dandelion-memory 1.0.0rc15__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- benchmarks/__init__.py +3 -0
- benchmarks/adapters/__init__.py +29 -0
- benchmarks/adapters/ablations.py +406 -0
- benchmarks/adapters/base.py +72 -0
- benchmarks/adapters/controlled/__init__.py +13 -0
- benchmarks/adapters/controlled/core_archival.py +115 -0
- benchmarks/adapters/controlled/extract_resolve.py +130 -0
- benchmarks/adapters/controlled/hybrid_rag.py +120 -0
- benchmarks/adapters/controlled/no_late_binding.py +113 -0
- benchmarks/adapters/dandelion.py +94 -0
- benchmarks/adapters/native/__init__.py +31 -0
- benchmarks/adapters/native/base.py +45 -0
- benchmarks/adapters/native/dandelion_native.py +164 -0
- benchmarks/adapters/native/dandelion_native_b0.py +165 -0
- benchmarks/adapters/native/dandelion_native_b1.py +170 -0
- benchmarks/adapters/native/dandelion_native_b2.py +169 -0
- benchmarks/adapters/native/dandelion_native_b3.py +171 -0
- benchmarks/adapters/native/dandelion_native_c4.py +261 -0
- benchmarks/adapters/native/dandelion_native_c4_1.py +281 -0
- benchmarks/adapters/native/dandelion_native_c4_2.py +435 -0
- benchmarks/adapters/native/event_evidence_b2.py +616 -0
- benchmarks/adapters/native/evidence_b2.py +609 -0
- benchmarks/adapters/native/letta_native.py +253 -0
- benchmarks/adapters/native/mem0_native.py +219 -0
- benchmarks/adapters/native/store_vector_b0.py +792 -0
- benchmarks/adapters/native/store_vector_b2.py +808 -0
- benchmarks/agent_adapter_stdio_acceptance.py +86 -0
- benchmarks/bootstrap_cross_stack.py +218 -0
- benchmarks/canonical_profile_instrumentation.py +167 -0
- benchmarks/canonical_public_baseline.py +288 -0
- benchmarks/cli.py +525 -0
- benchmarks/collector_public30_offline.py +1653 -0
- benchmarks/datasets/__init__.py +45 -0
- benchmarks/datasets/conversational_generator.py +238 -0
- benchmarks/datasets/generator.py +283 -0
- benchmarks/datasets/holdout_topology_generator.py +290 -0
- benchmarks/datasets/holdout_topology_generator_v2.py +294 -0
- benchmarks/datasets/multi_world.py +1643 -0
- benchmarks/datasets/scale_generator.py +289 -0
- benchmarks/datasets/schema.py +57 -0
- benchmarks/deep_collision_stress.py +199 -0
- benchmarks/evaluators/__init__.py +12 -0
- benchmarks/evaluators/manifest.py +142 -0
- benchmarks/evaluators/metrics.py +277 -0
- benchmarks/evaluators/runner.py +258 -0
- benchmarks/f1_step0/build_dev_measurement_manifest_v1.py +127 -0
- benchmarks/f1_step0/build_dev_measurement_manifest_v2.py +127 -0
- benchmarks/f1_step0/build_dev_measurement_manifest_v3.py +146 -0
- benchmarks/f1_step0/build_ledger.py +143 -0
- benchmarks/f1_step0/build_ledger_v2.py +296 -0
- benchmarks/f1_step0/compare_dev52_readiness_v1.py +106 -0
- benchmarks/f1_step0/locator_public_smoke.py +172 -0
- benchmarks/f1_step0/prepare_dev52_owner_v1.py +133 -0
- benchmarks/f1_step0/prepare_dev52_owner_v2.py +133 -0
- benchmarks/f1_step0/public_identity_probe.py +191 -0
- benchmarks/f1_step0/public_response_mapper.py +101 -0
- benchmarks/f1_step0/real_source_public_preflight.py +244 -0
- benchmarks/f1_step0/run_exact_fusion_dev_experiment_v1.py +83 -0
- benchmarks/f1_step0/scorer.py +110 -0
- benchmarks/f1_step0/source_aware_dev_measurement_v1.py +627 -0
- benchmarks/f1_step0/source_aware_dev_measurement_v2.py +659 -0
- benchmarks/f1_step0/source_aware_dev_measurement_v3.py +776 -0
- benchmarks/f1_step0/test_public_response_mapper.py +110 -0
- benchmarks/f1_step0/test_real_source_public_preflight.py +25 -0
- benchmarks/f1_step0/test_scorer.py +158 -0
- benchmarks/f1_step0/test_source_aware_dev_measurement.py +189 -0
- benchmarks/f1_step0/test_source_aware_dev_measurement_v2.py +263 -0
- benchmarks/f1_step0/test_source_aware_dev_measurement_v3.py +334 -0
- benchmarks/f1_step0/typed_readback_public_smoke.py +181 -0
- benchmarks/forensic_c4_1_dev.py +258 -0
- benchmarks/forensic_c4_dev.py +255 -0
- benchmarks/forensic_deep25_c4_1.py +478 -0
- benchmarks/legacy_migration_acceptance.py +29 -0
- benchmarks/longmemeval_actor.py +497 -0
- benchmarks/longmemeval_agent_judge.py +593 -0
- benchmarks/longmemeval_budget.py +139 -0
- benchmarks/longmemeval_full.py +410 -0
- benchmarks/longmemeval_judge.py +327 -0
- benchmarks/longmemeval_pilot.py +316 -0
- benchmarks/longmemeval_s.py +318 -0
- benchmarks/memory_bench.py +248 -0
- benchmarks/memory_formation/__init__.py +1 -0
- benchmarks/memory_formation/audit.py +801 -0
- benchmarks/memory_formation/real_audit.py +688 -0
- benchmarks/memory_session_acceptance.py +279 -0
- benchmarks/native_codex_host.py +513 -0
- benchmarks/owner_dense_acceptance.py +126 -0
- benchmarks/pilot_seeker_vault.py +266 -0
- benchmarks/prepare_public30_ripgrep_range_owner.py +115 -0
- benchmarks/profile_canonical.py +373 -0
- benchmarks/profile_canonical_lanes.py +405 -0
- benchmarks/profile_search_chunks.py +351 -0
- benchmarks/public_ranges_faq_proof.py +121 -0
- benchmarks/public_repos/build_history_scope_v3.py +330 -0
- benchmarks/public_repos/build_oracle.py +108 -0
- benchmarks/public_repos/build_oracle_v2.py +169 -0
- benchmarks/public_repos/build_oracle_v3.py +176 -0
- benchmarks/public_repos/map_public_v2.py +296 -0
- benchmarks/public_repos/project_prompts_v2.py +31 -0
- benchmarks/public_repos/project_prompts_v3.py +29 -0
- benchmarks/public_repos/ready_admin_v2.py +214 -0
- benchmarks/public_repos/score_public_v2.py +258 -0
- benchmarks/public_repos/test_map_public_v2.py +171 -0
- benchmarks/public_repos/test_public_repo_v2.py +44 -0
- benchmarks/public_repos/test_public_repo_v3.py +46 -0
- benchmarks/public_repos/test_ready_admin_v2.py +117 -0
- benchmarks/public_repos/test_score_public_v2.py +163 -0
- benchmarks/public_repos/test_wire_v2.py +292 -0
- benchmarks/public_repos/wire_v2.py +603 -0
