
direction: "higher_is_better"
}
};
export const P0_DEFAULT_FORBIDDEN_LEAKAGE_TOKENS = [
"graph_",
"domain_scope",
"relation_patterns",
"account_scope",
"semantic_profile",
"route",
"profile",
"store_canonical",
"hybrid_store_plus_live",
"deterministic_v2",
"candidate_evidence",
"problem_unit",
"raw_entities",
"graph_domain_scope"
];
export interface P0AcceptanceThresholds {
problem_first_answer_rate_min: number;
mechanism_coherence_score_min: number;
entity_leakage_rate_max: number;
accountant_actionability_score_min: number;
route_correctness_rate_min: number;
domain_purity_rate_min: number;
limitation_honesty_rate_min: number;
top_problem_unit_match_rate_min: number;
}
export interface P0QualityGapThresholds {
generic_explanation_rate_max: number;
false_confidence_rate_max: number;
mechanism_specificity_score_min: number;
followup_context_retention_score_min: number;
}
export const P0_DEFAULT_ACCEPTANCE_THRESHOLDS: P0AcceptanceThresholds = {
problem_first_answer_rate_min: 0.8,
mechanism_coherence_score_min: 3.5,
entity_leakage_rate_max: 0.08,
accountant_actionability_score_min: 3.5,
route_correctness_rate_min: 0.85,
domain_purity_rate_min: 0.9,
limitation_honesty_rate_min: 0.9,
top_problem_unit_match_rate_min: 0.65
};
export const P0_DEFAULT_QUALITY_GAP_THRESHOLDS: P0QualityGapThresholds = {
generic_explanation_rate_max: 0.2,
false_confidence_rate_max: 0.12,
mechanism_specificity_score_min: 3.0,
followup_context_retention_score_min: 0.75
};
function isProblemUnitType(value: unknown): value is ProblemUnitType {
return (
value === "document_conflict" ||
value === "broken_chain_segment" ||
value === "lifecycle_anomaly_node" ||
value === "unresolved_settlement_cluster" ||
value === "period_risk_cluster" ||
value === "cross_branch_inconsistency_cluster"
);
}
function isP0DomainId(value: unknown): value is P0DomainId {
return value === "settlements_60_62" || value === "vat_document_register_book" || value === "month_close_costs_20_44";
}
function isP0QueryClass(value: unknown): value is P0QueryClass {
return (
value === "symptom_first" ||
value === "lifecycle_first" ||
value === "causal_query" ||
value === "mixed_ambiguity" ||
value === "followup_investigation" ||
value === "noisy_input" ||
value === "translit_noisy" ||
value === "multi_intent"
);
}
function isRouteExpected(value: unknown): value is P0RouteExpected {
return (
value === "hybrid_store_plus_live" ||
value === "store_feature_risk" ||
value === "store_canonical" ||
value === "batch_refresh_then_store" ||
value === "live_mcp_drilldown" ||
value === "no_route"
);
}
function assertString(value: unknown, fieldPath: string): string {
if (typeof value !== "string" || !value.trim()) {
throw new Error(`Invalid corpus field '${fieldPath}': non-empty string is required.`);
}
return value.trim();
}
function assertBoolean(value: unknown, fieldPath: string): boolean {
if (typeof value !== "boolean") {
throw new Error(`Invalid corpus field '${fieldPath}': boolean is required.`);
}
return value;
}
function assertStringArray(value: unknown, fieldPath: string): string[] {
if (!Array.isArray(value)) {
throw new Error(`Invalid corpus field '${fieldPath}': string[] is required.`);
}
const output = value.map((item, index) => assertString(item, `${fieldPath}[${index}]`));
return output;
}
function assertRouteExpected(value: unknown, fieldPath: string): P0RouteExpected | P0RouteExpected[] {
if (Array.isArray(value)) {
const routes = value.map((item, index) => {
if (!isRouteExpected(item)) {
throw new Error(`Invalid corpus field '${fieldPath}[${index}]': unsupported route '${String(item)}'.`);
}
return item;
});
if (routes.length === 0) {
throw new Error(`Invalid corpus field '${fieldPath}': route list must not be empty.`);
}
return routes;
}
if (!isRouteExpected(value)) {
throw new Error(`Invalid corpus field '${fieldPath}': unsupported route '${String(value)}'.`);
}
return value;
}
function parseCase(raw: unknown, index: number): P0EvalCase {
if (!raw || typeof raw !== "object") {
throw new Error(`Invalid corpus case at index ${index}: object is required.`);
}
const row = raw as Record<string, unknown>;
const expected = row.expected_annotations;
if (!expected || typeof expected !== "object") {
throw new Error(`Invalid corpus case '${String(row.case_id ?? index)}': expected_annotations is required.`);
}
const annotations = expected as Record<string, unknown>;
const domain = row.domain;
if (!isP0DomainId(domain)) {
throw new Error(`Invalid corpus case '${String(row.case_id ?? index)}': unsupported domain '${String(domain)}'.`);
}
const queryClass = row.query_class;
if (!isP0QueryClass(queryClass)) {
throw new Error(`Invalid corpus case '${String(row.case_id ?? index)}': unsupported query_class '${String(queryClass)}'.`);
}
const expectedProblemType = row.expected_problem_unit_type;
if (!isProblemUnitType(expectedProblemType)) {
throw new Error(
`Invalid corpus case '${String(row.case_id ?? index)}': expected_problem_unit_type '${String(expectedProblemType)}' is not supported.`
);
}
const annotationProblemType = annotations.problem_unit_type_expected;
if (!isProblemUnitType(annotationProblemType)) {
throw new Error(
`Invalid corpus case '${String(row.case_id ?? index)}': expected_annotations.problem_unit_type_expected '${String(annotationProblemType)}' is not supported.`
);
}
const annotationDomain = annotations.domain_expected;
if (!isP0DomainId(annotationDomain)) {
throw new Error(
`Invalid corpus case '${String(row.case_id ?? index)}': expected_annotations.domain_expected '${String(annotationDomain)}' is not supported.`
);
}
const followupSeedQuery =
typeof row.followup_seed_query === "string" && row.followup_seed_query.trim() ? row.followup_seed_query.trim() : undefined;
const followupExpectedContextTokens = Array.isArray(row.followup_expected_context_tokens)
? assertStringArray(row.followup_expected_context_tokens, `cases[${index}].followup_expected_context_tokens`)
: undefined;
if (queryClass === "followup_investigation" && !followupSeedQuery) {
throw new Error(
`Invalid corpus case '${String(row.case_id ?? index)}': followup_investigation requires followup_seed_query.`
);
}
return {
case_id: assertString(row.case_id, `cases[${index}].case_id`),
domain,
user_query_raw: assertString(row.user_query_raw, `cases[${index}].user_query_raw`),
query_class: queryClass,
expected_problem_unit_type: expectedProblemType,
expected_mechanism_summary: assertString(row.expected_mechanism_summary, `cases[${index}].expected_mechanism_summary`),
expected_first_check: assertString(row.expected_first_check, `cases[${index}].expected_first_check`),
must_include_limitations: assertBoolean(row.must_include_limitations, `cases[${index}].must_include_limitations`),
forbidden_leakage_tokens: assertStringArray(row.forbidden_leakage_tokens, `cases[${index}].forbidden_leakage_tokens`),
followup_seed_query: followupSeedQuery,
followup_expected_context_tokens: followupExpectedContextTokens,
notes: assertString(row.notes, `cases[${index}].notes`),
expected_annotations: {
domain_expected: annotationDomain,
route_expected: assertRouteExpected(annotations.route_expected, `cases[${index}].expected_annotations.route_expected`),
problem_unit_type_expected: annotationProblemType,
mechanism_class_expected: assertString(
annotations.mechanism_class_expected,
`cases[${index}].expected_annotations.mechanism_class_expected`
),
evidence_expectation: assertString(annotations.evidence_expectation, `cases[${index}].expected_annotations.evidence_expectation`),
must_show_limitation: assertBoolean(annotations.must_show_limitation, `cases[${index}].expected_annotations.must_show_limitation`),
must_not_leak_internal: assertBoolean(
annotations.must_not_leak_internal,
`cases[${index}].expected_annotations.must_not_leak_internal`
),
first_check_action_expected: assertString(
annotations.first_check_action_expected,
`cases[${index}].expected_annotations.first_check_action_expected`
)
}
};
}
export function validateP0EvalCorpus(input: unknown): P0EvalCorpusFile {
if (!input || typeof input !== "object") {
throw new Error("Invalid P0 corpus: root object is required.");
}
const parsed = input as Record<string, unknown>;
const casesRaw = parsed.cases;
if (!Array.isArray(casesRaw)) {
throw new Error("Invalid P0 corpus: cases[] is required.");
}
const cases = casesRaw.map((item, index) => parseCase(item, index));
if (cases.length < 30 || cases.length > 150) {
throw new Error(`Invalid P0 corpus: expected 30..150 cases, got ${cases.length}.`);
}
const caseIds = assertStringArray(parsed.case_ids, "case_ids");
const declaredIds = [...caseIds].sort();
const actualIds = cases.map((item) => item.case_id).sort();
const idsMatch = declaredIds.length === actualIds.length && declaredIds.every((value, index) => value === actualIds[index]);
if (!idsMatch) {
throw new Error("Invalid P0 corpus: case_ids[] does not match cases[].");
}
const scenarioCount = Number(parsed.scenario_count ?? 0);
if (!Number.isFinite(scenarioCount) || scenarioCount !== cases.length) {
throw new Error("Invalid P0 corpus: scenario_count must match cases.length.");
}
const perDomain = cases.reduce<Record<P0DomainId, number>>(
(acc, item) => {
acc[item.domain] += 1;
return acc;
},
{
settlements_60_62: 0,
vat_document_register_book: 0,
month_close_costs_20_44: 0
}
);
for (const [domain, count] of Object.entries(perDomain)) {
if (count < 10) {
throw new Error(`Invalid P0 corpus: domain '${domain}' must contain at least 10 cases, got ${count}.`);
}
}
return {
suite_id: assertString(parsed.suite_id, "suite_id"),
suite_version: assertString(parsed.suite_version, "suite_version"),
schema_version: assertString(parsed.schema_version, "schema_version"),
scenario_count: scenarioCount,
case_ids: caseIds,
cases
};
}
export function normalizeExpectedRoutes(value: P0RouteExpected | P0RouteExpected[]): P0RouteExpected[] {
