Files
stonks-oracle/infra/migrations/041_v3_pipeline_tables.sql
Celes Renata a72f336ad1 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

351 lines
18 KiB
PL/PgSQL

-- Migration 041: V3 Pipeline Persistence Tables
-- Creates tables for the Intelligence Pipeline v3 staged evidence architecture:
-- v3_pipeline_runs, v3_stage_runs, v3_document_chunks, v3_evidence_spans,
-- v3_extracted_entities, v3_extracted_facts, v3_extracted_relations,
-- v3_rejected_candidates, v3_company_signal_candidates,
-- v3_adjudication_decisions, v3_routing_decisions, v3_stage_lineage
-- Includes idempotency keys, immutable-revision constraints, and appropriate indexes.
-- ═══════════════════════════════════════════════════════════════════════════════
-- 20.1: Pipeline runs and stage runs
-- ═══════════════════════════════════════════════════════════════════════════════
-- ─── v3_pipeline_runs ─────────────────────────────────────────────────────────
-- Top-level pipeline execution record for a document.
CREATE TABLE IF NOT EXISTS v3_pipeline_runs (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
document_id UUID NOT NULL,
pipeline_version TEXT NOT NULL DEFAULT 'v3.0',
status TEXT NOT NULL DEFAULT 'pending' CHECK (status IN ('pending', 'running', 'completed', 'failed')),
started_at TIMESTAMPTZ,
completed_at TIMESTAMPTZ,
idempotency_key TEXT NOT NULL UNIQUE,
error TEXT,
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_pipeline_runs_document
ON v3_pipeline_runs(document_id);
CREATE INDEX IF NOT EXISTS idx_v3_pipeline_runs_status
ON v3_pipeline_runs(status);
CREATE INDEX IF NOT EXISTS idx_v3_pipeline_runs_created
ON v3_pipeline_runs(created_at DESC);
-- ─── v3_stage_runs ────────────────────────────────────────────────────────────
-- Individual stage execution within a pipeline run.
CREATE TABLE IF NOT EXISTS v3_stage_runs (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
pipeline_run_id UUID NOT NULL REFERENCES v3_pipeline_runs(id) ON DELETE CASCADE,
stage TEXT NOT NULL CHECK (stage IN (
'segmentation', 'extraction', 'sentiment', 'novelty',
'routing', 'adjudication', 'impact', 'persistence'
)),
status TEXT NOT NULL DEFAULT 'pending' CHECK (status IN ('pending', 'running', 'completed', 'failed', 'skipped')),
started_at TIMESTAMPTZ,
completed_at TIMESTAMPTZ,
endpoint_id UUID REFERENCES inference_endpoints(id) ON DELETE SET NULL,
deployment_id UUID REFERENCES model_deployments(id) ON DELETE SET NULL,
model_version TEXT,
schema_version TEXT,
input_refs JSONB NOT NULL DEFAULT '[]',
output_refs JSONB NOT NULL DEFAULT '[]',
trace_id TEXT,
error TEXT,
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_stage_runs_pipeline
ON v3_stage_runs(pipeline_run_id);
CREATE INDEX IF NOT EXISTS idx_v3_stage_runs_stage
ON v3_stage_runs(stage);
CREATE INDEX IF NOT EXISTS idx_v3_stage_runs_status
ON v3_stage_runs(status);
-- ═══════════════════════════════════════════════════════════════════════════════
-- 20.2: Document chunks and evidence spans
-- ═══════════════════════════════════════════════════════════════════════════════
-- ─── v3_document_chunks ───────────────────────────────────────────────────────
-- Segmented document chunks with offset tracking.