- benchmarks/real_blocker_fix_acceptance.py +155 -0
- benchmarks/real_scan_rescan_acceptance.py +41 -0
- benchmarks/recall_v2_20k_smoke.py +112 -0
- benchmarks/recall_v2_e1_retrieve.py +70 -0
- benchmarks/recall_v2_e1_score.py +271 -0
- benchmarks/recall_v2_e2_prepare.py +69 -0
- benchmarks/recall_v2_e2_public_score.py +91 -0
- benchmarks/recall_v2_e2_score.py +108 -0
- benchmarks/recall_v2_navigation_mcp_replay.py +308 -0
- benchmarks/recall_v2_reader_capture.py +118 -0
- benchmarks/recall_v2_reader_replay.py +136 -0
- benchmarks/run_blind_h10_holdout.py +284 -0
- benchmarks/run_blind_h9_holdout.py +275 -0
- benchmarks/run_c4_2_nightly_scale.py +446 -0
- benchmarks/run_context_scale_simulation.py +189 -0
- benchmarks/run_conversational_bench.py +170 -0
- benchmarks/run_h11_evaluation.py +316 -0
- benchmarks/run_h11_v2_evaluation.py +431 -0
- benchmarks/run_native_fixed12.py +469 -0
- benchmarks/run_native_preflight.py +280 -0
- benchmarks/run_scale_stress.py +225 -0
- benchmarks/solanazation_brownfield_01/build_oracle.py +970 -0
- benchmarks/solanazation_brownfield_01/check_level_a.py +78 -0
- benchmarks/solanazation_brownfield_01/debug_scope.py +13 -0
- benchmarks/solanazation_brownfield_01/evaluate_bootstrap_02.py +570 -0
- benchmarks/solanazation_brownfield_01/evaluate_first_pass.py +306 -0
- benchmarks/solanazation_brownfield_01/generate_dryrun_report.py +222 -0
- benchmarks/solanazation_brownfield_01/generate_manifests.py +146 -0
- benchmarks/solanazation_brownfield_01/inspect_misses.py +29 -0
- benchmarks/solanazation_brownfield_01/run_miss_forensic.py +206 -0
- benchmarks/solanazation_brownfield_01/test_facts.py +15 -0
- benchmarks/statistics/__init__.py +15 -0
- benchmarks/statistics/significance.py +326 -0
- benchmarks/test_native_mcp_preflight.py +229 -0
- benchmarks/test_prepare_public30_range_owner.py +35 -0
- benchmarks/test_public30_source_reference.py +698 -0
- benchmarks/test_recall_v2_e1_reader_input.py +89 -0
- benchmarks/test_recall_v2_navigation_mcp_replay.py +155 -0
- benchmarks/test_recall_v2_reader_replay.py +41 -0
- dandelion/__init__.py +3 -0
- dandelion/__main__.py +11 -0
- dandelion/agent_adapter.py +2369 -0
- dandelion/bootstrap/__init__.py +872 -0
- dandelion/bootstrap/alias_policy.py +50 -0
- dandelion/bootstrap/authority.py +225 -0
- dandelion/bootstrap/canonical_history.py +412 -0
- dandelion/bootstrap/classifier.py +403 -0
- dandelion/bootstrap/component_graph.py +280 -0
- dandelion/bootstrap/discovery.py +279 -0
- dandelion/bootstrap/domain_induction.py +180 -0
- dandelion/bootstrap/domain_topology.py +673 -0
- dandelion/bootstrap/evidence_envelope.py +132 -0
- dandelion/bootstrap/extractors/base.py +171 -0
- dandelion/bootstrap/extractors/exact.py +1234 -0
- dandelion/bootstrap/extractors/generic.py +100 -0
- dandelion/bootstrap/extractors/git_temporal.py +111 -0
- dandelion/bootstrap/extractors/registry.py +345 -0
- dandelion/bootstrap/extractors/relations.py +89 -0
- dandelion/bootstrap/extractors/semantic.py +424 -0
- dandelion/bootstrap/extractors/status.py +218 -0
- dandelion/bootstrap/git_sweeper.py +135 -0
- dandelion/bootstrap/manifest.py +246 -0
- dandelion/bootstrap/quarantine.py +264 -0
- dandelion/bootstrap/report.py +259 -0
- dandelion/bootstrap/root_resolver.py +221 -0
- dandelion/bootstrap/tech_profiler.py +125 -0
- dandelion/bootstrap/topology.py +75 -0
- dandelion/bootstrap/topology_index.py +110 -0
- dandelion/bootstrap/transaction.py +315 -0
- dandelion/canonical_bootstrap.py +257 -0
- dandelion/canonical_memory.py +4130 -0
- dandelion/captured_input.py +295 -0
- dandelion/cli.py +767 -0
- dandelion/config.py +40 -0
- dandelion/derived_index.py +745 -0
- dandelion/domains.py +706 -0
- dandelion/embedder.py +229 -0
- dandelion/event_evidence.py +571 -0
- dandelion/evidence.py +812 -0
- dandelion/evidence_response.py +453 -0
- dandelion/evidence_workspace.py +2360 -0
- dandelion/exact_context.py +129 -0
- dandelion/hydrator.py +124 -0
- dandelion/inspector.py +442 -0
- dandelion/inspector_assets/dandelion-mark.svg +33 -0
- dandelion/inspector_assets/dandelion_graph.js +361 -0
- dandelion/inspector_assets/inspector.css +223 -0
- dandelion/inspector_assets/inspector.html +135 -0
- dandelion/inspector_assets/inspector.js +1059 -0
- dandelion/legacy_import.py +431 -0
- dandelion/live_source_grants.py +133 -0
- dandelion/local_owner.py +114 -0
- dandelion/memory_client.py +252 -0
- dandelion/memory_contract.py +107 -0
- dandelion/memory_owner.py +560 -0
- dandelion/reader_providers.py +118 -0
- dandelion/recall_v2.py +767 -0
- dandelion/reconciler.py +219 -0
- dandelion/request_budget.py +125 -0
- dandelion/retained_basis.py +223 -0
- dandelion/retrieval.py +485 -0
- dandelion/router.py +405 -0
- dandelion/search_terms.py +53 -0
- dandelion/security.py +124 -0
- dandelion/server.py +810 -0
- dandelion/source_content_fts.py +219 -0
- dandelion/source_delivery.py +89 -0
- dandelion/source_history.py +241 -0
- dandelion/source_identity.py +88 -0
- dandelion/source_navigator.py +5 -0
- dandelion/source_reader.py +444 -0
- dandelion/source_references.py +379 -0
- dandelion/source_sections.py +265 -0
- dandelion/store_kv.py +684 -0
- dandelion/store_vector.py +1289 -0
- dandelion/studio.py +325 -0
- dandelion/wal.py +227 -0
- dandelion_memory-1.0.0rc15.dist-info/METADATA +48 -0
- dandelion_memory-1.0.0rc15.dist-info/RECORD +232 -0
- dandelion_memory-1.0.0rc15.dist-info/WHEEL +5 -0
- dandelion_memory-1.0.0rc15.dist-info/entry_points.txt +5 -0
- dandelion_memory-1.0.0rc15.dist-info/licenses/LICENSE +21 -0
- dandelion_memory-1.0.0rc15.dist-info/top_level.txt +2 -0
|
@@ -0,0 +1,609 @@
|
|
|
1
|
+
"""Epistemic Evidence Plane for Dandelion memory architecture.