return Array.isArray(value) ? value : [value];
}
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
@@ -214,48 +214,56 @@ export class AssistantSessionLogger {
constructor(private readonly rootDir: string = ASSISTANT_SESSIONS_DIR) {}
public persistSession(session: AssistantSessionState): void {
ensureDir(this.rootDir);
const filePath = path.resolve(this.rootDir, `${session.session_id}.json`);
try {
ensureDir(this.rootDir);
const filePath = path.resolve(this.rootDir, `${session.session_id}.json`);
const startedAt = session.items[0]?.created_at ?? session.updated_at;
const userMessages = session.items.filter((item) => item.role === "user").length;
const assistantMessages = session.items.filter((item) => item.role === "assistant").length;
const assistantItems = session.items.filter((item) => item.role === "assistant");
const lastAssistant = assistantItems.length > 0 ? assistantItems[assistantItems.length - 1] : null;
const startedAt = session.items[0]?.created_at ?? session.updated_at;
const userMessages = session.items.filter((item) => item.role === "user").length;
const assistantMessages = session.items.filter((item) => item.role === "assistant").length;
const assistantItems = session.items.filter((item) => item.role === "assistant");
const lastAssistant = assistantItems.length > 0 ? assistantItems[assistantItems.length - 1] : null;
const traceIds = unique(session.items.map((item) => item.trace_id));
const replyTypes = Array.from(
new Set(
session.items
.map((item) => item.reply_type)
.filter((item): item is AssistantReplyType => typeof item === "string" && item.length > 0)
)
);
const turns = buildTurns(session.items);
const traceIds = unique(session.items.map((item) => item.trace_id));
const replyTypes = Array.from(
new Set(
session.items
.map((item) => item.reply_type)
.filter((item): item is AssistantReplyType => typeof item === "string" && item.length > 0)
)
);
const turns = buildTurns(session.items);
const record: AssistantSessionLogRecord = {
schema_version: "assistant_session_log_v1",
session_id: session.session_id,
started_at: startedAt,
updated_at: session.updated_at,
counters: {
total_messages: session.items.length,
user_messages: userMessages,
assistant_messages: assistantMessages
},
trace_ids: traceIds,
reply_types: replyTypes,
investigation_state: session.investigation_state,
turns,
conversation: session.items,
last_assistant: {
message_id: lastAssistant?.message_id ?? null,
reply_type: lastAssistant?.reply_type ?? null,
trace_id: lastAssistant?.trace_id ?? null,
created_at: lastAssistant?.created_at ?? null
const record: AssistantSessionLogRecord = {
schema_version: "assistant_session_log_v1",
session_id: session.session_id,
started_at: startedAt,
updated_at: session.updated_at,
counters: {
total_messages: session.items.length,
user_messages: userMessages,
assistant_messages: assistantMessages
},
trace_ids: traceIds,
reply_types: replyTypes,
investigation_state: session.investigation_state,
turns,
conversation: session.items,
last_assistant: {
message_id: lastAssistant?.message_id ?? null,
reply_type: lastAssistant?.reply_type ?? null,
trace_id: lastAssistant?.trace_id ?? null,
created_at: lastAssistant?.created_at ?? null
}
};
writeJsonFile(filePath, record);
} catch (error) {
const code = (error as { code?: unknown } | null)?.code;
if (code === "ENOSPC") {
return;
}
};
writeJsonFile(filePath, record);
throw error;
}
}
}
@@ -12,6 +12,7 @@ import {
FEATURE_ASSISTANT_STAGE2_EVAL_V1,
REPORTS_DIR
} from "../config";
import { P0EvalRunner } from "../eval/p0_eval_runner";
import type { AssistantMessageResponsePayload } from "../types/assistant";
import type {
EvalTarget,
@@ -328,6 +329,98 @@ const ASSISTANT_STAGE1_COMPARISON_SCHEMA_VERSION = "assistant_stage1_eval_compar
const DEFAULT_ASSISTANT_STAGE2_SUITE_FILE = "assistant_stage2_canonical_v0_1.json";
const ASSISTANT_STAGE2_RUN_SCHEMA_VERSION = "assistant_stage2_eval_run_v0_1";
const ASSISTANT_STAGE2_COMPARISON_SCHEMA_VERSION = "assistant_stage2_eval_comparison_v0_1";
const INMEM_EVAL_REPORT_PREFIX = "inmem_eval_report:";
const INMEM_EVAL_REPORTS = new Map<string, Record<string, unknown>>();
function isNoSpaceError(error: unknown): boolean {
const code = (error as { code?: unknown } | null)?.code;
return code === "ENOSPC";
}
function tryWriteJsonFile(pathname: string, value: unknown): boolean {
try {
writeJsonFile(pathname, value);
return true;
} catch (error) {
if (isNoSpaceError(error)) {
return false;
}
throw error;
}
}
function tryWriteTextFile(pathname: string, value: string): boolean {
try {
fs.writeFileSync(pathname, value, "utf-8");
return true;
} catch (error) {
if (isNoSpaceError(error)) {
return false;
}
throw error;
}
}
function putInMemoryEvalReport(report: Record<string, unknown>): string {
const key = `${INMEM_EVAL_REPORT_PREFIX}${nanoid(12)}`;
INMEM_EVAL_REPORTS.set(key, report);
return key;
}
function readEvalReportByRef(ref: string): { report: Record<string, unknown>; resolved_path: string } {
if (ref.startsWith(INMEM_EVAL_REPORT_PREFIX)) {
const report = INMEM_EVAL_REPORTS.get(ref);
if (!report) {
throw new Error(`In-memory eval report not found: ${ref}`);
}
return {
report,
resolved_path: ref
};
}
const resolvedPath = resolveReadablePath(ref);
const report = JSON.parse(fs.readFileSync(resolvedPath, "utf-8")) as Record<string, unknown>;
return {
report,
resolved_path: resolvedPath
};
}
function compactAssistantStage1Report(report: Record<string, unknown>): Record<string, unknown> {
const results = Array.isArray(report.results) ? (report.results as Array<Record<string, unknown>>) : [];
const compactResults = results.map((item) => ({
case_id: item.case_id ?? null,
scenario_tag: item.scenario_tag ?? null,
accountant_usefulness_score: item.accountant_usefulness_score ?? null,
accountant_metrics:
typeof item.accountant_metrics === "object" && item.accountant_metrics !== null ? item.accountant_metrics : null
}));
return {
...report,
results: compactResults
};
}
function compactAssistantStage2Report(report: Record<string, unknown>): Record<string, unknown> {
const results = Array.isArray(report.results) ? (report.results as Array<Record<string, unknown>>) : [];
const compactResults = results.map((item) => {
const metricSubscores = (item.metric_subscores ?? {}) as Record<string, unknown>;
return {
case_id: item.case_id ?? null,
metric_subscores: {
problem_clarity_score: metricSubscores.problem_clarity_score ?? null,
mechanism_coherence_score: metricSubscores.mechanism_coherence_score ?? null,
problem_first_answer_rate: metricSubscores.problem_first_answer_rate ?? null,
entity_leakage_rate: metricSubscores.entity_leakage_rate ?? null
}
};
});
return {
...report,
results: compactResults
};
}
type AssistantMetricKey = keyof AssistantEvalMetricVector;
type AssistantStage2MetricKey = keyof AssistantStage2MetricVector;
@@ -1140,7 +1233,7 @@ export class EvalService {
};
ensureDir(EVAL_CASES_DIR);
writeJsonFile(path.resolve(EVAL_CASES_DIR, `${runId}.report.json`), report);
tryWriteJsonFile(path.resolve(EVAL_CASES_DIR, `${runId}.report.json`), report);
return report;
}
@@ -1534,8 +1627,9 @@ export class EvalService {
currentReport: Record<string, unknown>;
baselineReportFile: string;
}): Record<string, unknown> {
const baselinePath = resolveReadablePath(input.baselineReportFile);
const baselineReport = JSON.parse(fs.readFileSync(baselinePath, "utf-8")) as Record<string, unknown>;
const baselineRef = readEvalReportByRef(input.baselineReportFile);
const baselinePath = baselineRef.resolved_path;
const baselineReport = baselineRef.report;
const currentReport = input.currentReport;
const metricKeys: AssistantMetricKey[] = [
"retrieval_differentiation_rate",
@@ -1633,14 +1727,15 @@ export class EvalService {
ensureDir(REPORTS_DIR);
const jsonPath = path.resolve(REPORTS_DIR, `${comparisonId}.json`);
const mdPath = path.resolve(REPORTS_DIR, `${comparisonId}.md`);
writeJsonFile(jsonPath, comparisonReport);
fs.writeFileSync(mdPath, buildAssistantComparisonMarkdownReport(comparisonReport), "utf-8");
const jsonWritten = tryWriteJsonFile(jsonPath, comparisonReport);
const mdWritten = tryWriteTextFile(mdPath, buildAssistantComparisonMarkdownReport(comparisonReport));
const comparisonRef = jsonWritten ? jsonPath : putInMemoryEvalReport(comparisonReport);
return {
...comparisonReport,
artifacts: {
comparison_report_json_path: jsonPath,
comparison_report_md_path: mdPath
comparison_report_json_path: comparisonRef,
comparison_report_md_path: mdWritten ? mdPath : null
}
};
}
@@ -1649,8 +1744,9 @@ export class EvalService {
currentReport: Record<string, unknown>;
baselineReportFile: string;
}): Record<string, unknown> {
const baselinePath = resolveReadablePath(input.baselineReportFile);
const baselineReport = JSON.parse(fs.readFileSync(baselinePath, "utf-8")) as Record<string, unknown>;
const baselineRef = readEvalReportByRef(input.baselineReportFile);
const baselinePath = baselineRef.resolved_path;
const baselineReport = baselineRef.report;
const currentReport = input.currentReport;
const metricKeys: AssistantStage2MetricKey[] = [
"problem_unit_precision",
@@ -1760,14 +1856,15 @@ export class EvalService {
ensureDir(REPORTS_DIR);
const jsonPath = path.resolve(REPORTS_DIR, `${comparisonId}.json`);
const mdPath = path.resolve(REPORTS_DIR, `${comparisonId}.md`);
writeJsonFile(jsonPath, comparisonReport);
fs.writeFileSync(mdPath, buildAssistantStage2ComparisonMarkdownReport(comparisonReport), "utf-8");
const jsonWritten = tryWriteJsonFile(jsonPath, comparisonReport);
const mdWritten = tryWriteTextFile(mdPath, buildAssistantStage2ComparisonMarkdownReport(comparisonReport));
const comparisonRef = jsonWritten ? jsonPath : putInMemoryEvalReport(comparisonReport);
return {
...comparisonReport,
artifacts: {
comparison_report_json_path: jsonPath,