CREATE TABLE IF NOT EXISTS v3_document_chunks (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
document_id UUID NOT NULL,
chunk_id TEXT NOT NULL,
section_path JSONB NOT NULL DEFAULT '[]',
speaker TEXT,
start_char INTEGER NOT NULL,
end_char INTEGER NOT NULL,
text TEXT NOT NULL,
overlap_left INTEGER NOT NULL DEFAULT 0,
overlap_right INTEGER NOT NULL DEFAULT 0,
boilerplate_score REAL NOT NULL DEFAULT 0.0,
document_type TEXT,
created_at TIMESTAMPTZ NOT NULL DEFAULT now(),
UNIQUE(document_id, chunk_id)
);
CREATE INDEX IF NOT EXISTS idx_v3_document_chunks_document
ON v3_document_chunks(document_id);
CREATE INDEX IF NOT EXISTS idx_v3_document_chunks_type
ON v3_document_chunks(document_type)
WHERE document_type IS NOT NULL;
-- ─── v3_evidence_spans ────────────────────────────────────────────────────────
-- Exact source text with character offsets for provenance.
CREATE TABLE IF NOT EXISTS v3_evidence_spans (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
document_id UUID NOT NULL,
chunk_id TEXT,
start_char INTEGER NOT NULL,
end_char INTEGER NOT NULL,
text TEXT NOT NULL,
checksum TEXT NOT NULL,
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_evidence_spans_document
ON v3_evidence_spans(document_id);
CREATE INDEX IF NOT EXISTS idx_v3_evidence_spans_checksum
ON v3_evidence_spans(checksum);
-- ═══════════════════════════════════════════════════════════════════════════════
-- 20.3: Extracted entities, facts, relations, and rejected candidates
-- ═══════════════════════════════════════════════════════════════════════════════
-- ─── v3_extracted_entities ────────────────────────────────────────────────────
-- Entities discovered during extraction (companies, people, orgs, etc.).
CREATE TABLE IF NOT EXISTS v3_extracted_entities (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
pipeline_run_id UUID NOT NULL REFERENCES v3_pipeline_runs(id) ON DELETE CASCADE,
entity_type TEXT NOT NULL,
literal_text TEXT NOT NULL,
canonical_id UUID,
evidence_span_id UUID REFERENCES v3_evidence_spans(id) ON DELETE SET NULL,
confidence REAL NOT NULL DEFAULT 0.0,
derivation TEXT NOT NULL DEFAULT 'specialist',
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_extracted_entities_pipeline
ON v3_extracted_entities(pipeline_run_id);
CREATE INDEX IF NOT EXISTS idx_v3_extracted_entities_canonical
ON v3_extracted_entities(canonical_id)
WHERE canonical_id IS NOT NULL;
CREATE INDEX IF NOT EXISTS idx_v3_extracted_entities_type
ON v3_extracted_entities(entity_type);
-- ─── v3_extracted_facts ───────────────────────────────────────────────────────
-- Structured facts (numeric values, dates, amounts, etc.).
CREATE TABLE IF NOT EXISTS v3_extracted_facts (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
pipeline_run_id UUID NOT NULL REFERENCES v3_pipeline_runs(id) ON DELETE CASCADE,
fact_type TEXT NOT NULL,
subject_entity_id UUID REFERENCES v3_extracted_entities(id) ON DELETE SET NULL,
predicate TEXT NOT NULL,
literal_value TEXT NOT NULL,
normalized_value JSONB,
unit TEXT,
period JSONB,
evidence_span_ids UUID[] NOT NULL DEFAULT '{}',
confidence REAL NOT NULL DEFAULT 0.0,
derivation TEXT NOT NULL DEFAULT 'deterministic',
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_extracted_facts_pipeline
ON v3_extracted_facts(pipeline_run_id);
CREATE INDEX IF NOT EXISTS idx_v3_extracted_facts_subject
ON v3_extracted_facts(subject_entity_id)
WHERE subject_entity_id IS NOT NULL;
CREATE INDEX IF NOT EXISTS idx_v3_extracted_facts_type
ON v3_extracted_facts(fact_type);
-- ─── v3_extracted_relations ───────────────────────────────────────────────────
-- Relations between entities (competes_with, supplies, etc.).