|
|
2
|
+
|
|
3
|
+
The Epistemic Plane separates semantic similarity from grounding support:
|
|
4
|
+
- Semantic Plane: dense vectors, asymmetric passage/query embeddings ("What happened and why?")
|
|
5
|
+
- Symbolic Plane: KVStore, exact state, revisions, code identifiers ("What exactly is true?")
|
|
6
|
+
- Evidence Plane: support level, component scope, memory kind, entity/topic grounding ("What does this memory actually support?")
|
|
7
|
+
|
|
8
|
+
Core Thesis:
|
|
9
|
+
"Similarity finds candidates. Evidence decides whether they are knowledge. Dandelion separates relevance from support."
|
|
10
|
+
|
|
11
|
+
Operational constraint:
|
|
12
|
+
- 0 LLM calls on core memory path
|
|
13
|
+
- Deterministic, local, sub-millisecond execution
|
|
14
|
+
"""
|
|
15
|
+
|
|
16
|
+
from __future__ import annotations
|
|
17
|
+
|
|
18
|
+
import enum
|
|
19
|
+
import re
|
|
20
|
+
from dataclasses import dataclass, field, asdict
|
|
21
|
+
from typing import Optional, Any
|
|
22
|
+
|
|
23
|
+
from dandelion.domains import extract_component_root, extract_domain_kind
|
|
24
|
+
|
|
25
|
+
# Technical Code Identifier Pattern
|
|
26
|
+
TECHNICAL_CODE_ID: re.Pattern = re.compile(
|
|
27
|
+
r"\b0x[a-fA-F0-9]{4,}\b|" # Hex addresses / hashes
|
|
28
|
+
r"\b[a-zA-Z0-9_]+/[a-zA-Z0-9_./\-]+\.[a-zA-Z0-9]+\b|" # File paths with extension (src/worker.py)
|
|
29
|
+
r"/(?:[a-zA-Z0-9_.\-]+/)+[a-zA-Z0-9_.\-]*\b|" # Absolute endpoints (/api/v1/...)
|
|
30
|
+
r"\b[a-zA-Z0-9_-]+\.[a-zA-Z0-9_-]+\b|" # Dotted symbols (api.port, config.yaml)
|
|
31
|
+
r"\b[A-Z0-9_]{2,}-[A-Z0-9_]{2,}\b|" # Hyphenated tags (POC-157, ADR-01)
|
|
32
|
+
r"\b[A-Z0-9]+_[A-Z0-9_]+\b|" # ALL_CAPS with underscore (ERR_TIMEOUT, HTTP_500)
|
|
33
|
+
r"\b[a-z]+[A-Z][a-zA-Z0-9]*\b|" # camelCase (getUserProfile)
|
|
34
|
+
r"\b[a-z0-9]+_[a-z0-9_]+\b|" # snake_case (droop_control)
|
|
35
|
+
r"\bADR\b|\bRFC\b" # Architectural decision / request for comment tags
|
|
36
|
+
)
|
|
37
|
+
|
|
38
|
+
QUERY_STOP_WORDS: set[str] = {
|
|
39
|
+
"why", "how", "what", "when", "where", "which", "who", "whom",
|
|
40
|
+
"this", "that", "with", "from", "into", "does", "have", "will",
|
|
41
|
+
"the", "for", "our", "and", "was", "were", "been", "about", "after",
|
|
42
|
+
"show", "find", "list", "check", "extract", "without", "during",
|
|
43
|
+
"not", "all", "can", "get", "set", "use", "has", "had", "are",
|
|
44
|
+
"you", "out", "per", "via", "but", "also", "during", "service",
|
|
45
|
+
"services", "system", "systems", "implementation", "details", "explain",
|
|
46
|
+
"describe", "overview", "note", "notes", "policy", "policies", "review",
|
|
47
|
+
"каков", "какой", "какая", "какие", "для", "после", "при", "как", "что",
|
|
48
|
+
"где", "когда", "почему", "это", "или", "все", "был", "была", "было",
|
|
49
|
+
}
|
|
50
|
+
|
|
51
|
+
COMMON_ACRONYM_EXCLUSIONS: set[str] = {
|
|
52
|
+
"CPU", "RAM", "RCA", "URL", "API", "SQL", "DNS", "SSH", "TCP", "UDP",
|
|
53
|
+
"AWS", "GCP", "SDK", "CLI", "IDE", "HTTP", "REST", "JSON", "YAML", "SIMD", "POST", "GET"
|
|
54
|
+
}
|
|
55
|
+
|
|
56
|
+
KNOWN_SOFTWARE_COMPONENTS: set[str] = {
|
|
57
|
+
"ecommerce", "identity", "streaming", "legaltech", "devtools", "fintech",
|
|
58
|
+
"telemetry", "search", "logistics", "healthcare", "energy", "gaming",
|
|
59
|
+
"iot", "biotech", "robotics", "edtech", "agritech", "aerospace",
|
|
60
|
+
"automotive", "quantum", "insurtech", "marinetech", "cleantech",
|
|
61
|
+
"miningtech", "govtech", "devops", "web3", "secops", "mlops", "mobile",
|
|
62
|
+
"dataeng", "messaging", "crm", "avionics", "genomics", "social",
|
|
63
|
+
"compiler", "hrtech", "traveltech", "audio", "autonomous", "soc",
|
|
64
|
+
"publishing", "maritime", "semiconductor", "pharmatech", "supplychain",
|
|
65
|
+
"space", "neurotech", "realestate",
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
# Cross-lingual equivalence pairs for bilingual support matching
|
|
69
|
+
CROSS_LINGUAL_PAIRS: list[tuple[str, str]] = [
|
|
70
|
+
("streaming", "streaming"),
|
|
71
|
+
("streaming", "асинхрон"),
|
|
72
|
+
("queue", "очеред"),
|
|
73
|
+
("grpc", "grpc"),
|
|
74
|
+
("approved", "утвердил"),
|
|
75
|
+
("decision", "решение"),
|
|
76
|
+
("decision", "переход"),
|
|
77
|
+
("asynchronous", "асинхрон"),
|
|
78
|
+
("communication", "streaming"),
|
|
79
|
+
("communication", "grpc"),
|
|
80
|
+
("processing", "обработк"),
|
|
81
|
+
("incident", "сбой"),
|
|
82
|
+
("incident", "авария"),
|
|
83
|
+
("incident", "падение"),
|
|
84
|
+
("incident", "отказ"),
|
|
85
|
+
]
|
|
86
|
+
|
|
87
|
+
|
|
88
|
+
COMMON_CAMEL_CASE_EXCLUSIONS: set[str] = {
|
|
89
|
+
"ediscovery", "ecommerce", "ebusiness", "egov"
|
|
90
|
+
}
|
|
91
|
+
|
|
92
|
+
# Domain title keyword associations for canonical component mapping
|
|