comparison_report_md_path: mdPath
comparison_report_json_path: comparisonRef,
comparison_report_md_path: mdWritten ? mdPath : null
}
};
}
@@ -1803,7 +1900,7 @@ export class EvalService {
try {
for (const turn of suiteCase.turns) {
const response = await assistantService.handleMessage({
const response = (await assistantService.handleMessage({
session_id: sessionId,
user_message: turn.user_message,
message: turn.user_message,
@@ -1819,7 +1916,7 @@ export class EvalService {
domainPrompt: payload.normalizeConfig.domainPrompt,
fewShotExamples: payload.normalizeConfig.fewShotExamples,
useMock: payload.useMock
});
})) as AssistantMessageResponsePayload;
turnResponses.push(response);
requestsTotal += 1;
}
@@ -2099,12 +2196,14 @@ export class EvalService {
ensureDir(REPORTS_DIR);
const runJsonPath = path.resolve(REPORTS_DIR, `${runId}.json`);
const runMdPath = path.resolve(REPORTS_DIR, `${runId}.md`);
writeJsonFile(runJsonPath, report);
fs.writeFileSync(runMdPath, buildAssistantEvalMarkdownReport(report), "utf-8");
const compactReport = compactAssistantStage1Report(report);
const jsonWritten = tryWriteJsonFile(runJsonPath, compactReport);
const mdWritten = tryWriteTextFile(runMdPath, buildAssistantEvalMarkdownReport(compactReport));
const runReportRef = jsonWritten ? runJsonPath : putInMemoryEvalReport(compactReport);
report.artifacts = {
run_report_json_path: runJsonPath,
run_report_md_path: runMdPath
run_report_json_path: runReportRef,
run_report_md_path: mdWritten ? runMdPath : null
};
if (payload.compareWithReportFile) {
@@ -2151,7 +2250,7 @@ export class EvalService {
try {
for (const turn of suiteCase.turns) {
const response = await assistantService.handleMessage({
const response = (await assistantService.handleMessage({
session_id: sessionId,
user_message: turn.user_message,
message: turn.user_message,
@@ -2167,7 +2266,7 @@ export class EvalService {
domainPrompt: payload.normalizeConfig.domainPrompt,
fewShotExamples: payload.normalizeConfig.fewShotExamples,
useMock: payload.useMock
});
})) as AssistantMessageResponsePayload;
turnResponses.push(response);
requestsTotal += 1;
}
@@ -2393,12 +2492,14 @@ export class EvalService {
ensureDir(REPORTS_DIR);
const runJsonPath = path.resolve(REPORTS_DIR, `${runId}.json`);
const runMdPath = path.resolve(REPORTS_DIR, `${runId}.md`);
writeJsonFile(runJsonPath, report);
fs.writeFileSync(runMdPath, buildAssistantStage2EvalMarkdownReport(report), "utf-8");
const compactReport = compactAssistantStage2Report(report);
const jsonWritten = tryWriteJsonFile(runJsonPath, compactReport);
const mdWritten = tryWriteTextFile(runMdPath, buildAssistantStage2EvalMarkdownReport(compactReport));
const runReportRef = jsonWritten ? runJsonPath : putInMemoryEvalReport(compactReport);
report.artifacts = {
run_report_json_path: runJsonPath,
run_report_md_path: runMdPath
run_report_json_path: runReportRef,
run_report_md_path: mdWritten ? runMdPath : null
};
if (payload.compareWithReportFile) {
@@ -2411,6 +2512,33 @@ export class EvalService {
return report;
}
private async runAssistantP0(payload: {
normalizeConfig: Omit<NormalizeRequestPayload, "userQuestion" | "context">;
caseIds?: string[];
useMock?: boolean;
mode: EvalRunMode;
caseSetFile?: string;
compareWithReportFile?: string;
}): Promise<Record<string, unknown>> {
if (!FEATURE_ASSISTANT_STAGE2_EVAL_V1) {
throw new ApiError(
"ASSISTANT_P0_EVAL_DISABLED",
"Assistant P0 eval target is disabled by FEATURE_ASSISTANT_STAGE2_EVAL_V1.",
409
);
}
const runner = new P0EvalRunner(this.normalizerService);
return runner.run({
normalizeConfig: payload.normalizeConfig,
caseIds: payload.caseIds,
useMock: payload.useMock,
mode: payload.mode,
caseSetFile: payload.caseSetFile,
compareWithReportFile: payload.compareWithReportFile
});
}
public async run(payload: {
normalizeConfig: Omit<NormalizeRequestPayload, "userQuestion" | "context">;
caseIds?: string[];
@@ -2446,6 +2574,17 @@ export class EvalService {
});
}
if (evalTarget === "assistant_p0") {
return this.runAssistantP0({
normalizeConfig: payload.normalizeConfig,
caseIds: payload.caseIds,
useMock: payload.useMock,
mode,
caseSetFile: payload.caseSetFile,
compareWithReportFile: payload.compareWithReportFile
});
}
const promptVersion = String(payload.normalizeConfig.promptVersion ?? "").toLowerCase();
const schemaVersion = String(payload.normalizeConfig.schemaVersion ?? "").toLowerCase();
const isV2 =
@@ -2659,7 +2798,7 @@ export class EvalService {
};
ensureDir(EVAL_CASES_DIR);
writeJsonFile(path.resolve(EVAL_CASES_DIR, `${runId}.report.json`), report);
tryWriteJsonFile(path.resolve(EVAL_CASES_DIR, `${runId}.report.json`), report);
const shouldWriteV11Artifacts =
mode === "single-pass-strict" &&
@@ -2668,14 +2807,13 @@ export class EvalService {
if (shouldWriteV11Artifacts) {
ensureDir(REPORTS_DIR);
writeJsonFile(path.resolve(REPORTS_DIR, "normalizer_eval_v1_1_run.json"), report);
fs.writeFileSync(
tryWriteJsonFile(path.resolve(REPORTS_DIR, "normalizer_eval_v1_1_run.json"), report);
tryWriteTextFile(
path.resolve(REPORTS_DIR, "normalizer_eval_v1_1_run.md"),
buildMarkdownReport({
...report,
report_title: "LLM Normalizer v1.1 Eval Run"
}),
"utf-8"
})
);
}
@@ -2687,14 +2825,13 @@ export class EvalService {
if (shouldWriteV1121EvalArtifacts) {
ensureDir(REPORTS_DIR);
writeJsonFile(path.resolve(REPORTS_DIR, "normalizer_v1_1_2_1_eval.json"), report);
fs.writeFileSync(
tryWriteJsonFile(path.resolve(REPORTS_DIR, "normalizer_v1_1_2_1_eval.json"), report);
tryWriteTextFile(
path.resolve(REPORTS_DIR, "normalizer_v1_1_2_1_eval.md"),
buildMarkdownReport({
...report,
report_title: "LLM Normalizer v1.1.2.1 Eval Run"
}),
"utf-8"
})
);
}
@@ -2705,14 +2842,13 @@ export class EvalService {
if (shouldWriteV111MicroArtifacts) {
ensureDir(REPORTS_DIR);
writeJsonFile(path.resolve(REPORTS_DIR, "normalizer_v1_1_1_micro_eval.json"), report);
fs.writeFileSync(
tryWriteJsonFile(path.resolve(REPORTS_DIR, "normalizer_v1_1_1_micro_eval.json"), report);
tryWriteTextFile(
path.resolve(REPORTS_DIR, "normalizer_v1_1_1_micro_eval.md"),
buildMarkdownReport({
...report,
report_title: "LLM Normalizer v1.1.1 Micro Eval"
}),
"utf-8"
})
);
}
@@ -2723,14 +2859,13 @@ export class EvalService {
if (shouldWriteV112MicroArtifacts) {
ensureDir(REPORTS_DIR);
writeJsonFile(path.resolve(REPORTS_DIR, "normalizer_v1_1_2_micro_eval.json"), report);
fs.writeFileSync(
tryWriteJsonFile(path.resolve(REPORTS_DIR, "normalizer_v1_1_2_micro_eval.json"), report);
tryWriteTextFile(
path.resolve(REPORTS_DIR, "normalizer_v1_1_2_micro_eval.md"),
buildMarkdownReport({
...report,
report_title: "LLM Normalizer v1.1.2 Micro Eval"
}),
"utf-8"
})
);
}
@@ -127,6 +127,103 @@ function collectOpenUncertainties(
return capStrings([...requirementNotes, ...limitationNotes], INVESTIGATION_MAX_UNCERTAINTIES);
}
function normalizeAccountPrefix(value: string): string | null {
const account = String(value ?? "").trim();
if (!account) {
return null;
}
const match = account.match(/^(\d{2})/);
return match?.[1] ?? null;
}
function isSettlementAccount(value: string): boolean {
const prefix = normalizeAccountPrefix(value);
return prefix === "60" || prefix === "62" || prefix === "51" || prefix === "76";
}
function isVatAccount(value: string): boolean {
const prefix = normalizeAccountPrefix(value);
return prefix === "19" || prefix === "68";
}
function isCloseCostsAccount(value: string): boolean {
const prefix = normalizeAccountPrefix(value);
if (!prefix) {
return false;
}
const account = Number(prefix);
return (account >= 20 && account <= 44) || prefix === "97";
}
function inferFollowupActiveDomain(input: {
userMessage: string;
focusAccounts: string[];
routeSummary: RouteHintSummary | null;
previous: InvestigationStateWithProblemUnits;
}): string | null {
const corpus = `${input.userMessage} ${input.previous.focus.active_query_subject ?? ""}`.toLowerCase();
const hasSettlementSignal =
input.focusAccounts.some((item) => isSettlementAccount(item)) ||
/(60(?:\.\d{2})?|62(?:\.\d{2})?|оплат|расчет|расч[её]т|зачет|зач[её]т|аванс|долг|поставщ|покупат|settlement|payment|supplier|customer)/i.test(
corpus
);
if (hasSettlementSignal) {
return "settlements_60_62";
}
const hasVatSignal =
input.focusAccounts.some((item) => isVatAccount(item)) ||
/(ндс|счет[\s-]?фактур|сч[её]т[\s-]?фактур|книг[аи]|vat|invoice|book|register)/i.test(corpus);
if (hasVatSignal) {
return "vat_document_register_book";
}
const hasCloseSignal =
input.focusAccounts.some((item) => isCloseCostsAccount(item)) ||
/(закрыти|закрытие|месяц|затрат|распредел|списан|period\s*close|month\s*close|allocation|residual|cost)/i.test(corpus);
if (hasCloseSignal) {
return "month_close_costs_20_44";
}
const routeDomain = deriveDomain(input.routeSummary);
if (routeDomain && routeDomain !== "no_route") {
return routeDomain;
}
return input.previous.followup_context?.active_domain ?? input.previous.focus.domain ?? null;
}
function collectUncoveredRequirementIds(coverageReport: RequirementCoverageReport): string[] {
return capStrings(
[
...coverageReport.requirements_uncovered,
...coverageReport.requirements_partially_covered,
...coverageReport.clarification_needed_for,
...coverageReport.out_of_scope_requirements
],
INVESTIGATION_MAX_REQUIREMENT_LINKS
);
}
function collectEvidenceSummary(retrievalResults: UnifiedRetrievalResult[]): string[] {
const lines = retrievalResults.map((result) => {
const requirementRef = result.requirement_ids[0] ?? result.fragment_id;
return `${requirementRef}:${result.status}:${result.route}`;
});
return capStrings(lines, 6);
}
function settlementFocusActions(activeDomain: string | null): string[] {
if (activeDomain !== "settlements_60_62") {
return [];
}
return [
"Проверьте договор и объект расчетов по платежу.",
"Сверьте регистр расчетов и привязку платежа к закрывающему документу.",
"Проверьте зачет аванса или взаимозачет по связке 60/62."
