Initial import NDC_1C

This commit is contained in:
2026-03-26 10:38:25 +03:00
commit a162d77ef7
2943 changed files with 3615871 additions and 0 deletions
+2
View File
@@ -0,0 +1,2 @@
"""Router and orchestration decision helpers for accounting analytics flows."""
+51
View File
@@ -0,0 +1,51 @@
from __future__ import annotations
from dataclasses import asdict, dataclass
from typing import Any
from router.query_classifier import RouteDecisionFlags
from router.route_selector import RouteSelectionResult
from router.store_sufficiency import StoreSufficiencyResult
@dataclass
class RouteDecisionLog:
question_id: str
question_text: str
parsed_class: str
decision_flags: dict[str, Any]
sufficiency_snapshot: dict[str, Any]
candidate_routes: list[str]
rejected_routes: dict[str, str]
chosen_route: str
execution_mode: str
batch_job_id: str | None
def to_dict(self) -> dict[str, Any]:
return asdict(self)
def build_route_decision_log(
*,
question_id: str,
question_text: str,
parsed_class: str,
flags: RouteDecisionFlags,
suff: StoreSufficiencyResult,
selection: RouteSelectionResult,
execution_mode: str,
batch_job_id: str | None,
) -> RouteDecisionLog:
return RouteDecisionLog(
question_id=question_id,
question_text=question_text,
parsed_class=parsed_class,
decision_flags=flags.to_dict(),
sufficiency_snapshot=suff.to_dict(),
candidate_routes=selection.candidate_routes,
rejected_routes=selection.rejected_routes,
chosen_route=selection.chosen_route,
execution_mode=execution_mode,
batch_job_id=batch_job_id,
)
+177
View File
@@ -0,0 +1,177 @@
from __future__ import annotations
from dataclasses import asdict, dataclass
import re
from typing import Any
ACCOUNT_TOKEN_RE = re.compile(r"\b\d{2}(?:\.\d{2})?\b")
@dataclass
class RouteDecisionFlags:
needs_exact_object_trace: bool
needs_causal_chain: bool
needs_cross_entity_join: bool
needs_full_period_aggregation: bool
needs_ranking: bool
needs_anomaly_summary: bool
needs_runtime_truth: bool
freshness_sensitive: bool
ambiguous_object_scope: bool
store_sufficiency_confident: bool
precomputed_aggregate_available: bool
def to_dict(self) -> dict[str, Any]:
return asdict(self)
def _norm(text: str) -> str:
return text.lower().strip()
def _contains_any(text: str, tokens: list[str]) -> bool:
return any(token in text for token in tokens)
def _has_account_token(text: str) -> bool:
return bool(ACCOUNT_TOKEN_RE.search(text))
def _aggregate_available_for_shape(
*,
available: set[str],
needs_ranking: bool,
needs_anomaly_summary: bool,
needs_full_period_aggregation: bool,
text: str,
) -> bool:
if needs_ranking:
ranking_tokens = {
"risk_account_ranking",
"risk_counterparty_ranking",
"risk_ranking",
}
return bool(available.intersection(ranking_tokens))
if needs_anomaly_summary:
anomaly_tokens = {
"company_anomaly_summary",
}
return bool(available.intersection(anomaly_tokens))
if needs_full_period_aggregation:
if "baseline" in text:
return "baseline_period_summary" in available
return "full_period_aggregation" in available
return False
def classify_query_for_route(
question_text: str,
parsed_intent: dict[str, Any],
store_metadata: dict[str, Any],
) -> RouteDecisionFlags:
text = _norm(question_text)
question_class = str(parsed_intent.get("question_class", "")).strip().lower()
exact_markers = [
"документ по номеру",
"source-of-record",
"источник",
"цепочка",
"почему",
"subconto3",
"субконто3",
]
causal_markers = [
"свяжи",
"цепочка",
"через",
"объясни",
"почему",
"источник",
"регистр",
"первич",
]
cross_markers = [
"свяжи",
"документ",
"провод",
"контрагент",
"договор",
"регистр",
]
ranking_markers = ["рейтинг", "ranking", "топ", "top"]
anomaly_markers = ["аномал", "summary", "срез", "risk-slice", "риск-срез"]
needs_exact_object_trace = _contains_any(text, exact_markers) and (
question_class in {"drilldown_explain", "simple_factual", "cross_entity"}
)
if question_class == "simple_factual" and "документ по номеру" in text:
needs_exact_object_trace = True
needs_causal_chain = _contains_any(text, causal_markers) and question_class in {
"drilldown_explain",
"cross_entity",
}
needs_cross_entity_join = (
question_class == "cross_entity"
or (_contains_any(text, cross_markers) and " и " in text and "->" not in text)
)
needs_ranking = _contains_any(text, ranking_markers) and question_class in {
"heavy_analytical",
"period_trend",
"anomaly_control",
}
needs_anomaly_summary = _contains_any(text, anomaly_markers)
is_heavy = question_class == "heavy_analytical"
is_baseline_heavy = is_heavy and "baseline" in text
needs_full_period_aggregation = is_heavy and not is_baseline_heavy
needs_runtime_truth = needs_exact_object_trace or _contains_any(
text, ["runtime", "source-of-record", "источник регистра"]
)
freshness_sensitive = question_class in {
"period_trend",
"anomaly_control",
"heavy_analytical",
}
ambiguous_object_scope = question_class == "ambiguous_fuzzy"
if ambiguous_object_scope and _has_account_token(text):