93
|
+
DOMAIN_TITLE_KEYWORDS: dict[str, set[str]] = {
|
|
94
|
+
"energy": {"smart", "grid", "scada"},
|
|
95
|
+
"legaltech": {"legal", "compliance", "ediscovery"},
|
|
96
|
+
"fintech": {"fintech", "payments", "ledger", "iso8583"},
|
|
97
|
+
"ecommerce": {"ecommerce", "retail", "catalog", "cart"},
|
|
98
|
+
"healthcare": {"healthcare", "hipaa", "ehr", "hl7", "fhir"},
|
|
99
|
+
"web3": {"web3", "defi", "staking", "evm"},
|
|
100
|
+
"secops": {"secops", "cybersecurity", "threat", "bastion"},
|
|
101
|
+
"mlops": {"mlops", "inference", "triton", "vllm"},
|
|
102
|
+
"mobile": {"mobile", "apns", "deep", "sqlite"},
|
|
103
|
+
"iot": {"iot", "embedded", "firmware", "mqtt"},
|
|
104
|
+
"dataeng": {"dataeng", "lakehouse", "clickhouse", "spark"},
|
|
105
|
+
"gaming": {"gaming", "matchmaking", "lobby"},
|
|
106
|
+
"search": {"search", "opensearch", "bm25"},
|
|
107
|
+
"messaging": {"messaging", "webrtc", "synapse"},
|
|
108
|
+
"identity": {"identity", "iam", "keycloak", "saml", "jwks"},
|
|
109
|
+
"telemetry": {"telemetry", "opentelemetry", "observability", "prometheus"},
|
|
110
|
+
"streaming": {"streaming", "transcoding", "hls", "ffmpeg"},
|
|
111
|
+
"crm": {"crm", "salesforce", "integrations"},
|
|
112
|
+
"logistics": {"logistics", "fleet", "telematics"},
|
|
113
|
+
"devtools": {"devtools", "runners", "docker"},
|
|
114
|
+
"avionics": {"avionics", "drone", "autopilot", "mavlink"},
|
|
115
|
+
"genomics": {"genomics", "bioinformatics", "nextflow", "blast"},
|
|
116
|
+
"social": {"social", "graph", "newsfeed", "neo4j"},
|
|
117
|
+
"quantum": {"quantum", "qiskit", "qubit"},
|
|
118
|
+
"devops": {"devops", "tooling", "infrastructure"},
|
|
119
|
+
}
|
|
120
|
+
|
|
121
|
+
|
|
122
|
+
class SupportLevel(str, enum.Enum):
|
|
123
|
+
"""Epistemic support level evaluating candidate memory against query evidence."""
|
|
124
|
+
STRONG = "STRONG" # Grounded structural / exact proof (fact ref, code ID match, component+topic match)
|
|
125
|
+
MEDIUM = "MEDIUM" # Grounded semantic match (component match + topic/action overlap, or high semantic similarity)
|
|
126
|
+
WEAK = "WEAK" # Bare semantic similarity without verified component/topic grounding
|
|
127
|
+
UNSUPPORTED = "UNSUPPORTED" # Contradicted component, ungrounded topic query, or missing mandatory identifier
|
|
128
|
+
|
|
129
|
+
|
|
130
|
+
@dataclass
|
|
131
|
+
class MemoryEvidence:
|
|
132
|
+
"""Evidence envelope stored in memory payload and resolved during recall."""
|
|
133
|
+
domain_root: str = ""
|
|
134
|
+
component_ids: list[str] = field(default_factory=list)
|
|
135
|
+
memory_kind: str = "general"
|
|
136
|
+
entities: list[str] = field(default_factory=list)
|
|
137
|
+
topics: list[str] = field(default_factory=list)
|
|
138
|
+
actions: list[str] = field(default_factory=list)
|
|
139
|
+
fact_refs: list[str] = field(default_factory=list)
|
|
140
|
+
template: str = ""
|
|
141
|
+
event_id: str = ""
|
|
142
|
+
event_sig: Optional[Any] = None
|
|
143
|
+
|
|
144
|
+
def to_dict(self) -> dict[str, Any]:
|
|
145
|
+
d = asdict(self)
|
|
146
|
+
d.pop("event_sig", None)
|
|
147
|
+
return d
|
|
148
|
+
|
|
149
|
+
@classmethod
|
|
150
|
+
def from_dict(cls, data: dict[str, Any]) -> MemoryEvidence:
|
|
151
|
+
return cls(
|
|
152
|
+
domain_root=data.get("domain_root", ""),
|
|
153
|
+
component_ids=data.get("component_ids", []),
|
|
154
|
+
memory_kind=data.get("memory_kind", "general"),
|
|
155
|
+
entities=data.get("entities", []),
|
|
156
|
+
topics=data.get("topics", []),
|
|
157
|
+
actions=data.get("actions", []),
|
|
158
|
+
fact_refs=data.get("fact_refs", []),
|
|
159
|
+
template=data.get("template", ""),
|
|
160
|
+
event_id=data.get("event_id", ""),
|
|
161
|
+
)
|
|
162
|
+
|
|
163
|
+
|
|
164
|
+
@dataclass
|
|
165
|
+
class QueryEvidence:
|
|
166
|
+
"""Structured evidence requirements parsed from a recall query."""
|
|
167
|
+
query: str
|
|
168
|
+
explicit_domains: list[str] = field(default_factory=list)
|
|
169
|
+
explicit_identifiers: list[str] = field(default_factory=list)
|
|
170
|
+
explicit_revs: list[tuple[str, int]] = field(default_factory=list)
|
|
171
|
+
requested_kind: Optional[str] = None
|
|
172
|
+
scoped_prefix: Optional[str] = None
|
|
173
|
+
scope_tokens: list[str] = field(default_factory=list)
|
|
174
|
+
substantive_topics: list[str] = field(default_factory=list)
|
|
175
|
+
is_cyrillic: bool = False
|
|
176
|
+
event_sig: Optional[Any] = None
|
|
177
|
+
|
|
178
|
+
|
|
179
|
+
def extract_query_evidence(
|
|
180
|
+
query: str,
|
|
181
|
+
domain: Optional[str] = None,
|
|
182
|
+
explicit_revs: Optional[list[tuple[str, int]]] = None,
|
|
183
|
+
known_domains: Optional[set[str]] = None,
|
|
184
|
+
) -> QueryEvidence:
|
|
185
|
+
"""Extract structured evidence requirements from a query."""