];
}
function normalizeEntityBacklinks(values: ProblemUnitEntityBacklink[]): ProblemUnitEntityBacklink[] {
const result: ProblemUnitEntityBacklink[] = [];
const seen = new Set<string>();
@@ -273,7 +370,27 @@ export function cloneInvestigationState(state: InvestigationStateWithProblemUnit
followup_context: state.followup_context
? {
...state.followup_context,
referenced_requirement_ids: [...state.followup_context.referenced_requirement_ids]
referenced_requirement_ids: [...state.followup_context.referenced_requirement_ids],
...(state.followup_context.active_requirement_ids
? {
active_requirement_ids: [...state.followup_context.active_requirement_ids]
}
: {}),
...(state.followup_context.uncovered_requirement_ids
? {
uncovered_requirement_ids: [...state.followup_context.uncovered_requirement_ids]
}
: {}),
...(state.followup_context.settlement_next_actions
? {
settlement_next_actions: [...state.followup_context.settlement_next_actions]
}
: {}),
...(state.followup_context.evidence_summary
? {
evidence_summary: [...state.followup_context.evidence_summary]
}
: {})
}
: null
};
@@ -320,12 +437,27 @@ export function createEmptyInvestigationState(
export function updateInvestigationState(input: UpdateInvestigationStateInput): InvestigationStateWithProblemUnits {
const previous = input.previous;
const focusFromMessage = capStrings(detectAccounts(input.userMessage), INVESTIGATION_MAX_PRIMARY_ACCOUNTS);
const mergedFocusAccounts = capStrings(
[...focusFromMessage, ...previous.focus.primary_accounts],
INVESTIGATION_MAX_PRIMARY_ACCOUNTS
);
const requirementIds = capStrings(
input.requirements.map((item) => item.requirement_id),
INVESTIGATION_MAX_REQUIREMENT_LINKS
);
const mainRequirement = input.requirements[0]?.requirement_text ?? input.userMessage;
const problemUnitState = updateProblemUnitState(previous, input.retrievalResults);
const uncoveredRequirementIds = collectUncoveredRequirementIds(input.coverageReport);
const activeDomain = inferFollowupActiveDomain({
userMessage: input.userMessage,
focusAccounts: mergedFocusAccounts,
routeSummary: input.routeSummary,
previous
});
const focusDomain = activeDomain ?? deriveDomain(input.routeSummary) ?? previous.focus.domain;
const settlementNextActions = settlementFocusActions(activeDomain);
const lastProblemUnitId = problemUnitState?.active_problem_units[0] ?? null;
const evidenceSummary = collectEvidenceSummary(input.retrievalResults);
return {
schema_version: INVESTIGATION_STATE_SCHEMA_VERSION,
@@ -335,12 +467,9 @@ export function updateInvestigationState(input: UpdateInvestigationStateInput):
updated_at: input.timestamp,
question_id: input.questionId,
focus: {
domain: deriveDomain(input.routeSummary) ?? previous.focus.domain,
domain: focusDomain,
period: detectPeriod(input.userMessage) ?? previous.focus.period,
primary_accounts: capStrings(
[...focusFromMessage, ...previous.focus.primary_accounts],
INVESTIGATION_MAX_PRIMARY_ACCOUNTS
),
primary_accounts: mergedFocusAccounts,
active_query_subject: mainRequirement.slice(0, 180)
},
narrowing_status: deriveNarrowingStatus(input.routeSummary, input.coverageReport),
@@ -353,7 +482,13 @@ export function updateInvestigationState(input: UpdateInvestigationStateInput):
followup_context: {
previous_question_id: previous.question_id,
last_user_message: input.userMessage.slice(0, 240),
referenced_requirement_ids: requirementIds
referenced_requirement_ids: requirementIds,
active_domain: activeDomain,
active_requirement_ids: requirementIds,
uncovered_requirement_ids: uncoveredRequirementIds,
last_problem_unit_id: lastProblemUnitId,
settlement_next_actions: settlementNextActions,
evidence_summary: evidenceSummary
},
query_mode_hint: deriveQueryModeHint(input.routeSummary),
...(problemUnitState
@@ -4,7 +4,11 @@ import type {
RetrievalResultType,
UnifiedRetrievalResult
} from "../types/assistant";
import { FEATURE_ASSISTANT_EVIDENCE_ENRICHMENT_V1, FEATURE_ASSISTANT_PROBLEM_UNITS_V1 } from "../config";
import {
FEATURE_ASSISTANT_EVIDENCE_ENRICHMENT_V1,
FEATURE_ASSISTANT_GRAPH_RUNTIME_V1,
FEATURE_ASSISTANT_PROBLEM_UNITS_V1
} from "../config";
import { EVIDENCE_SOURCE_REF_SCHEMA_VERSION } from "../types/stage1Contracts";
import type {
EvidenceConfidence,
@@ -15,6 +19,8 @@ import type {
EvidenceSourceRef
} from "../types/stage1Contracts";
import { assembleProblemUnits } from "./problemUnitAssembler";
import { buildAccountingGraph } from "./stage4GraphRuntime";
import type { GraphSignalSummary } from "../types/stage4Graph";
interface RawRetrievalResult {
status?: string;
@@ -124,6 +130,22 @@ function mergeSummaryWithProblemUnitMeta(
};
}
function mergeSummaryWithGraphMeta(summary: Record<string, unknown>, graphSummary: GraphSignalSummary): Record<string, unknown> {
return {
...summary,
graph_runtime_enabled: true,
graph_total_units: graphSummary.total_units,
graph_bound_units: graphSummary.bound_units,
graph_nodes_count: graphSummary.node_count,
graph_edges_count: graphSummary.edge_count,
graph_missing_links_count: graphSummary.missing_links_count,
graph_conflicting_links_count: graphSummary.conflicting_links_count,
graph_coverage_grade: graphSummary.graph_coverage_grade,
graph_domain_distribution: graphSummary.domain_distribution,
graph_relation_distribution: graphSummary.relation_distribution
};
}
function normalizeConfidence(value: unknown): RetrievalConfidence {
if (value === "high" || value === "medium" || value === "low") {
return value;
@@ -526,24 +548,60 @@ export function normalizeRetrievalResult(
business_interpretation: baseResult.business_interpretation
});
const enrichedSummary = mergeSummaryWithProblemUnitMeta(summary, {
candidateEvidenceCount: assembled.candidate_evidence.length,
problemUnitsCount: assembled.problem_units.length,
unitTypes: assembled.problem_unit_summary.unit_types,
duplicateCollapses: assembled.problem_unit_summary.duplicate_collapses,
severityDistribution: assembled.problem_unit_summary.severity_distribution,
confidenceDistribution: assembled.problem_unit_summary.confidence_distribution,
lifecycleEnrichedUnits: assembled.problem_unit_summary.lifecycle_enriched_units ?? 0,
lifecycleDomainDistribution: (assembled.problem_unit_summary.lifecycle_domain_distribution ?? {}) as Record<string, number>,
lifecycleDefectDistribution: (assembled.problem_unit_summary.lifecycle_defect_distribution ?? {}) as Record<string, number>
const graphBuild = FEATURE_ASSISTANT_GRAPH_RUNTIME_V1
? buildAccountingGraph({
route,
candidateEvidence: assembled.candidate_evidence,
problemUnits: assembled.problem_units
})
: null;
const graphBindingByUnitId = new Map(graphBuild?.unit_bindings.map((item) => [item.problem_unit_id, item] as const) ?? []);
const graphBoundProblemUnits = assembled.problem_units.map((unit) => {
const binding = graphBindingByUnitId.get(unit.problem_unit_id);
if (!binding) {
return unit;
}
return {
...unit,
graph_binding: binding
};
});
const graphBoundSummary = graphBuild
? {
...assembled.problem_unit_summary,
graph_summary: graphBuild.summary
}
: assembled.problem_unit_summary;
let enrichedSummary = mergeSummaryWithProblemUnitMeta(summary, {
candidateEvidenceCount: assembled.candidate_evidence.length,
problemUnitsCount: graphBoundProblemUnits.length,
unitTypes: graphBoundSummary.unit_types,
duplicateCollapses: graphBoundSummary.duplicate_collapses,
severityDistribution: graphBoundSummary.severity_distribution,
confidenceDistribution: graphBoundSummary.confidence_distribution,
lifecycleEnrichedUnits: graphBoundSummary.lifecycle_enriched_units ?? 0,
lifecycleDomainDistribution: (graphBoundSummary.lifecycle_domain_distribution ?? {}) as Record<string, number>,
lifecycleDefectDistribution: (graphBoundSummary.lifecycle_defect_distribution ?? {}) as Record<string, number>
});
if (graphBuild) {
enrichedSummary = mergeSummaryWithGraphMeta(enrichedSummary, graphBuild.summary);
}
return {
...baseResult,
summary: enrichedSummary,
raw_entities: items,
candidate_evidence: assembled.candidate_evidence,
problem_units: assembled.problem_units,
problem_unit_summary: assembled.problem_unit_summary
problem_units: graphBoundProblemUnits,
problem_unit_summary: graphBoundSummary,
...(graphBuild
? {
accounting_graph: graphBuild
}
: {})
};
}
@@ -1,4 +1,5 @@
import type {
DeterministicRouteHint,
NoRouteReason,
NormalizedPayload,
NormalizedQueryV1,
@@ -42,6 +43,244 @@ function toRouteHintSummaryV1(normalized: NormalizedQueryV1): RouteHintSummaryV1
type V2Family = NormalizedQueryV2 | NormalizedQueryV2_0_1 | NormalizedQueryV2_0_2;
type V2FamilyFragment = V2Family["fragments"][number];
type RouteQueryClass =
| "exact_object_trace"
| "ranking_or_period_summary"
| "symptom_first"
| "lifecycle_first"
| "chain_break"
| "period_impact"
| "causal_query"
| "mixed_ambiguity"
| "rule_check_without_symptom"
| "canonical_fact_lookup";
interface RouteDisciplineRule {
query_class: RouteQueryClass;
required_route: Exclude<DeterministicRouteHint, "no_route">;
allowed_fallback: DeterministicRouteHint[];
forbidden_fallback: DeterministicRouteHint[];
description: string;
}
const ACCOUNT_HINT_PATTERN =
/(?:\b(?:account|acct|schet|счет|сч)\s*[:#]?\s*(?:[1-9][0-9](?:[./-][0-9]{1,2})?)|\b(?:19|20|21|23|25|26|28|29|44|51|60|62|68)\b)/i;
const PERIOD_PATTERN = /\b20\d{2}(?:[-./](?:0[1-9]|1[0-2]))?\b/i;
const SYMPTOM_MARKER_PATTERN =
/(?:\bsymptom\b|\banomaly\b|\bproblem\b|\bissue\b|\btail\b|\bhanging\b|\bblocked\b|\bincomplete\b|remains?\s+open|not\s+(?:confirmed|observed|resolved|closed)|не\s+(?:подтвержден|закрыт|наблюдается)|хвост|сбой|проблем)/i;
const LIFECYCLE_MARKER_PATTERN =
/(?:\blifecycle\b|\bchain\b|\btransition\b|\bstep\b|\btrace\b|цепоч|этап|переход|связк|где\s+разрыв)/i;
const CHAIN_BREAK_PATTERN =
/(?:\bbreak\b|\bbroken\b|\bgap\b|missing\s+(?:transition|step|link)|chain\s+break|разрыв|обрыв|нет\s+переход|не\s+дошл|не\s+наблюд)/i;
const PERIOD_IMPACT_PATTERN =
/(?:period\s*close|month\s*close|month-end|residual|allocation|20[/-]44|закрыти|остатк|распредел|конец\s+месяц)/i;
const CAUSAL_PATTERN = /(?:\bwhy\b|\bbecause\b|\breason\b|explain\s+mechanism|почему|объясни|механизм|причин)/i;
const AMBIGUITY_PATTERN =
/(?:\bmaybe\b|\bperhaps\b|not\s+sure|i\s+only\s+know|part\s+may\s+be\s+missing|возможно|может\s+быть|не\s+уверен|не\s+знаю|часть\s+цепочки\s+не\s+подтвержд)/i;
const TRANSLIT_PROBLEM_PATTERN = /(?:raschet|oplata|zakryt|nds|vychet|zatrat|ostatok|cepoch|perehod|pochemu|prichin|period)/i;
const DOMAIN_LEXICAL_ANCHOR_PATTERN =
/(?:\b(?:settlement|payment|bank|supplier|customer|vat|nds|invoice|register|book|period\s*close|month\s*close|close\s*operation|allocation|residual|cost|expenses?)\b|оплат|расчет|РЅРґСЃ|СЃС‡[её]С‚.?фактур|РєРЅРёРі[аи]|затрат|закрыт|остатк)/i;
export const ROUTE_DISCIPLINE_RULE_TABLE: RouteDisciplineRule[] = [
{
query_class: "exact_object_trace",
required_route: "live_mcp_drilldown",
allowed_fallback: ["no_route"],
forbidden_fallback: ["store_canonical", "hybrid_store_plus_live", "store_feature_risk", "batch_refresh_then_store"],
description: "Exact object trace queries always run via live drilldown."
