feat: Intelligence Pipeline v3 — full implementation

Multi-stage evidence-grounded inference architecture replacing the monolithic 9B model extraction pipeline. CPU-first specialist services handle routine extraction while the 9B vLLM model is preserved for semantic adjudication of ambiguous cases. Key components: - Capability-aware inference gateway (OpenAI-compatible + Ollama) - Endpoint registry with DB migrations and REST API - Sentence-aware document segmenter (property tests) - Deterministic financial parsing with offset integrity - Symbol resolution with ambiguity detection - Specialist service (GLiNER2, dynamic batching, K8s deployment) - Company-specific sentiment (FinBERT, calibration) - Retrieval-based novelty and duplicate detection - Confidence calibration pipeline - Deterministic routing engine (property tests) - 9B adjudication layer with VRAM gating - Stock-specific impact model (features, labels, baseline, trained) - Pipeline orchestrator (state machine, queues, leases, feature flags) - Bounded parallelism (async workers, semaphore, load shedding) - Observability (tracing, metrics, alerts) - Compatibility adapter (v3→v2 golden mapping tests) - Shadow/canary promotion framework - Active learning and fine-tuning pipeline Test results: 1,161 tests pass, ruff lint clean. All 282 spec tasks completed.
2026-07-13 02:14:59 +00:00
parent 84634a365e
commit a72f336ad1
227 changed files with 50403 additions and 0 deletions
@@ -0,0 +1,100 @@
+"""Data models for evidence verification results and rejected candidates.
+
+Provides structured types for offset verification, entity-evidence association,
+numeric consistency checks, rejected-candidate storage, and verification reports.
+"""
+
+from __future__ import annotations
+
+from datetime import datetime, timezone
+from enum import Enum
+
+from pydantic import BaseModel, Field
+
+
+class RejectionReason(str, Enum):
+    """Structured reason codes for candidate rejection during verification."""
+
+    INVALID_OFFSET = "invalid_offset"
+    TEXT_MISMATCH = "text_mismatch"
+    ENTITY_NOT_IN_EVIDENCE = "entity_not_in_evidence"
+    NUMERIC_INCONSISTENCY = "numeric_inconsistency"
+    UNSUPPORTED_CLAIM = "unsupported_claim"
+    SCHEMA_VIOLATION = "schema_violation"
+    CONFIDENCE_BELOW_THRESHOLD = "confidence_below_threshold"
+
+
+class OffsetVerification(BaseModel):
+    """Result of verifying a single evidence span's offsets against source text."""
+
+    span_id: str = Field(description="ID of the evidence span being verified.")
+    valid: bool = Field(description="Whether the span text matches source at offsets.")
+    reason: str | None = Field(
+        default=None,
+        description="Explanation when invalid (e.g., 'text mismatch at offset 42').",
+    )
+
+
+class AssociationVerification(BaseModel):
+    """Result of verifying that an entity appears in at least one linked evidence span."""
+
+    entity_id: str = Field(description="ID of the entity being verified.")
+    valid: bool = Field(description="Whether entity text was found in any linked span.")
+    reason: str | None = Field(
+        default=None,
+        description="Explanation when invalid (e.g., 'entity text not in any linked span').",
+    )
+
+
+class NumericVerification(BaseModel):
+    """Result of verifying a numeric fact against its evidence spans."""
+
+    fact_id: str = Field(description="ID of the numeric fact being verified.")
+    valid: bool = Field(description="Whether the numeric value was found in evidence text.")
+    found_value: str | None = Field(
+        default=None,
+        description="The value found in the evidence text, if any.",
+    )
+    expected_value: str = Field(description="The expected normalized value from the fact.")
+    reason: str | None = Field(
+        default=None,
+        description="Explanation when invalid (e.g., 'value 3.14 not found in evidence').",
+    )
+
+
+class RejectedCandidate(BaseModel):
+    """A candidate that was rejected during verification, stored for audit and learning."""
+
+    candidate_type: str = Field(
+        description="Type of candidate: entity, fact, event, relation, sentiment."
+    )
+    candidate_data: dict = Field(
+        description="Serialized candidate data for audit trail."
+    )
+    rejection_reason: RejectionReason = Field(
+        description="Structured reason code for rejection."
+    )
+    stage: str = Field(
+        description="Pipeline stage where rejection occurred (e.g., 'offset_verification')."
+    )
+    timestamp: datetime = Field(
+        default_factory=lambda: datetime.now(tz=timezone.utc),
+        description="When the rejection was recorded.",
+    )
+
+
+class VerificationReport(BaseModel):
+    """Summary report from a full verification pass over extraction candidates."""
+
+    total_candidates: int = Field(ge=0, description="Total candidates evaluated.")
+    verified: int = Field(ge=0, description="Candidates that passed verification.")
+    rejected: int = Field(ge=0, description="Candidates that failed verification.")
+    coverage_rate: float = Field(
+        ge=0.0,
+        le=1.0,
+        description="Proportion of candidates with valid evidence support.",
+    )
+    rejection_breakdown: dict[str, int] = Field(
+        default_factory=dict,
+        description="Count of rejections per RejectionReason code.",
+    )