|
|
186
|
+
q_lower = query.lower()
|
|
187
|
+
raw_tokens = set(re.findall(r"[a-zA-Z0-9\u0400-\u04FF]+", q_lower))
|
|
188
|
+
is_cyrillic = bool(re.search(r"[\u0400-\u04FF]", query))
|
|
189
|
+
|
|
190
|
+
# Incorporate dynamically known domain roots
|
|
191
|
+
components_pool = set(KNOWN_SOFTWARE_COMPONENTS)
|
|
192
|
+
known_roots: set[str] = set()
|
|
193
|
+
if known_domains:
|
|
194
|
+
for d in known_domains:
|
|
195
|
+
root = extract_component_root(d)
|
|
196
|
+
if root:
|
|
197
|
+
known_roots.add(root)
|
|
198
|
+
components_pool.add(root)
|
|
199
|
+
|
|
200
|
+
# Detect scoped component
|
|
201
|
+
scoped_prefix: Optional[str] = None
|
|
202
|
+
if domain:
|
|
203
|
+
scoped_prefix = extract_component_root(domain)
|
|
204
|
+
|
|
205
|
+
# 1. Target scope in query tail (supports English and Russian prepositions: in, for, about, для, в)
|
|
206
|
+
tail_tokens: set[str] = set()
|
|
207
|
+
norm_tail: str = ""
|
|
208
|
+
tail_match = re.search(
|
|
209
|
+
r"\b(?:in|for|about|для|в)\s+((?:(?!\b(?:in|for|about|для|в)\b)[A-Za-z0-9_& /,\-()\u0400-\u04FF])+?)(?:\.|\?|\s*)$",
|
|
210
|
+
query,
|
|
211
|
+
re.IGNORECASE,
|
|
212
|
+
)
|
|
213
|
+
if tail_match:
|
|
214
|
+
tail_str = tail_match.group(1).strip()
|
|
215
|
+
tail_tokens = set(re.findall(r"[a-zA-Z0-9\u0400-\u04FF]+", tail_str.lower()))
|
|
216
|
+
norm_tail = re.sub(r"[\s\-]+", "", tail_str.lower())
|
|
217
|
+
|
|
218
|
+
if not scoped_prefix and tail_tokens:
|
|
219
|
+
# Step 1: Multi-keyword title matches (overlap >= 2, e.g. "Smart Grid & SCADA Telemetry" -> energy)
|
|
220
|
+
best_dom = None
|
|
221
|
+
max_kw_overlap = 1
|
|
222
|
+
for dom, kw_set in DOMAIN_TITLE_KEYWORDS.items():
|
|
223
|
+
if known_roots and dom not in known_roots:
|
|
224
|
+
continue
|
|
225
|
+
overlap = len(kw_set & tail_tokens)
|
|
226
|
+
if overlap > max_kw_overlap and overlap >= 2:
|
|
227
|
+
max_kw_overlap = overlap
|
|
228
|
+
best_dom = dom
|
|
229
|
+
if best_dom:
|
|
230
|
+
scoped_prefix = best_dom
|
|
231
|
+
|
|
232
|
+
# Step 2: Exact known root match in tail tokens or normalized tail
|
|
233
|
+
if not scoped_prefix and known_roots:
|
|
234
|
+
for kr in sorted(known_roots, key=len, reverse=True):
|
|
235
|
+
if kr in tail_tokens or kr == norm_tail:
|
|
236
|
+
scoped_prefix = kr
|
|
237
|
+
break
|
|
238
|
+
|
|
239
|
+
# Step 3: Partial / substring root match with substantive tail tokens
|
|
240
|
+
if not scoped_prefix and known_roots:
|
|
241
|
+
for kr in sorted(known_roots, key=len, reverse=True):
|
|
242
|
+
if any(len(tok) >= 4 and (tok in kr or kr.startswith(tok)) for tok in tail_tokens):
|
|
243
|
+
scoped_prefix = kr
|
|
244
|
+
break
|
|
245
|
+
|
|
246
|
+
# Step 4: Single-keyword domain title match or generic components pool
|
|
247
|
+
if not scoped_prefix:
|
|
248
|
+
best_dom = None
|
|
249
|
+
max_kw_overlap = 0
|
|
250
|
+
for dom, kw_set in DOMAIN_TITLE_KEYWORDS.items():
|
|
251
|
+
if known_roots and dom not in known_roots:
|
|
252
|
+
continue
|
|
253
|
+
overlap = len(kw_set & tail_tokens)
|
|
254
|
+
if overlap > max_kw_overlap and overlap >= 1:
|
|
255
|
+
max_kw_overlap = overlap
|
|
256
|
+
best_dom = dom
|
|
257
|
+
if best_dom:
|
|
258
|
+
scoped_prefix = best_dom
|
|
259
|
+
else:
|
|
260
|
+
for comp in sorted(components_pool, key=len, reverse=True):
|
|
261
|
+
if known_roots and comp not in known_roots:
|
|
262
|
+
continue
|
|
263
|
+
if comp in tail_tokens or comp in norm_tail:
|
|
264
|
+
scoped_prefix = comp
|
|
265
|
+
break
|
|
266
|
+
|
|
267
|
+
# 2. Uppercase prefix match in query (e.g. ECOMMERCE-401, FINTECH, DEVTOOLS)
|
|
268
|
+
if not scoped_prefix:
|
|
269
|
+
upper_matches = re.findall(r"\b[A-Z]{3,}(?:-[A-Z0-9]+)?\b", query)
|
|
270
|
+
for m in upper_matches:
|
|
271
|
+
tag = m.split("-")[0]
|
|
272
|
+
if tag not in COMMON_ACRONYM_EXCLUSIONS:
|
|
273
|
+
t_low = tag.lower()
|
|
274
|
+
if (not known_roots or t_low in known_roots) and (t_low in components_pool or t_low in known_roots):
|
|
275
|
+
scoped_prefix = t_low
|
|
276
|
+
break
|
|
277
|
+
|
|
278
|
+
# 3. Known component token in raw query (excluding ambiguous technical words like streaming)
|
|
279
|
+
if not scoped_prefix:
|
|
280
|
+
ambiguous = {"streaming", "search", "mobile", "gaming", "social", "energy"}
|
|
281
|
+
search_pool = known_roots if known_roots else components_pool
|
|
282
|
+
for comp in sorted(search_pool, key=len, reverse=True):