# Ambiguous account prompts should avoid hard downcast into canonical-only answers.
needs_runtime_truth = True
available_aggregates = {
str(item).strip().lower() for item in store_metadata.get("precomputed_aggregates", [])
}
precomputed_aggregate_available = _aggregate_available_for_shape(
available=available_aggregates,
needs_ranking=needs_ranking,
needs_anomaly_summary=needs_anomaly_summary,
needs_full_period_aggregation=needs_full_period_aggregation,
text=text,
)
store_sufficiency_confident = (
question_class == "simple_factual"
and not needs_runtime_truth
and not needs_causal_chain
and not needs_cross_entity_join
)
return RouteDecisionFlags(
needs_exact_object_trace=needs_exact_object_trace,
needs_causal_chain=needs_causal_chain,
needs_cross_entity_join=needs_cross_entity_join,
needs_full_period_aggregation=needs_full_period_aggregation,
needs_ranking=needs_ranking,
needs_anomaly_summary=needs_anomaly_summary,
needs_runtime_truth=needs_runtime_truth,
freshness_sensitive=freshness_sensitive,
ambiguous_object_scope=ambiguous_object_scope,
store_sufficiency_confident=store_sufficiency_confident,
precomputed_aggregate_available=precomputed_aggregate_available,
)
+113
View File
@@ -0,0 +1,113 @@
from __future__ import annotations
from dataclasses import asdict, dataclass
from typing import Any
from router.query_classifier import RouteDecisionFlags
from router.store_sufficiency import StoreSufficiencyResult
ALL_ROUTES = [
"live_mcp_drilldown",
"store_canonical",
"store_feature_risk",
"hybrid_store_plus_live",
"batch_refresh_then_store",
]
@dataclass
class RouteSelectionResult:
chosen_route: str
candidate_routes: list[str]
rejected_routes: dict[str, str]
def to_dict(self) -> dict[str, Any]:
return asdict(self)
def choose_route(
flags: RouteDecisionFlags,
suff: StoreSufficiencyResult,
*,
parsed_as_trend_or_risk: bool,
) -> RouteSelectionResult:
rejected: dict[str, str] = {}
if flags.needs_exact_object_trace:
rejected["store_canonical"] = "exact_trace_requires_live"
rejected["store_feature_risk"] = "exact_trace_requires_live"
rejected["hybrid_store_plus_live"] = "exact_trace_prefers_direct_live"
rejected["batch_refresh_then_store"] = "exact_trace_not_batch"
return RouteSelectionResult(
chosen_route="live_mcp_drilldown",
candidate_routes=list(ALL_ROUTES),
rejected_routes=rejected,
)
heavy_shape = (
flags.needs_full_period_aggregation
or flags.needs_ranking
or (flags.needs_anomaly_summary and not parsed_as_trend_or_risk)
)
if heavy_shape:
aggregate_ok = (
flags.precomputed_aggregate_available
and suff.freshness_ok
and suff.aggregate_level_ok
and (not flags.needs_ranking or suff.ranking_ready)
)
if not aggregate_ok:
rejected["store_feature_risk"] = "aggregate_not_sufficient"
rejected["store_canonical"] = "wrong_query_shape"
rejected["live_mcp_drilldown"] = "heavy_query_not_live"
return RouteSelectionResult(
chosen_route="batch_refresh_then_store",
candidate_routes=list(ALL_ROUTES),
rejected_routes=rejected,
)
if flags.needs_cross_entity_join and flags.needs_causal_chain:
if not suff.explanation_ready:
rejected["store_canonical"] = "cross_entity_causal_needs_live_stitching"
return RouteSelectionResult(
chosen_route="hybrid_store_plus_live",
candidate_routes=list(ALL_ROUTES),
rejected_routes=rejected,
)
if parsed_as_trend_or_risk:
if suff.feature_sufficient and not flags.needs_runtime_truth:
rejected["store_canonical"] = "trend_risk_query_prefers_feature"
return RouteSelectionResult(
chosen_route="store_feature_risk",
candidate_routes=list(ALL_ROUTES),
rejected_routes=rejected,
)
rejected["store_feature_risk"] = "feature_or_freshness_not_sufficient"
if (
suff.canonical_sufficient
and not flags.needs_causal_chain
and not flags.ambiguous_object_scope
and not flags.needs_runtime_truth
):
if flags.needs_cross_entity_join and not suff.explanation_ready:
rejected["store_canonical"] = "cross_entity_needs_explanation_stitching"
else:
return RouteSelectionResult(
chosen_route="store_canonical",
candidate_routes=list(ALL_ROUTES),
rejected_routes=rejected,
)
if not suff.canonical_sufficient:
rejected["store_canonical"] = "canonical_not_sufficient"