CREATE TABLE IF NOT EXISTS v3_extracted_relations (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
pipeline_run_id UUID NOT NULL REFERENCES v3_pipeline_runs(id) ON DELETE CASCADE,
relation_type TEXT NOT NULL,
source_entity_id UUID NOT NULL REFERENCES v3_extracted_entities(id) ON DELETE CASCADE,
target_entity_id UUID NOT NULL REFERENCES v3_extracted_entities(id) ON DELETE CASCADE,
evidence_span_ids UUID[] NOT NULL DEFAULT '{}',
confidence REAL NOT NULL DEFAULT 0.0,
derivation TEXT NOT NULL DEFAULT 'specialist',
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_extracted_relations_pipeline
ON v3_extracted_relations(pipeline_run_id);
CREATE INDEX IF NOT EXISTS idx_v3_extracted_relations_source
ON v3_extracted_relations(source_entity_id);
CREATE INDEX IF NOT EXISTS idx_v3_extracted_relations_target
ON v3_extracted_relations(target_entity_id);
-- ─── v3_rejected_candidates ──────────────────────────────────────────────────
-- Candidates that failed validation or were rejected by a stage.
CREATE TABLE IF NOT EXISTS v3_rejected_candidates (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
pipeline_run_id UUID NOT NULL REFERENCES v3_pipeline_runs(id) ON DELETE CASCADE,
candidate_type TEXT NOT NULL,
candidate_data JSONB NOT NULL,
rejection_reason TEXT NOT NULL,
stage TEXT NOT NULL,
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_rejected_candidates_pipeline
ON v3_rejected_candidates(pipeline_run_id);
CREATE INDEX IF NOT EXISTS idx_v3_rejected_candidates_stage
ON v3_rejected_candidates(stage);
-- ═══════════════════════════════════════════════════════════════════════════════
-- 20.4: Company signal candidates and probability distributions
-- ═══════════════════════════════════════════════════════════════════════════════
-- ─── v3_company_signal_candidates ─────────────────────────────────────────────
-- Per-company signal output with full probability distributions.
CREATE TABLE IF NOT EXISTS v3_company_signal_candidates (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
pipeline_run_id UUID NOT NULL REFERENCES v3_pipeline_runs(id) ON DELETE CASCADE,
company_id UUID NOT NULL REFERENCES companies(id) ON DELETE CASCADE,
relevance_probability REAL NOT NULL DEFAULT 0.0,
event_probabilities JSONB NOT NULL DEFAULT '{}',
sentiment_probabilities JSONB NOT NULL DEFAULT '{}',
direction_probabilities JSONB NOT NULL DEFAULT '{}',
horizon_probabilities JSONB NOT NULL DEFAULT '{}',
expected_magnitude REAL,
evidence_span_ids UUID[] NOT NULL DEFAULT '{}',
routing_reasons TEXT[] NOT NULL DEFAULT '{}',
adjudicated BOOLEAN NOT NULL DEFAULT FALSE,
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_signal_candidates_pipeline
ON v3_company_signal_candidates(pipeline_run_id);
CREATE INDEX IF NOT EXISTS idx_v3_signal_candidates_company
ON v3_company_signal_candidates(company_id);
CREATE INDEX IF NOT EXISTS idx_v3_signal_candidates_adjudicated
ON v3_company_signal_candidates(adjudicated)
WHERE adjudicated = TRUE;
-- ═══════════════════════════════════════════════════════════════════════════════
-- 20.5: Adjudication decisions, routing reasons, calibration, and model lineage
-- ═══════════════════════════════════════════════════════════════════════════════
-- ─── v3_stage_lineage ─────────────────────────────────────────────────────────
-- Detailed model/endpoint lineage for each stage invocation.
-- Created before adjudication_decisions because it is referenced as a FK.
CREATE TABLE IF NOT EXISTS v3_stage_lineage (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
stage_run_id UUID NOT NULL REFERENCES v3_stage_runs(id) ON DELETE CASCADE,
endpoint_id UUID REFERENCES inference_endpoints(id) ON DELETE SET NULL,
deployment_id UUID REFERENCES model_deployments(id) ON DELETE SET NULL,
model TEXT,
protocol TEXT,
structured_mode TEXT,
request_id TEXT,
latency_ms INTEGER,
retries INTEGER NOT NULL DEFAULT 0,
trace_id TEXT,
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_stage_lineage_stage_run
ON v3_stage_lineage(stage_run_id);
CREATE INDEX IF NOT EXISTS idx_v3_stage_lineage_endpoint
ON v3_stage_lineage(endpoint_id)
WHERE endpoint_id IS NOT NULL;
-- ─── v3_adjudication_decisions ────────────────────────────────────────────────
-- Decisions made by the 9B adjudicator for ambiguous documents.