|
|
283
|
+
if comp in raw_tokens and comp not in ambiguous:
|
|
284
|
+
scoped_prefix = comp
|
|
285
|
+
break
|
|
286
|
+
|
|
287
|
+
# Scope tokens associated with component
|
|
288
|
+
scope_tokens: set[str] = set()
|
|
289
|
+
if scoped_prefix:
|
|
290
|
+
scope_tokens.add(scoped_prefix)
|
|
291
|
+
if scoped_prefix in DOMAIN_TITLE_KEYWORDS:
|
|
292
|
+
scope_tokens.update(DOMAIN_TITLE_KEYWORDS[scoped_prefix])
|
|
293
|
+
if tail_tokens:
|
|
294
|
+
scope_tokens.update(tail_tokens)
|
|
295
|
+
|
|
296
|
+
# Code identifiers (excluding title words, acronyms, and non-code camelCase like eDiscovery)
|
|
297
|
+
code_ids = []
|
|
298
|
+
for ident in TECHNICAL_CODE_ID.findall(query):
|
|
299
|
+
cleaned = ident.strip(".:,-/")
|
|
300
|
+
if (
|
|
301
|
+
cleaned.lower() not in QUERY_STOP_WORDS
|
|
302
|
+
and cleaned not in COMMON_ACRONYM_EXCLUSIONS
|
|
303
|
+
and cleaned.lower() not in scope_tokens
|
|
304
|
+
and cleaned.lower() not in COMMON_CAMEL_CASE_EXCLUSIONS
|
|
305
|
+
):
|
|
306
|
+
code_ids.append(cleaned)
|
|
307
|
+
|
|
308
|
+
# Memory kind intent
|
|
309
|
+
requested_kind: Optional[str] = None
|
|
310
|
+
if any(k in q_lower for k in ("incident", "outage", "failure", "crash", "postmortem", "post-mortem", "downtime", "alert", "500", "сбой", "инцидент", "авария")):
|
|
311
|
+
requested_kind = "incident"
|
|
312
|
+
elif any(k in q_lower for k in ("architecture", "architectural", "adr", "rfc", "design", "архитектур")):
|
|
313
|
+
requested_kind = "architecture"
|
|
314
|
+
elif any(k in q_lower for k in ("config", "settings", "port", "host", "endpoint", "url", "key", "конфиг")):
|
|
315
|
+
requested_kind = "config"
|
|
316
|
+
elif any(k in q_lower for k in ("bug", "bugfix", "fix", "patch", "issue", "ошибк", "исправлен")):
|
|
317
|
+
requested_kind = "bugfix"
|
|
318
|
+
|
|
319
|
+
# Substantive question topics (tokens not in stop words, not in scope tokens, not kind keywords)
|
|
320
|
+
kind_noise = {
|
|
321
|
+
"incident", "outage", "failure", "crash", "postmortem", "downtime", "alert",
|
|
322
|
+
"architecture", "architectural", "design",
|
|
323
|
+
"config", "configuration", "settings", "options", "setup",
|
|
324
|
+
"bugfix", "reference", "general"
|
|
325
|
+
}
|
|
326
|
+
substantive_topics = set()
|
|
327
|
+
for tok in raw_tokens:
|
|
328
|
+
if (
|
|
329
|
+
len(tok) > 2
|
|
330
|
+
and tok not in QUERY_STOP_WORDS
|
|
331
|
+
and tok not in scope_tokens
|
|
332
|
+
and tok not in kind_noise
|
|
333
|
+
and not tok.isdigit()
|
|
334
|
+
):
|
|
335
|
+
substantive_topics.add(tok)
|
|
336
|
+
|
|
337
|
+
explicit_domains = [domain] if domain else []
|
|
338
|
+
from benchmarks.adapters.native.event_evidence_b2 import PredicateNormalizer
|
|
339
|
+
ev_sig = PredicateNormalizer.extract_signature(
|
|
340
|
+
text=query,
|
|
341
|
+
domain_root=scoped_prefix or (explicit_domains[0].split(".")[0] if explicit_domains else ""),
|
|
342
|
+
)
|
|
343
|
+
|
|
344
|
+
return QueryEvidence(
|
|
345
|
+
query=query,
|
|
346
|
+
explicit_domains=explicit_domains,
|
|
347
|
+
explicit_identifiers=sorted(code_ids),
|
|
348
|
+
explicit_revs=explicit_revs or [],
|
|
349
|
+
requested_kind=requested_kind,
|
|
350
|
+
scoped_prefix=scoped_prefix,
|
|
351
|
+
scope_tokens=sorted(scope_tokens),
|
|
352
|
+
substantive_topics=sorted(substantive_topics),
|
|
353
|
+
is_cyrillic=is_cyrillic,
|
|
354
|
+
event_sig=ev_sig,
|
|
355
|
+
)
|
|
356
|
+
|
|
357
|
+
|
|
358
|
+
def extract_memory_evidence(
|
|
359
|
+
domain: str,
|
|
360
|
+
template: str,
|
|
361
|
+
referenced_facts: Optional[list[str]] = None,
|
|
362
|
+
event_id: str = "",
|
|
363
|
+
) -> MemoryEvidence:
|
|
364
|
+
"""Extract evidence envelope from memory domain and template."""
|
|
365
|
+
domain_root = extract_component_root(domain)
|
|
366
|
+
domain_kind = extract_domain_kind(domain)
|
|
367
|
+
|
|
368
|
+
# Component IDs from domain and template
|
|
369
|
+
comp_ids: set[str] = set()
|
|
370
|
+
if domain_root:
|
|
371
|
+
comp_ids.add(domain_root)
|
|
372
|
+
|
|
373
|
+
# Extract [PREFIX-123] or service name from template
|
|
374
|
+
prefix_match = re.findall(r"\b([A-Z]{3,})-\d+\b", template)
|
|
375
|
+
for p in prefix_match:
|
|
376
|
+
if p not in COMMON_ACRONYM_EXCLUSIONS:
|
|
377
|
+
comp_ids.add(p.lower())
|
|
378
|
+
|
|
379
|
+
t_lower = template.lower()
|
|
380
|
+
domain_segments = set(domain.lower().split("."))