},
{
query_class: "ranking_or_period_summary",
required_route: "batch_refresh_then_store",
allowed_fallback: ["no_route"],
forbidden_fallback: ["store_canonical", "hybrid_store_plus_live"],
description: "Ranking and period summary queries require analytical batch path."
},
{
query_class: "symptom_first",
required_route: "hybrid_store_plus_live",
allowed_fallback: ["no_route"],
forbidden_fallback: ["store_canonical"],
description: "Symptom-first intents are deterministically promoted to hybrid path."
},
{
query_class: "lifecycle_first",
required_route: "hybrid_store_plus_live",
allowed_fallback: ["no_route"],
forbidden_fallback: ["store_canonical"],
description: "Lifecycle-first intents are deterministically promoted to hybrid path."
},
{
query_class: "chain_break",
required_route: "hybrid_store_plus_live",
allowed_fallback: ["no_route"],
forbidden_fallback: ["store_canonical"],
description: "Chain-break intents are deterministically promoted to hybrid path."
},
{
query_class: "period_impact",
required_route: "hybrid_store_plus_live",
allowed_fallback: ["no_route"],
forbidden_fallback: ["store_canonical"],
description: "Period-impact problem intents are deterministically promoted to hybrid path."
},
{
query_class: "causal_query",
required_route: "hybrid_store_plus_live",
allowed_fallback: ["no_route"],
forbidden_fallback: ["store_canonical"],
description: "Causal/mechanism intents are deterministically promoted to hybrid path."
},
{
query_class: "mixed_ambiguity",
required_route: "hybrid_store_plus_live",
allowed_fallback: ["no_route"],
forbidden_fallback: ["store_canonical"],
description: "Mixed ambiguity keeps hybrid as primary route, with explicit no-route fallback."
},
{
query_class: "rule_check_without_symptom",
required_route: "store_feature_risk",
allowed_fallback: ["no_route"],
forbidden_fallback: ["store_canonical"],
description: "Rule checks without symptom/lifecycle signals run via risk profile path."
},
{
query_class: "canonical_fact_lookup",
required_route: "store_canonical",
allowed_fallback: ["no_route"],
forbidden_fallback: ["hybrid_store_plus_live"],
description: "Only plain factual lookups are allowed to stay on canonical path."
}
];
const ROUTE_DISCIPLINE_RULE_MAP = new Map<RouteQueryClass, RouteDisciplineRule>(
ROUTE_DISCIPLINE_RULE_TABLE.map((item) => [item.query_class, item])
);
function mergedFragmentText(fragment: V2FamilyFragment): string {
return `${fragment.raw_fragment_text ?? ""} ${fragment.normalized_fragment_text ?? ""}`.toLowerCase();
}
function hasLifecycleDomainHint(fragment: V2FamilyFragment, lowerText: string): boolean {
const accountHints = Array.isArray(fragment.account_hints) ? fragment.account_hints.map((item) => String(item)) : [];
if (accountHints.some((item) => /^(97|01|02|08|19|20|21|23|25|26|28|29|44|68(?:\.\d+)?|51|60|62)$/.test(item))) {
return true;
}
return (
fragment.candidate_labels.includes("anomaly_probe") ||
fragment.candidate_labels.includes("period_close_risk") ||
PERIOD_IMPACT_PATTERN.test(lowerText)
);
}
function hasSymptomSignal(fragment: V2FamilyFragment, lowerText: string): boolean {
return (
fragment.flags.asks_for_anomaly_scan ||
fragment.candidate_labels.includes("anomaly_probe") ||
fragment.candidate_labels.includes("period_close_risk") ||
SYMPTOM_MARKER_PATTERN.test(lowerText) ||
TRANSLIT_PROBLEM_PATTERN.test(lowerText)
);
}
function hasLifecycleSignal(fragment: V2FamilyFragment, lowerText: string): boolean {
return (
fragment.flags.asks_for_chain_explanation ||
fragment.flags.mentions_period_close_context ||
LIFECYCLE_MARKER_PATTERN.test(lowerText) ||
hasLifecycleDomainHint(fragment, lowerText)
);
}
function hasChainBreakSignal(lowerText: string): boolean {
return CHAIN_BREAK_PATTERN.test(lowerText);
}
function hasPeriodImpactSignal(lowerText: string): boolean {
return PERIOD_IMPACT_PATTERN.test(lowerText);
}
function hasCausalSignal(lowerText: string): boolean {
return CAUSAL_PATTERN.test(lowerText);
}
function hasAmbiguitySignal(fragment: V2FamilyFragment, lowerText: string): boolean {
return (
AMBIGUITY_PATTERN.test(lowerText) ||
fragment.confidence === "low" ||
fragment.domain_relevance === "unclear" ||
fragment.business_scope === "unclear"
);
}
function hasAccountOrPeriodAnchor(fragment: V2FamilyFragment, lowerText: string): boolean {
return fragment.account_hints.length > 0 || ACCOUNT_HINT_PATTERN.test(lowerText) || PERIOD_PATTERN.test(lowerText);
}
function resolveRouteClass(fragment: V2FamilyFragment): RouteDisciplineRule {
const lowerText = mergedFragmentText(fragment);
const symptomSignal = hasSymptomSignal(fragment, lowerText);
const lifecycleSignal = hasLifecycleSignal(fragment, lowerText);
const chainBreakSignal = hasChainBreakSignal(lowerText);
const periodImpactSignal = hasPeriodImpactSignal(lowerText);
const causalSignal = hasCausalSignal(lowerText);
const ambiguitySignal = hasAmbiguitySignal(fragment, lowerText);
const accountOrPeriodAnchor = hasAccountOrPeriodAnchor(fragment, lowerText);
if (fragment.flags.asks_for_exact_object_trace) {
return ROUTE_DISCIPLINE_RULE_MAP.get("exact_object_trace")!;
}
if (fragment.flags.asks_for_ranking_or_top || fragment.flags.asks_for_period_summary) {
return ROUTE_DISCIPLINE_RULE_MAP.get("ranking_or_period_summary")!;
}
if (ambiguitySignal && (symptomSignal || lifecycleSignal || chainBreakSignal || periodImpactSignal || causalSignal)) {
return ROUTE_DISCIPLINE_RULE_MAP.get("mixed_ambiguity")!;
}
if (chainBreakSignal) {
return ROUTE_DISCIPLINE_RULE_MAP.get("chain_break")!;
}
if (periodImpactSignal && accountOrPeriodAnchor) {
return ROUTE_DISCIPLINE_RULE_MAP.get("period_impact")!;
}
if (lifecycleSignal) {
return ROUTE_DISCIPLINE_RULE_MAP.get("lifecycle_first")!;
}
if (symptomSignal) {
return ROUTE_DISCIPLINE_RULE_MAP.get("symptom_first")!;
}
if (causalSignal && accountOrPeriodAnchor) {
return ROUTE_DISCIPLINE_RULE_MAP.get("causal_query")!;
}
if (fragment.flags.asks_for_rule_check) {
return ROUTE_DISCIPLINE_RULE_MAP.get("rule_check_without_symptom")!;
}
return ROUTE_DISCIPLINE_RULE_MAP.get("canonical_fact_lookup")!;
}
function shouldPromoteFromNoRoute(fragment: V2FamilyFragment, rule: RouteDisciplineRule): boolean {
if (rule.required_route === "store_canonical") {
return false;
}
if (explicitNoRouteReason(fragment) === "out_of_scope") {
return false;
}
const lowerText = mergedFragmentText(fragment);
const hasProblemSignal =
hasSymptomSignal(fragment, lowerText) ||
hasLifecycleSignal(fragment, lowerText) ||
hasChainBreakSignal(lowerText) ||
hasPeriodImpactSignal(lowerText) ||
hasCausalSignal(lowerText);
const hasAnchor =
hasAccountOrPeriodAnchor(fragment, lowerText) ||
fragment.candidate_labels.includes("cross_entity") ||
DOMAIN_LEXICAL_ANCHOR_PATTERN.test(lowerText);
return hasProblemSignal && hasAnchor;
}
function reasonForNoRoute(noRouteReason: NoRouteReason | null | undefined): string {
if (noRouteReason === "out_of_scope") {
return "Fragment is out-of-scope for company-specific accounting contour.";
@@ -98,37 +337,33 @@ function decideRouteForFragment(fragment: V2FamilyFragment): RouteDecisionV2 {
const readiness = executionReadiness(fragment);
const clarification = clarificationReason(fragment);
const soft = softAssumptions(fragment);
const routeRule = resolveRouteClass(fragment);
if (status === "no_route") {
return buildNoRouteDecision(fragment, noRouteReason);
}
if (readiness === "needs_clarification" || readiness === "no_route") {
return buildNoRouteDecision(fragment, noRouteReason ?? "insufficient_specificity");
}
if (fragment.domain_relevance !== "in_scope") {
if (fragment.domain_relevance === "out_of_scope") {
return buildNoRouteDecision(fragment, "out_of_scope");
}
if (fragment.flags.asks_for_exact_object_trace) {
return {
fragment_id: fragment.fragment_id,
domain_relevance: fragment.domain_relevance,
business_scope: fragment.business_scope,
candidate_labels: fragment.candidate_labels,
decision_flags: fragment.flags,
execution_readiness: readiness,
clarification_reason: clarification,
soft_assumption_used: soft,
route_status: "routed",
no_route_reason: null,
route: "live_mcp_drilldown",
reason: "Exact object trace requested."