if flags.ambiguous_object_scope:
rejected["store_canonical"] = "ambiguous_scope_requires_hybrid_or_feature"
return RouteSelectionResult(
chosen_route="hybrid_store_plus_live",
candidate_routes=list(ALL_ROUTES),
rejected_routes=rejected,
)
+111
View File
@@ -0,0 +1,111 @@
from __future__ import annotations
from dataclasses import asdict, dataclass
from typing import Any
from router.query_classifier import RouteDecisionFlags
@dataclass
class StoreSufficiencyResult:
canonical_sufficient: bool
feature_sufficient: bool
risk_sufficient: bool
freshness_ok: bool
aggregate_level_ok: bool
ranking_ready: bool
explanation_ready: bool
reason_codes: list[str]
def to_dict(self) -> dict[str, Any]:
return asdict(self)
def _to_float(value: Any, default: float = 0.0) -> float:
try:
return float(value)
except (TypeError, ValueError):
return default
def check_store_sufficiency(
question_shape: RouteDecisionFlags,
store_metadata: dict[str, Any],
) -> StoreSufficiencyResult:
reason_codes: list[str] = []
freshness_threshold_hours = _to_float(store_metadata.get("freshness_threshold_hours", 6.0), default=6.0)
refresh_age = _to_float(store_metadata.get("refresh_age_hours", 0.0), default=0.0)
feature_age = _to_float(store_metadata.get("feature_age_hours", refresh_age), default=refresh_age)
risk_age = _to_float(store_metadata.get("risk_age_hours", refresh_age), default=refresh_age)
canonical_semantic_coverage = _to_float(store_metadata.get("canonical_semantic_coverage", 0.0), default=0.0)
canonical_relation_types = int(store_metadata.get("canonical_relation_types", 0))
canonical_links_total = int(store_metadata.get("canonical_links_total", 0))
canonical_entities_total = int(store_metadata.get("canonical_entities_total", 0))
feature_ready = bool(store_metadata.get("feature_ready", False))
risk_ready = bool(store_metadata.get("risk_ready", False))
ranking_ready = bool(store_metadata.get("ranking_ready", False))
aggregate_ready = bool(store_metadata.get("aggregate_ready", False))
freshness_ok = refresh_age <= freshness_threshold_hours
if question_shape.freshness_sensitive and not freshness_ok:
reason_codes.append("refresh_stale")
canonical_sufficient = (
canonical_entities_total > 0
and canonical_links_total > 0
and canonical_semantic_coverage >= 0.85
and canonical_relation_types >= 10
and (freshness_ok or not question_shape.freshness_sensitive)
)
if not canonical_sufficient:
reason_codes.append("canonical_not_sufficient")
feature_sufficient = feature_ready and (
feature_age <= freshness_threshold_hours or not question_shape.freshness_sensitive
)
if not feature_sufficient and question_shape.needs_anomaly_summary:
reason_codes.append("feature_not_sufficient")
risk_sufficient = risk_ready and (
risk_age <= freshness_threshold_hours or not question_shape.freshness_sensitive
)
if not risk_sufficient and (question_shape.needs_anomaly_summary or question_shape.needs_ranking):
reason_codes.append("risk_not_sufficient")
if question_shape.needs_ranking:
if not ranking_ready:
reason_codes.append("ranking_not_ready")
aggregate_level_ok = aggregate_ready and (
(not question_shape.needs_ranking or ranking_ready)
) and question_shape.precomputed_aggregate_available
if not aggregate_level_ok and (
question_shape.needs_full_period_aggregation
or question_shape.needs_ranking
or question_shape.needs_anomaly_summary
):
reason_codes.append("aggregate_not_sufficient")
explanation_ready = (
canonical_sufficient
and canonical_semantic_coverage >= 0.90
and canonical_relation_types >= 20
and not question_shape.needs_runtime_truth
and not (question_shape.needs_cross_entity_join and question_shape.needs_causal_chain)
)
if not explanation_ready and (question_shape.needs_causal_chain or question_shape.needs_cross_entity_join):
reason_codes.append("explanation_not_ready")
return StoreSufficiencyResult(
canonical_sufficient=canonical_sufficient,
feature_sufficient=feature_sufficient,
risk_sufficient=risk_sufficient,
freshness_ok=freshness_ok,
aggregate_level_ok=aggregate_level_ok,
ranking_ready=ranking_ready,
explanation_ready=explanation_ready,
reason_codes=reason_codes,
)