Knowledge Ingestion
File, URL, and text import with chunking, normalization, and provenance-aware persistence. Entry point for all unstructured input.
Architecture
Deterministic Knowledge Infrastructure
Atlas implements a linear pipeline from unstructured documents to governed knowledge. Each layer has dedicated Prisma models, service modules, API routes, UI surfaces, and benchmark metrics. Nothing is implicit — provenance, verification, and governance are first-class.
End-to-end pipeline
Documents
Import
Normalization
Claim Extraction
Evidence
Relationships
Entities
Knowledge Graph
Semantic Layer
Verification
AI Reasoning
Governance
Knowledge Graph
Claims, entities, relationships, and evidence form a queryable directed graph. Builder and service layers expose statistics and node overlays.
Evolution Engine
Snapshots and change detectors track knowledge mutations. Evolution runs emit append-only events for audit and diff inspection.
Semantic Layer
BM25 lexical search combined with pgvector embeddings. Hybrid ranking with configurable OpenAI or Voyage AI embedding providers.
Verification
Deterministic verification runs label claims before they enter reasoning or canonical mutation paths. Results persist per claim.
Reasoning
Retrieval-grounded interpretation via OpenAI, Anthropic, or local providers. Suggestions require human review before acceptance.
Governance (AIGov)
Deterministic policy engine with no AI in decision paths. Checks gate import, verification, reasoning, and suggestion acceptance.
Evaluation
Gold-standard benchmark datasets, thirteen metrics, offline runner, and regression-gated CI. Scientifically reproducible measurement.
Data boundaries
- Provenance: every document, claim, and entity carries source metadata and governance labels.
- Verification: claims receive explicit labels before reasoning uses them as primary evidence.
- Governance: policy decisions are deterministic, append-only, and auditable — no silent canonical mutation.
- Evaluation: benchmark metrics are computed offline against gold-standard datasets with regression detection in CI.