};
if (status === "no_route" || readiness === "needs_clarification" || readiness === "no_route") {
if (shouldPromoteFromNoRoute(fragment, routeRule)) {
return {
fragment_id: fragment.fragment_id,
domain_relevance: fragment.domain_relevance,
business_scope: fragment.business_scope,
candidate_labels: fragment.candidate_labels,
decision_flags: fragment.flags,
execution_readiness: readiness,
clarification_reason: clarification,
soft_assumption_used: soft,
route_status: "routed",
no_route_reason: null,
route: routeRule.required_route,
reason: `${routeRule.description} Query class: ${routeRule.query_class}. Promoted from no-route by anchor/symptom guardrail.`
};
}
return buildNoRouteDecision(fragment, noRouteReason);
}
if (fragment.flags.asks_for_ranking_or_top || fragment.flags.asks_for_period_summary) {
if (status === "routed" || status === null) {
return {
fragment_id: fragment.fragment_id,
domain_relevance: fragment.domain_relevance,
@@ -140,80 +375,8 @@ function decideRouteForFragment(fragment: V2FamilyFragment): RouteDecisionV2 {
soft_assumption_used: soft,
route_status: "routed",
no_route_reason: null,
route: "batch_refresh_then_store",
reason: "Ranking/summary semantics require batch analytical route."
};
}
if (fragment.flags.has_multi_entity_scope && fragment.flags.asks_for_chain_explanation) {
return {
fragment_id: fragment.fragment_id,
domain_relevance: fragment.domain_relevance,
business_scope: fragment.business_scope,
candidate_labels: fragment.candidate_labels,
decision_flags: fragment.flags,
execution_readiness: readiness,
clarification_reason: clarification,
soft_assumption_used: soft,
route_status: "routed",
no_route_reason: null,
route: "hybrid_store_plus_live",
reason: "Multi-entity causal chain requested."
};
}
if (fragment.flags.asks_for_rule_check && !fragment.flags.asks_for_chain_explanation) {
return {
fragment_id: fragment.fragment_id,
domain_relevance: fragment.domain_relevance,
business_scope: fragment.business_scope,
candidate_labels: fragment.candidate_labels,
decision_flags: fragment.flags,
execution_readiness: readiness,
clarification_reason: clarification,
soft_assumption_used: soft,
route_status: "routed",
no_route_reason: null,
route: "store_feature_risk",
reason: "Rule-control check without causal decomposition."
};
}
if (
fragment.flags.asks_for_anomaly_scan &&
!fragment.flags.asks_for_ranking_or_top &&
!(fragment.flags.has_multi_entity_scope && fragment.flags.asks_for_chain_explanation)
) {
return {
fragment_id: fragment.fragment_id,
domain_relevance: fragment.domain_relevance,
business_scope: fragment.business_scope,
candidate_labels: fragment.candidate_labels,
decision_flags: fragment.flags,
execution_readiness: readiness,
clarification_reason: clarification,
soft_assumption_used: soft,
route_status: "routed",
no_route_reason: null,
route: "store_feature_risk",
reason: "Anomaly scan without heavy ranking or causal chain."
};
}
if (status === "routed") {
return {
fragment_id: fragment.fragment_id,
domain_relevance: fragment.domain_relevance,
business_scope: fragment.business_scope,
candidate_labels: fragment.candidate_labels,
decision_flags: fragment.flags,
execution_readiness: readiness,
clarification_reason: clarification,
soft_assumption_used: soft,
route_status: "routed",
no_route_reason: null,
route: "store_canonical",
reason: "Routed fragment without deep analytical or causal signals."
route: routeRule.required_route,
reason: `${routeRule.description} Query class: ${routeRule.query_class}. Allowed fallback: ${routeRule.allowed_fallback.join(", ")}. Forbidden fallback: ${routeRule.forbidden_fallback.join(", ")}.`
};
}
@@ -0,0 +1,662 @@
import type { CandidateEvidenceItem, ProblemUnit } from "../types/stage2ProblemUnits";
import type { LifecycleDomain } from "../types/stage3Lifecycle";
import type {
AccountingGraphBuildResult,
AccountingGraphEdge,
AccountingGraphEdgeFlag,
AccountingGraphNode,
AccountingGraphRelationType,
AccountingGraphProvenance,
GraphConfidenceGrade,
GraphDomainKey,
GraphSignalSummary,
ProblemUnitGraphBinding
} from "../types/stage4Graph";
import { ACCOUNTING_GRAPH_SCHEMA_VERSION } from "../types/stage4Graph";
interface BuildAccountingGraphInput {
route: string;
candidateEvidence: CandidateEvidenceItem[];
problemUnits: ProblemUnit[];
}
interface GraphNodeCreateInput {
node_type: AccountingGraphNode["node_type"];
domain: GraphDomainKey;
stable_key: string;
label: string;
confidence: GraphConfidenceGrade;
attributes?: Record<string, unknown>;
provenance?: Partial<AccountingGraphProvenance>;
}
interface GraphEdgeCreateInput {
relation_type: AccountingGraphRelationType;
from_node_id: string;
to_node_id: string;
domain: GraphDomainKey;
confidence: GraphConfidenceGrade;
flags?: AccountingGraphEdgeFlag[];
provenance?: Partial<AccountingGraphProvenance>;
}
const GRAPH_CONFIDENCE_ORDER: Record<GraphConfidenceGrade, number> = {
low: 1,
medium: 2,
high: 3
};
const DOMAIN_PATH_HINTS: Record<LifecycleDomain, string[]> = {
bank_settlement: ["payment_to_settlement", "wrong_closing_document_type"],
customer_settlement: ["invoice_to_payment", "payment_to_closure"],
deferred_expense: ["deferred_expense_to_writeoff", "writeoff_sequence"],
fixed_asset: ["asset_card_to_depreciation", "card_document_register_alignment"],
vat_flow: ["invoice_to_vat_register", "cross_branch_alignment"],
period_close: ["period_close_dependency_chain", "closure_blocker_transition"]
};
function uniqueStrings(values: Array<string | null | undefined>, limit = 16): string[] {
return Array.from(new Set(values.map((item) => String(item ?? "").trim()).filter(Boolean))).slice(0, limit);
}
function compactToken(value: string): string {
const normalized = value.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-+|-+$/g, "");
return normalized.length > 0 ? normalized.slice(0, 48) : "x";
}
function stableNodeId(type: AccountingGraphNode["node_type"], domain: GraphDomainKey, stableKey: string): string {
return `gnd-${compactToken(type)}-${compactToken(domain)}-${compactToken(stableKey)}`;
}
function stableEdgeId(relation: AccountingGraphRelationType, fromNode: string, toNode: string): string {
return `ged-${compactToken(relation)}-${compactToken(fromNode)}-${compactToken(toNode)}`;
}
function mergeConfidence(left: GraphConfidenceGrade, right: GraphConfidenceGrade): GraphConfidenceGrade {
return GRAPH_CONFIDENCE_ORDER[right] > GRAPH_CONFIDENCE_ORDER[left] ? right : left;
}
function mergeProvenance(
left: AccountingGraphProvenance,
right: Partial<AccountingGraphProvenance> | undefined,
routeFallback: string
): AccountingGraphProvenance {
return {
route: String(right?.route ?? left.route ?? routeFallback),
candidate_ids: uniqueStrings([...(left.candidate_ids ?? []), ...(right?.candidate_ids ?? [])], 24),
evidence_ids: uniqueStrings([...(left.evidence_ids ?? []), ...(right?.evidence_ids ?? [])], 24)
};
}
class GraphAccumulator {
private readonly nodesById = new Map<string, AccountingGraphNode>();
private readonly edgesById = new Map<string, AccountingGraphEdge>();
constructor(private readonly route: string) {}
public upsertNode(input: GraphNodeCreateInput): AccountingGraphNode {
const node_id = stableNodeId(input.node_type, input.domain, input.stable_key);
const existing = this.nodesById.get(node_id);
if (!existing) {
const created: AccountingGraphNode = {
node_id,
node_type: input.node_type,
domain: input.domain,
label: input.label,
confidence: input.confidence,
attributes: input.attributes ?? {},
provenance: {
route: String(input.provenance?.route ?? this.route),
candidate_ids: uniqueStrings(input.provenance?.candidate_ids ?? [], 24),
evidence_ids: uniqueStrings(input.provenance?.evidence_ids ?? [], 24)
}
};
this.nodesById.set(node_id, created);
return created;
}
existing.confidence = mergeConfidence(existing.confidence, input.confidence);
existing.provenance = mergeProvenance(existing.provenance, input.provenance, this.route);
if (Object.keys(input.attributes ?? {}).length > 0) {
existing.attributes = {
...existing.attributes,
...input.attributes
};
}
return existing;
}
public upsertEdge(input: GraphEdgeCreateInput): AccountingGraphEdge {
const edge_id = stableEdgeId(input.relation_type, input.from_node_id, input.to_node_id);
const existing = this.edgesById.get(edge_id);
if (!existing) {
const created: AccountingGraphEdge = {
edge_id,
relation_type: input.relation_type,
from_node_id: input.from_node_id,
to_node_id: input.to_node_id,
domain: input.domain,
confidence: input.confidence,
flags: uniqueStrings(input.flags ?? [], 8) as AccountingGraphEdgeFlag[],
provenance: {
route: String(input.provenance?.route ?? this.route),
candidate_ids: uniqueStrings(input.provenance?.candidate_ids ?? [], 24),
evidence_ids: uniqueStrings(input.provenance?.evidence_ids ?? [], 24)
}
};
this.edgesById.set(edge_id, created);
return created;
}
existing.confidence = mergeConfidence(existing.confidence, input.confidence);
existing.flags = uniqueStrings([...(existing.flags ?? []), ...(input.flags ?? [])], 8) as AccountingGraphEdgeFlag[];
existing.provenance = mergeProvenance(existing.provenance, input.provenance, this.route);
return existing;
}
public export(): { nodes: AccountingGraphNode[]; edges: AccountingGraphEdge[] } {
return {
nodes: Array.from(this.nodesById.values()).sort((left, right) => left.node_id.localeCompare(right.node_id)),
edges: Array.from(this.edgesById.values()).sort((left, right) => left.edge_id.localeCompare(right.edge_id))
};
}
}
function inferDomainFromUnit(unit: ProblemUnit): GraphDomainKey {
if (unit.lifecycle_domain) {
return unit.lifecycle_domain;
}