CREATE TABLE IF NOT EXISTS v3_adjudication_decisions (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
pipeline_run_id UUID NOT NULL REFERENCES v3_pipeline_runs(id) ON DELETE CASCADE,
question_codes TEXT[] NOT NULL DEFAULT '{}',
candidates JSONB NOT NULL DEFAULT '{}',
decision JSONB NOT NULL DEFAULT '{}',
evidence_span_ids UUID[] NOT NULL DEFAULT '{}',
model_lineage_id UUID REFERENCES v3_stage_lineage(id) ON DELETE SET NULL,
created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_adjudication_pipeline
ON v3_adjudication_decisions(pipeline_run_id);
-- ─── v3_routing_decisions ─────────────────────────────────────────────────────
-- Records of fast-path vs adjudication routing decisions.
CREATE TABLE IF NOT EXISTS v3_routing_decisions (
id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
pipeline_run_id UUID NOT NULL REFERENCES v3_pipeline_runs(id) ON DELETE CASCADE,
document_id UUID NOT NULL,
route TEXT NOT NULL CHECK (route IN ('fast_path', 'adjudication')),
reason_codes TEXT[] NOT NULL DEFAULT '{}',
confidence_features JSONB NOT NULL DEFAULT '{}',
decided_at TIMESTAMPTZ NOT NULL DEFAULT now()
);
CREATE INDEX IF NOT EXISTS idx_v3_routing_pipeline
ON v3_routing_decisions(pipeline_run_id);
CREATE INDEX IF NOT EXISTS idx_v3_routing_route
ON v3_routing_decisions(route);
-- ═══════════════════════════════════════════════════════════════════════════════
-- 20.6: Idempotency and immutable-revision constraints
-- ═══════════════════════════════════════════════════════════════════════════════
-- Pipeline runs: idempotency_key UNIQUE is already defined above in the table.
-- Stage runs: unique per pipeline_run_id + stage to prevent duplicate stage execution.
CREATE UNIQUE INDEX IF NOT EXISTS idx_v3_stage_runs_idempotent
ON v3_stage_runs(pipeline_run_id, stage);
-- Company signal candidates: unique per pipeline_run_id + company_id.
CREATE UNIQUE INDEX IF NOT EXISTS idx_v3_signal_candidates_idempotent
ON v3_company_signal_candidates(pipeline_run_id, company_id);
-- Routing decisions: unique per pipeline_run_id (one routing decision per run).
CREATE UNIQUE INDEX IF NOT EXISTS idx_v3_routing_idempotent
ON v3_routing_decisions(pipeline_run_id);
-- Evidence spans: unique by document + checksum to avoid storing duplicates.
CREATE UNIQUE INDEX IF NOT EXISTS idx_v3_evidence_spans_idempotent
ON v3_evidence_spans(document_id, checksum);
-- Immutable revision rule: pipeline_runs and stage_runs cannot be updated once completed.
-- Enforced via trigger: reject updates to rows where status = 'completed' or 'failed'.
CREATE OR REPLACE FUNCTION v3_immutable_completed_row()
RETURNS TRIGGER AS $$
BEGIN
IF OLD.status IN ('completed', 'failed') THEN
RAISE EXCEPTION 'Cannot modify a % record with status=%', TG_TABLE_NAME, OLD.status;
END IF;
RETURN NEW;
END;
$$ LANGUAGE plpgsql;
DROP TRIGGER IF EXISTS trg_v3_pipeline_runs_immutable ON v3_pipeline_runs;
CREATE TRIGGER trg_v3_pipeline_runs_immutable
BEFORE UPDATE ON v3_pipeline_runs
FOR EACH ROW EXECUTE FUNCTION v3_immutable_completed_row();
DROP TRIGGER IF EXISTS trg_v3_stage_runs_immutable ON v3_stage_runs;
CREATE TRIGGER trg_v3_stage_runs_immutable
BEFORE UPDATE ON v3_stage_runs
FOR EACH ROW EXECUTE FUNCTION v3_immutable_completed_row();