Bernt
754c89506b
- Layer 1: ArchiveArtifact — immutable original with retention policy - Layer 2: KnowledgeArtifact — extracted knowledge (observations, segmentations, feature vectors, relations) - Layer 3: Knowledge Graph / Ontology (documented, not implemented) - Layer 4: Decision Intelligence (existing DecisionCase) - DataLifecycle: tracks every step with artifact lineage - Key principle: AI models trained on curated datasets, not whole archive - Ontology answers 'what does it mean in our domain?' Long-term goal: Every observation converted once to structured knowledge, reused infinitely for analysis, decisions, training. Next: PR-005A — Minimal Mission Import UI for MVP-0
2.1 KiB
2.1 KiB
ADR-011: Four-Layer Data Architecture
Status
Accepted
Context
We need a clear separation between raw data, extracted knowledge, and decision intelligence to avoid building expensive, tightly-coupled AI systems.
Decision
Separate into four layers:
Layer 1: Raw Archive (cheap, immutable)
- Original files never changed
- Hash guarantees integrity
- Storage policy: active (30 days) → archive (long-term, cheaper)
Layer 2: Knowledge Extraction
- Objects, geometry, GPS, classifications, bounding boxes, embeddings
- Observations and Evidence created
- Structured knowledge used for most queries, not video files
Layer 3: Knowledge Graph / Ontology
- Road A12 → has Observation → Crack → belongs to Area → Stockholm
- Answers: "Show all main roads where cracking increased last 12 months"
- Comes from knowledge graph, not by reading video files
Layer 4: Decision Intelligence
- Verified Decision Cases
- Curated datasets for AI training
- Business impact tracking
Key Principles
- Archive Artifact = Original file, never modified
- Knowledge Artifact = Extracted knowledge, can be re-generated
- Ontology before model — taxonomy answers "what does it mean in our domain?"
- AI models trained on curated datasets, not whole archive
Data Lifecycle
Upload → Immutable Archive → Metadata Extraction → Knowledge Extraction
→ Ontology Mapping → Decision Pipeline → Learning → Archive Retention
Each step produces a new Artifact with version history.
Long-term Goal
"Every real observation is converted once to structured knowledge and can then be reused infinitely for analysis, decision support, history, training, and future AI models."
Consequences
Positive
- Raw data = revisionable archive
- Real product = growing knowledge base + verified Decision Cases
- Can swap AI models without losing originals
- Cheaper storage (archive vs active)
Negative
- More complex pipeline
- Need to manage extraction models
- Knowledge graph maintenance
Related
- ADR-002: Artifact is common base contract
- ADR-003: Event Sourcing for traceability