const accountText = unit.affected_accounts.join(" ").toLowerCase();
if (/\b97\b/.test(accountText) || unit.problem_unit_type === "lifecycle_anomaly_node") {
return "deferred_expense";
}
if (/\b(01|02|08)\b/.test(accountText)) {
return "fixed_asset";
}
if (/\b(19|68)\b/.test(accountText) || unit.problem_unit_type === "cross_branch_inconsistency_cluster") {
return "vat_flow";
}
if (unit.problem_unit_type === "period_risk_cluster" || unit.period_impact?.impact_class === "close_risk") {
return "period_close";
}
if (/\b62\b/.test(accountText)) {
return "customer_settlement";
}
if (/\b(51|60|76)\b/.test(accountText) || unit.problem_unit_type === "unresolved_settlement_cluster") {
return "bank_settlement";
}
return "unknown";
}
function relationPathHints(domain: GraphDomainKey, unit: ProblemUnit): string[] {
const path: string[] = [`domain:${domain}`];
if (unit.current_lifecycle_state && unit.expected_lifecycle_state) {
path.push(`state:${unit.current_lifecycle_state}->${unit.expected_lifecycle_state}`);
} else if (unit.current_lifecycle_state) {
path.push(`state:${unit.current_lifecycle_state}`);
}
if (domain !== "unknown") {
path.push(...(DOMAIN_PATH_HINTS[domain] ?? []));
}
if (unit.missing_transition) {
path.push(`missing:${unit.missing_transition}`);
}
if (unit.invalid_transition) {
path.push(`conflict:${unit.invalid_transition}`);
}
return uniqueStrings(path, 10);
}
function graphConfidenceFromUnit(unit: ProblemUnit): GraphConfidenceGrade {
return unit.lifecycle_confidence?.grade ?? unit.confidence.grade;
}
function coverageGrade(boundUnits: number, totalUnits: number): GraphConfidenceGrade {
if (totalUnits <= 0) {
return "low";
}
const ratio = boundUnits / totalUnits;
if (ratio >= 0.8) return "high";
if (ratio >= 0.4) return "medium";
return "low";
}
function buildSummary(input: {
nodes: AccountingGraphNode[];
edges: AccountingGraphEdge[];
bindings: ProblemUnitGraphBinding[];
totalUnits: number;
}): GraphSignalSummary {
const domain_distribution: GraphSignalSummary["domain_distribution"] = {};
const relation_distribution: GraphSignalSummary["relation_distribution"] = {};
for (const binding of input.bindings) {
const domainMarker = binding.relation_path.find((item) => item.startsWith("domain:")) ?? "domain:unknown";
const domainKey = domainMarker.replace(/^domain:/, "") as GraphDomainKey;
domain_distribution[domainKey] = (domain_distribution[domainKey] ?? 0) + 1;
}
for (const edge of input.edges) {
relation_distribution[edge.relation_type] = (relation_distribution[edge.relation_type] ?? 0) + 1;
}
const missing_links_count = input.bindings.reduce((acc, item) => acc + item.missing_links.length, 0);
const conflicting_links_count = input.bindings.reduce((acc, item) => acc + item.conflicting_links.length, 0);
const bound_units = input.bindings.length;
return {
total_units: input.totalUnits,
bound_units,
node_count: input.nodes.length,
edge_count: input.edges.length,
missing_links_count,
conflicting_links_count,
graph_coverage_grade: coverageGrade(bound_units, input.totalUnits),
domain_distribution,
relation_distribution
};
}
export function buildAccountingGraph(input: BuildAccountingGraphInput): AccountingGraphBuildResult {
const accumulator = new GraphAccumulator(input.route);
const candidateById = new Map(input.candidateEvidence.map((item) => [item.candidate_id, item] as const));
const bindings: ProblemUnitGraphBinding[] = [];
const issues: string[] = [];
if (input.problemUnits.length === 0) {
issues.push("no_problem_units_for_graph_build");
}
for (const unit of input.problemUnits) {
const domain = inferDomainFromUnit(unit);
const confidence = graphConfidenceFromUnit(unit);
const candidateIds = uniqueStrings(unit.evidence_pack, 12).filter((item) => candidateById.has(item));
const evidenceIds = uniqueStrings(unit.evidence_pack, 12);
const domainNode = accumulator.upsertNode({
node_type: "domain",
domain,
stable_key: `domain:${domain}`,
label: domain,
confidence,
attributes: {
domain
},
provenance: {
route: input.route,
candidate_ids: candidateIds,
evidence_ids: evidenceIds
}
});
const problemNode = accumulator.upsertNode({
node_type: "problem_unit",
domain,
stable_key: `problem:${unit.problem_unit_id}`,
label: unit.title || unit.problem_unit_id,
confidence,
attributes: {
problem_unit_id: unit.problem_unit_id,
problem_unit_type: unit.problem_unit_type,
business_defect_class: unit.business_defect_class,
lifecycle_defect_type: unit.lifecycle_defect_type ?? null
},
provenance: {
route: input.route,
candidate_ids: candidateIds,
evidence_ids: evidenceIds
}
});
accumulator.upsertEdge({
relation_type: "belongs_to_domain",
from_node_id: problemNode.node_id,
to_node_id: domainNode.node_id,
domain,
confidence,
flags: ["actual_link"],
provenance: {
route: input.route,
candidate_ids: candidateIds,
evidence_ids: evidenceIds
}
});
for (const account of uniqueStrings(unit.affected_accounts, 4)) {
const accountNode = accumulator.upsertNode({
node_type: "account",
domain,
stable_key: `account:${account}`,
label: account,
confidence,
attributes: {
account
},
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
accumulator.upsertEdge({
relation_type: "affects_account",
from_node_id: problemNode.node_id,
to_node_id: accountNode.node_id,
domain,
confidence,
flags: ["actual_link"],
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
}
for (const document of uniqueStrings(unit.affected_documents, 3)) {
const documentNode = accumulator.upsertNode({
node_type: "document",
domain,
stable_key: `document:${document}`,
label: document,
confidence,
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
accumulator.upsertEdge({
relation_type: "affects_document",
from_node_id: problemNode.node_id,
to_node_id: documentNode.node_id,
domain,
confidence,
flags: ["actual_link"],
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
}
for (const counterparty of uniqueStrings(unit.affected_counterparties, 3)) {
const counterpartyNode = accumulator.upsertNode({
node_type: "counterparty",
domain,
stable_key: `counterparty:${counterparty}`,
label: counterparty,
confidence,
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
accumulator.upsertEdge({
relation_type: "affects_counterparty",
from_node_id: problemNode.node_id,
to_node_id: counterpartyNode.node_id,
domain,
confidence,
flags: ["actual_link"],
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
}
if (unit.current_lifecycle_state) {
const currentNode = accumulator.upsertNode({
node_type: "lifecycle_state",
domain,
stable_key: `current_state:${unit.current_lifecycle_state}`,
label: unit.current_lifecycle_state,
confidence,
attributes: {
state_role: "current"
},
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
accumulator.upsertEdge({
relation_type: "current_state",
from_node_id: problemNode.node_id,
to_node_id: currentNode.node_id,
domain,
confidence,
flags: ["actual_link"],
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
}
if (unit.expected_lifecycle_state) {
const expectedNode = accumulator.upsertNode({
node_type: "lifecycle_state",
domain,
stable_key: `expected_state:${unit.expected_lifecycle_state}`,
label: unit.expected_lifecycle_state,
confidence,
attributes: {
state_role: "expected"
},
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
accumulator.upsertEdge({
relation_type: "expected_state",
from_node_id: problemNode.node_id,
to_node_id: expectedNode.node_id,
domain,
confidence,
flags: ["expected_link"],
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
}
if (unit.missing_transition) {
const missingNode = accumulator.upsertNode({
node_type: "transition",
domain,
stable_key: `missing_transition:${unit.missing_transition}`,
label: unit.missing_transition,
confidence,
attributes: {
transition_role: "missing"
},
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
accumulator.upsertEdge({
relation_type: "missing_transition",
from_node_id: problemNode.node_id,
to_node_id: missingNode.node_id,
domain,
confidence,
flags: ["missing_link"],
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
}
if (unit.invalid_transition) {
const invalidNode = accumulator.upsertNode({
node_type: "transition",
domain,
stable_key: `invalid_transition:${unit.invalid_transition}`,
label: unit.invalid_transition,
confidence,
attributes: {
transition_role: "invalid"
},
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
accumulator.upsertEdge({
relation_type: "invalid_transition",
from_node_id: problemNode.node_id,
to_node_id: invalidNode.node_id,
domain,
confidence,
flags: ["conflict_link"],
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
}
if (unit.lifecycle_defect_type) {
const defectNode = accumulator.upsertNode({
node_type: "defect",
domain,
stable_key: `defect:${unit.lifecycle_defect_type}`,
label: unit.lifecycle_defect_type,
confidence,
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
accumulator.upsertEdge({
relation_type: "has_defect",
from_node_id: problemNode.node_id,
to_node_id: defectNode.node_id,
domain,
confidence,
flags: unit.lifecycle_defect_type === "cross_branch_state_conflict" ? ["conflict_link"] : ["actual_link"],
provenance: {
route: input.route,
evidence_ids: evidenceIds
}
});
}
for (const candidateId of candidateIds.slice(0, 6)) {
const candidate = candidateById.get(candidateId);
const evidenceLabel = candidate
? `${candidate.source_ref.entity}:${candidate.source_ref.id}`
: `candidate:${candidateId}`;
const evidenceNode = accumulator.upsertNode({
node_type: "evidence",
domain,
stable_key: `evidence:${candidateId}`,
label: evidenceLabel,
confidence,
attributes: {
candidate_id: candidateId
},
provenance: {
route: input.route,
candidate_ids: [candidateId],