|
|
381
|
+
for comp in KNOWN_SOFTWARE_COMPONENTS:
|
|
382
|
+
if comp == domain_root or comp in domain_segments:
|
|
383
|
+
comp_ids.add(comp)
|
|
384
|
+
|
|
385
|
+
# Memory kind
|
|
386
|
+
mem_kind = domain_kind
|
|
387
|
+
if any(k in t_lower for k in ("incident", "outage", "failure", "crash", "postmortem", "alert", "500")):
|
|
388
|
+
if mem_kind == "general":
|
|
389
|
+
mem_kind = "incident"
|
|
390
|
+
elif any(k in t_lower for k in ("adr", "rfc", "architecture", "decision")):
|
|
391
|
+
if mem_kind == "general":
|
|
392
|
+
mem_kind = "architecture"
|
|
393
|
+
|
|
394
|
+
# Entities
|
|
395
|
+
entities = [
|
|
396
|
+
ident.strip(".:,-/")
|
|
397
|
+
for ident in TECHNICAL_CODE_ID.findall(template)
|
|
398
|
+
if ident.lower() not in QUERY_STOP_WORDS and ident not in COMMON_ACRONYM_EXCLUSIONS
|
|
399
|
+
]
|
|
400
|
+
|
|
401
|
+
# Fact refs
|
|
402
|
+
fact_refs = list(referenced_facts) if referenced_facts else []
|
|
403
|
+
in_tmpl_refs = re.findall(r"\{\{ref:([\w.]+)(?:@rev:\d+)?\}\}", template)
|
|
404
|
+
for r in in_tmpl_refs:
|
|
405
|
+
if r not in fact_refs:
|
|
406
|
+
fact_refs.append(r)
|
|
407
|
+
|
|
408
|
+
# Actions and Topics
|
|
409
|
+
raw_tokens = [w.lower().strip(".:,-/?!\"'(){}[]") for w in re.split(r"[\s\-._/]+", template)]
|
|
410
|
+
topics = [
|
|
411
|
+
tok for tok in raw_tokens
|
|
412
|
+
if (
|
|
413
|
+
len(tok) > 2
|
|
414
|
+
and tok not in QUERY_STOP_WORDS
|
|
415
|
+
and tok not in comp_ids
|
|
416
|
+
and not tok.isdigit()
|
|
417
|
+
)
|
|
418
|
+
]
|
|
419
|
+
|
|
420
|
+
from benchmarks.adapters.native.event_evidence_b2 import PredicateNormalizer
|
|
421
|
+
m_sig = PredicateNormalizer.extract_signature(
|
|
422
|
+
text=template,
|
|
423
|
+
domain_root=domain_root,
|
|
424
|
+
event_id=event_id,
|
|
425
|
+
)
|
|
426
|
+
|
|
427
|
+
return MemoryEvidence(
|
|
428
|
+
domain_root=domain_root,
|
|
429
|
+
component_ids=sorted(comp_ids),
|
|
430
|
+
memory_kind=mem_kind,
|
|
431
|
+
entities=sorted(set(entities)),
|
|
432
|
+
topics=sorted(set(topics)),
|
|
433
|
+
actions=[],
|
|
434
|
+
fact_refs=sorted(set(fact_refs)),
|
|
435
|
+
template=template,
|
|
436
|
+
event_id=event_id,
|
|
437
|
+
event_sig=m_sig,
|
|
438
|
+
)
|
|
439
|
+
|
|
440
|
+
|
|
441
|
+
def resolve_evidence(
|
|
442
|
+
query_ev: QueryEvidence,
|
|
443
|
+
mem_ev: MemoryEvidence,
|
|
444
|
+
vector_score: float,
|
|
445
|
+
is_fact_boosted: bool = False,
|
|
446
|
+
enable_event_gate: bool = True,
|
|
447
|
+
) -> tuple[SupportLevel, float, str]:
|
|
448
|
+
"""Resolve whether candidate memory provides epistemic support for the query.
|
|
449
|
+
|
|
450
|
+
Returns:
|
|
451
|
+
(SupportLevel, evidence_score, reason)
|
|
452
|
+
"""
|
|
453
|
+
from benchmarks.adapters.native.event_evidence_b2 import (
|
|
454
|
+
PredicateNormalizer,
|
|
455
|
+
EventRelation,
|
|
456
|
+
resolve_event_relation,
|
|
457
|
+
)
|
|
458
|
+
|
|
459
|
+
q_sig = getattr(query_ev, "event_sig", None)
|
|
460
|
+
if q_sig is None:
|
|
461
|
+
q_sig = PredicateNormalizer.extract_signature(
|
|
462
|
+
text=query_ev.query,
|
|
463
|
+
domain_root=query_ev.scoped_prefix or (query_ev.explicit_domains[0].split(".")[0] if query_ev.explicit_domains else ""),
|
|
464
|
+
)
|
|
465
|
+
|
|
466
|
+
m_sig = getattr(mem_ev, "event_sig", None)
|
|
467
|
+
if m_sig is None:
|
|
468
|
+
m_sig = PredicateNormalizer.extract_signature(
|
|
469
|
+
text=mem_ev.template or "",
|
|
470
|
+
domain_root=mem_ev.domain_root,
|
|
471
|
+
event_id=mem_ev.event_id or "",
|
|
472
|
+
)
|
|
473
|
+
|
|
474
|
+
has_event_semantics = (
|
|
475
|
+
q_sig.event_kind != "general"
|
|
476
|
+
or m_sig.event_kind != "general"
|
|
477
|
+
or bool(q_sig.trigger)
|
|
478
|
+
or bool(q_sig.failure_mode)
|
|
479
|
+
)
|
|
480
|
+
|
|
481
|
+
# 1. Symbolic Plane: Fact reverse lookup match
|
|
482
|
+
if is_fact_boosted:
|
|
483
|
+
if enable_event_gate and has_event_semantics:
|
|
484
|
+
rel = resolve_event_relation(q_sig, m_sig)
|
|
485
|
+
if rel == EventRelation.CONFLICTING_EVENT:
|
|
486
|
+
return SupportLevel.UNSUPPORTED, 0.0, f"Event evidence rejection: {rel.value} (ungrounded)"
|
|
487
|
+
return SupportLevel.STRONG, 1.0, "Symbolic fact match"
|
|
488
|
+
|
|
489
|
+
# 2. Symbolic Plane: Explicit code identifier gate
|
|
490
|
+
if query_ev.explicit_identifiers:
|
|
491
|
+
mem_ident_set = set(ident.lower() for ident in mem_ev.entities)
|
|
492
|
+
mem_topics_set = set(t.lower() for t in mem_ev.topics)
|
|
493
|
+
all_mem_tokens = mem_ident_set | mem_topics_set | set(mem_ev.component_ids)
|
|
494
|
+
|
|
495
|
+
has_ident_match = any(
|
|
496
|
+
qid.lower() in all_mem_tokens or any(qid.lower() in tok for tok in all_mem_tokens)
|
|
497
|
+
for qid in query_ev.explicit_identifiers
|
|
498
|
+
)
|
|
499
|
+
if has_ident_match:
|
|
500
|
+
if enable_event_gate and has_event_semantics:
|
|
501
|
+
rel = resolve_event_relation(q_sig, m_sig)
|
|
502
|
+
if rel == EventRelation.CONFLICTING_EVENT:
|
|
503
|
+
return SupportLevel.UNSUPPORTED, 0.0, f"Event evidence rejection: {rel.value} (ungrounded)"
|
|
504
|
+
return SupportLevel.STRONG, 0.95, "Code identifier match"
|
|
505
|
+
else:
|
|
506
|
+
return SupportLevel.UNSUPPORTED, 0.0, "Missing required code identifier (ungrounded)"
|
|
507
|
+
|
|
508
|
+
# 3. Event Plane: ARCH-EVENT-01 Structured Event Resolution
|
|
509
|
+
if enable_event_gate and has_event_semantics:
|
|
510
|
+
rel = resolve_event_relation(q_sig, m_sig)
|
|
511
|
+
if rel == EventRelation.CONFLICTING_EVENT:
|
|
512
|
+
return SupportLevel.UNSUPPORTED, 0.0, f"Event evidence rejection: {rel.value} (ungrounded)"
|
|
513
|
+
elif rel == EventRelation.SAME_EVENT:
|
|
514
|
+
return SupportLevel.STRONG, 0.95, "Verified event identity match"
|
|
515
|
+
elif rel == EventRelation.RELATED_EVENT:
|
|
516
|
+
if (q_sig.failure_mode and q_sig.failure_mode == m_sig.failure_mode and vector_score >= 0.60):
|
|
517
|
+
pass
|
|
518
|
+
else:
|
|
519
|
+
return SupportLevel.UNSUPPORTED, 0.0, f"Event evidence rejection: {rel.value} (ungrounded)"
|
|
520
|
+
elif rel == EventRelation.INSUFFICIENT_EVIDENCE:
|
|
521
|
+
if query_ev.requested_kind == "incident" or q_sig.event_kind == "incident":
|
|
522
|
+
mem_tokens = set(mem_ev.topics) | set(mem_ev.component_ids) | set(mem_ev.entities)
|
|
523
|
+
has_text_overlap = any(qt in mem_tokens for qt in query_ev.substantive_topics)
|
|
524
|
+
if not has_text_overlap:
|
|
525
|
+
return SupportLevel.UNSUPPORTED, 0.0, f"Event evidence rejection: {rel.value} (ungrounded)"
|
|
526
|
+
|
|
527
|
+
|
|
528
|
+
mem_tokens = set(mem_ev.topics) | set(mem_ev.component_ids) | set(mem_ev.entities)
|
|
529
|
+
|
|
530
|
+
# 3. Evidence Plane: Component-Scope Anchoring (SELECT-A)
|
|
531
|
+
scoped_prefix = query_ev.scoped_prefix
|
|
532
|
+
if scoped_prefix:
|
|
533
|
+
mem_comp_set = set(mem_ev.component_ids)
|
|
534
|
+
if mem_ev.domain_root:
|
|
535
|
+
mem_comp_set.add(mem_ev.domain_root)
|
|
536
|
+
|
|
537
|
+
if scoped_prefix not in mem_comp_set:
|
|
538
|
+
# Component mismatch: query explicitly scoped to a different component!
|
|
539
|
+
return SupportLevel.UNSUPPORTED, 0.0, f"Component mismatch ({mem_ev.domain_root} != {scoped_prefix})"
|
|
540
|
+
|
|
541
|
+
# Kind check: If query requested an incident, memory MUST be an incident
|
|
542
|
+
if query_ev.requested_kind == "incident":
|
|
543
|
+
if mem_ev.memory_kind != "incident":
|
|
544
|
+
return SupportLevel.UNSUPPORTED, 0.0, "Requested incident but memory is not an incident"
|
|
545
|
+
|
|
546
|
+
# Check substantive question topics
|
|
547
|
+
q_topics = query_ev.substantive_topics
|
|
548
|
+
topic_hits = 0
|
|
549
|
+
for qt in q_topics:
|
|
550
|
+
if qt in mem_tokens:
|
|
551
|
+
topic_hits += 1
|
|
552
|
+
else:
|
|
553
|
+
for mt in mem_tokens:
|
|
554
|
+
if len(qt) >= 4 and len(mt) >= 4 and (qt[:4] == mt[:4] or qt in mt or mt in qt):
|
|
555
|
+
topic_hits += 1
|
|
556
|
+
break
|
|
557
|
+
# Cross-lingual equivalence check
|
|
558
|
+
matched_cl = False
|
|
559
|
+
for eng, rus in CROSS_LINGUAL_PAIRS:
|
|
560
|
+
if (qt == eng and rus in mt) or (rus in qt and mt == eng):
|
|
561
|
+
topic_hits += 1
|
|
562
|
+
matched_cl = True
|
|
563
|
+
break
|
|
564
|
+
if matched_cl:
|
|
565
|
+
break
|
|
566
|
+
|
|
567
|
+
if q_topics and topic_hits == 0:
|
|
568
|
+
# Component matched, but substantive question topics are completely ungrounded!
|
|
569
|
+
return SupportLevel.UNSUPPORTED, 0.0, "Substantive question topics ungrounded in component memory"
|
|
570
|
+
|
|
571
|
+
# SELECT-B: Grounded Component + Topic Match
|
|
572
|
+
if vector_score >= 0.55:
|
|
573
|
+
return SupportLevel.STRONG, 0.90 + min(0.08, topic_hits * 0.02), f"Grounded component and topic match ({topic_hits} hits)"
|
|
574
|
+
elif vector_score >= 0.45:
|
|
575
|
+
return SupportLevel.MEDIUM, 0.70 + min(0.08, topic_hits * 0.02), f"Plausible component and topic match ({topic_hits} hits)"
|
|
576
|
+
elif query_ev.explicit_domains and (topic_hits >= 1 or not q_topics):
|
|
577
|
+
return SupportLevel.MEDIUM, 0.60 + min(0.08, topic_hits * 0.02), f"Domain-scoped topic match ({topic_hits} hits)"
|
|
578
|
+
else:
|
|
579
|
+
return SupportLevel.WEAK, 0.40, "Low vector similarity"
|
|
580
|
+
|
|
581
|
+
# 4. Pure semantic query without explicit component scope
|
|
582
|
+
if query_ev.requested_kind == "incident" and mem_ev.memory_kind != "incident":
|
|
583
|
+
return SupportLevel.UNSUPPORTED, 0.0, "Query requested incident but memory is not an incident"
|
|
584
|
+
|
|
585
|
+
# Cross-lingual / Cyrillic
|
|
586
|
+
if query_ev.is_cyrillic:
|
|
587
|
+
cyr_q = {w.lower() for w in query_ev.substantive_topics if re.search(r"[\u0400-\u04FF]", w)}
|
|
588
|
+
cyr_t = {w.lower() for w in mem_tokens if re.search(r"[\u0400-\u04FF]", w)}
|
|
589
|
+
stem_q = {w[:4] for w in cyr_q if len(w) >= 4}
|
|
590
|
+
stem_t = {w[:4] for w in cyr_t if len(w) >= 4}
|
|
591
|
+
stem_hits = len(stem_q & stem_t)
|
|
592
|
+
cyr_overlap = len(cyr_q & cyr_t)
|
|
593
|
+
if vector_score >= 0.60 and (cyr_overlap > 0 or stem_hits > 0):
|
|
594
|
+
return SupportLevel.MEDIUM, 0.75, "Cyrillic semantic match with topic grounding"
|
|
595
|
+
else:
|
|
596
|
+
return SupportLevel.UNSUPPORTED, 0.0, "Cyrillic match lacks lexical grounding"
|
|
597
|
+
|
|
598
|
+
topic_hits = sum(1 for qt in query_ev.substantive_topics if any(qt in mt or mt in qt for mt in mem_tokens))
|
|
599
|
+
if query_ev.substantive_topics and topic_hits == 0:
|
|
600
|
+
return SupportLevel.UNSUPPORTED, 0.0, "Substantive question topics ungrounded in memory"
|
|
601
|
+
|
|
602
|
+
if vector_score >= 0.65:
|
|
603
|
+
return SupportLevel.MEDIUM, 0.80, "High cosine match"
|
|
604
|
+
elif vector_score >= 0.60:
|
|
605
|
+
return SupportLevel.MEDIUM, 0.70, "Medium cosine match"
|
|
606
|
+
elif query_ev.explicit_domains and (topic_hits >= 1 or not query_ev.substantive_topics):
|
|
607
|
+
return SupportLevel.MEDIUM, 0.65, f"Domain-scoped semantic match ({topic_hits} hits)"
|
|
608
|
+
else:
|
|
609
|
+
return SupportLevel.WEAK, 0.40, "Low similarity"
|