evidence_ids: [candidateId]
}
});
accumulator.upsertEdge({
relation_type: "supported_by_evidence",
from_node_id: problemNode.node_id,
to_node_id: evidenceNode.node_id,
domain,
confidence,
flags: ["actual_link"],
provenance: {
route: input.route,
candidate_ids: [candidateId],
evidence_ids: [candidateId]
}
});
if (candidate && candidate.relation_pattern_hits.length > 0) {
for (const relationHint of uniqueStrings(candidate.relation_pattern_hits, 2)) {
const hintNode = accumulator.upsertNode({
node_type: "transition",
domain,
stable_key: `hint:${relationHint}`,
label: relationHint,
confidence,
attributes: {
transition_role: "hint"
},
provenance: {
route: input.route,
candidate_ids: [candidateId],
evidence_ids: [candidateId]
}
});
accumulator.upsertEdge({
relation_type: "supports_path",
from_node_id: evidenceNode.node_id,
to_node_id: hintNode.node_id,
domain,
confidence,
flags: ["actual_link"],
provenance: {
route: input.route,
candidate_ids: [candidateId],
evidence_ids: [candidateId]
}
});
}
}
}
const missing_links = uniqueStrings([
unit.missing_transition,
...(unit.lifecycle_resolution?.missing_transitions ?? [])
]);
const conflicting_links = uniqueStrings([
unit.invalid_transition,
...(unit.lifecycle_resolution?.invalid_transitions ?? [])
]);
bindings.push({
problem_unit_id: unit.problem_unit_id,
graph_node_id: problemNode.node_id,
relation_path: relationPathHints(domain, unit),
missing_links,
conflicting_links,
provenance_evidence_ids: evidenceIds,
graph_confidence: confidence
});
}
const exported = accumulator.export();
const summary = buildSummary({
nodes: exported.nodes,
edges: exported.edges,
bindings,
totalUnits: input.problemUnits.length
});
if (summary.bound_units < summary.total_units) {
issues.push("some_problem_units_not_bound_to_graph");
}
if (summary.node_count === 0 || summary.edge_count === 0) {
issues.push("graph_runtime_empty");
}
return {
schema_version: ACCOUNTING_GRAPH_SCHEMA_VERSION,
nodes: exported.nodes,
edges: exported.edges,
unit_bindings: bindings,
summary,
issues: uniqueStrings(issues, 8)
};
}
@@ -52,10 +52,22 @@ function redactSecrets(payload: Record<string, unknown>): Record<string, unknown
return output;
}
function isNoSpaceError(error: unknown): boolean {
const code = (error as { code?: unknown } | null)?.code;
return code === "ENOSPC";
}
export function saveTrace(record: TraceRecord): void {
ensureDir(TRACES_DIR);
const target = path.resolve(TRACES_DIR, `${record.trace_id}.json`);
writeJsonFile(target, record);
try {
ensureDir(TRACES_DIR);
const target = path.resolve(TRACES_DIR, `${record.trace_id}.json`);
writeJsonFile(target, record);
} catch (error) {
if (isNoSpaceError(error)) {
return;
}
throw error;
}
}
export function listTraces(limit = 100): HistoryListItem[] {
@@ -97,8 +109,15 @@ export function getTrace(traceId: string): TraceRecord | null {
}
export function savePreset(preset: PromptPreset): void {
ensureDir(PRESETS_DIR);
writeJsonFile(path.resolve(PRESETS_DIR, `${preset.id}.json`), preset);
try {
ensureDir(PRESETS_DIR);
writeJsonFile(path.resolve(PRESETS_DIR, `${preset.id}.json`), preset);
} catch (error) {
if (isNoSpaceError(error)) {
return;
}
throw error;
}
}
export function listPresets(): PromptPreset[] {
@@ -114,9 +133,15 @@ export function listPresets(): PromptPreset[] {
}
export function saveEvalCase(casePayload: Record<string, unknown>): string {
ensureDir(EVAL_CASES_DIR);
const id = String(casePayload.case_id ?? `NQ-${Date.now()}`);
writeJsonFile(path.resolve(EVAL_CASES_DIR, `${id}.json`), casePayload);
try {
ensureDir(EVAL_CASES_DIR);
writeJsonFile(path.resolve(EVAL_CASES_DIR, `${id}.json`), casePayload);
} catch (error) {
if (!isNoSpaceError(error)) {
throw error;
}
}
return id;
}
@@ -6,6 +6,7 @@ import type {
ProblemUnit,
ProblemUnitSummary
} from "./stage2ProblemUnits";
import type { AccountingGraphBuildResult } from "./stage4Graph";
export type AssistantFallbackType = "none" | "out_of_scope" | "clarification" | "partial" | "unknown";
export type AssistantReplyType =
@@ -99,6 +100,7 @@ export interface UnifiedRetrievalResult {
candidate_evidence?: CandidateEvidenceItem[];
problem_units?: ProblemUnit[];
problem_unit_summary?: ProblemUnitSummary | null;
accounting_graph?: AccountingGraphBuildResult;
summary: Record<string, unknown>;
evidence: EvidenceItem[];
why_included: string[];
@@ -138,6 +140,10 @@ export interface AssistantDebugPayload {
answer_structure_v11: AnswerStructureV11 | null;
investigation_state_snapshot: InvestigationStateWithProblemUnits | null;
normalized: NormalizeResponsePayload["normalized"];
normalizer_output?: {
contains_multiple_tasks?: boolean;
scope_confidence?: string | null;
} | null;
}
export interface AssistantConversationItem {
@@ -1,7 +1,7 @@
import type { AssistantEvalBroadnessLevel, AssistantEvalQuestionType } from "./stage1Contracts";
import type { ProblemUnitType } from "./stage2ProblemUnits";
export type EvalTarget = "normalizer" | "assistant_stage1" | "assistant_stage2";
export type EvalTarget = "normalizer" | "assistant_stage1" | "assistant_stage2" | "assistant_p0";
export interface AssistantStage1SuiteCaseTurn {
user_message: string;
@@ -33,6 +33,12 @@ export interface InvestigationFollowupContext {
previous_question_id: string | null;
last_user_message: string;
referenced_requirement_ids: string[];
active_domain?: string | null;
active_requirement_ids?: string[];
uncovered_requirement_ids?: string[];
last_problem_unit_id?: string | null;
settlement_next_actions?: string[];
evidence_summary?: string[];
}
export interface InvestigationState {
@@ -1,6 +1,7 @@
import type { InvestigationState } from "./stage1Contracts";
import type { EvidenceSourceRef } from "./stage1Contracts";
import type { LifecycleConfidence, LifecycleDefectType, LifecycleDomain, LifecycleResolution } from "./stage3Lifecycle";
import type { GraphSignalSummary, ProblemUnitGraphBinding } from "./stage4Graph";
export const CANDIDATE_EVIDENCE_SCHEMA_VERSION = "candidate_evidence_v0_1" as const;
export const PROBLEM_UNIT_SCHEMA_VERSION = "problem_unit_v0_1" as const;
@@ -89,6 +90,7 @@ export interface ProblemUnit {
lifecycle_resolution?: LifecycleResolution;
lifecycle_ranking_score?: number;
lifecycle_ranking_basis?: string[];
graph_binding?: ProblemUnitGraphBinding;
}
export interface ProblemUnitSummary {
@@ -103,6 +105,7 @@ export interface ProblemUnitSummary {
lifecycle_enriched_units?: number;
lifecycle_domain_distribution?: Partial<Record<LifecycleDomain, number>>;
lifecycle_defect_distribution?: Partial<Record<LifecycleDefectType, number>>;
graph_summary?: GraphSignalSummary;
}
export interface InvestigationProblemUnitState {
@@ -0,0 +1,95 @@
import type { LifecycleDomain } from "./stage3Lifecycle";
export const ACCOUNTING_GRAPH_SCHEMA_VERSION = "accounting_graph_v0_1" as const;
export type GraphConfidenceGrade = "low" | "medium" | "high";
export type GraphDomainKey = LifecycleDomain | "unknown";
export const ACCOUNTING_GRAPH_NODE_TYPES = [
"domain",
"problem_unit",
"account",
"document",
"counterparty",
"lifecycle_state",
"transition",
"defect",
"evidence"
] as const;
export type AccountingGraphNodeType = (typeof ACCOUNTING_GRAPH_NODE_TYPES)[number];
export const ACCOUNTING_GRAPH_RELATION_TYPES = [
"belongs_to_domain",
"affects_account",
"affects_document",
"affects_counterparty",
"current_state",
"expected_state",
"missing_transition",
"invalid_transition",
"has_defect",
"supported_by_evidence",
"supports_path"
] as const;
export type AccountingGraphRelationType = (typeof ACCOUNTING_GRAPH_RELATION_TYPES)[number];
export type AccountingGraphEdgeFlag = "expected_link" | "actual_link" | "missing_link" | "conflict_link";
export interface AccountingGraphProvenance {
route: string;
candidate_ids: string[];
evidence_ids: string[];
}
export interface AccountingGraphNode {
node_id: string;
node_type: AccountingGraphNodeType;
label: string;
domain: GraphDomainKey;
confidence: GraphConfidenceGrade;
attributes: Record<string, unknown>;
provenance: AccountingGraphProvenance;
}
export interface AccountingGraphEdge {
edge_id: string;
relation_type: AccountingGraphRelationType;
from_node_id: string;
to_node_id: string;
domain: GraphDomainKey;
confidence: GraphConfidenceGrade;
flags: AccountingGraphEdgeFlag[];
provenance: AccountingGraphProvenance;
}
export interface ProblemUnitGraphBinding {
problem_unit_id: string;
graph_node_id: string;
relation_path: string[];
missing_links: string[];
conflicting_links: string[];
provenance_evidence_ids: string[];
graph_confidence: GraphConfidenceGrade;
}
export interface GraphSignalSummary {
total_units: number;
bound_units: number;
node_count: number;
edge_count: number;
missing_links_count: number;
conflicting_links_count: number;
graph_coverage_grade: GraphConfidenceGrade;
domain_distribution: Partial<Record<GraphDomainKey, number>>;
relation_distribution: Partial<Record<AccountingGraphRelationType, number>>;
}
export interface AccountingGraphBuildResult {
schema_version: typeof ACCOUNTING_GRAPH_SCHEMA_VERSION;
nodes: AccountingGraphNode[];
edges: AccountingGraphEdge[];
unit_bindings: ProblemUnitGraphBinding[];
summary: GraphSignalSummary;
issues: string[];
}
+9 -1
View File
@@ -37,5 +37,13 @@ export function logJson(entry: JsonLogEntry): void {
details: redactObject(entry.details)
};
// Structured JSON logs for diagnostics/trace aggregation.
process.stdout.write(JSON.stringify(safe) + "\n");
try {
process.stdout.write(JSON.stringify(safe) + "\n");
} catch (error) {
const code = (error as { code?: unknown } | null)?.code;
if (code === "ENOSPC" || code === "EPIPE") {
return;
}
throw error;
}
}