Deterministic Knowledge Infrastructure
Deterministic Knowledge Infrastructure
Transform documents into verifiable knowledge with provenance, semantic retrieval, verification, reasoning and governance.
Atlas transforms unstructured information into verifiable knowledge through deterministic extraction, semantic enrichment, verification, AI reasoning and governance.
Atlas is not
- ✕a chatbot
- ✕a note-taking app
- ✕document search
- ✕generic RAG
Atlas is
- ✓a deterministic knowledge pipeline
- ✓a knowledge graph platform
- ✓provenance-first architecture
- ✓a verification-first system
- ✓governance-native infrastructure
Knowledge pipeline
Every stage is explicit, auditable, and implemented as a separate subsystem with API routes, persistence, and tests.
Subsystems
knowledge ingestion
Knowledge Ingestion
Import documents from files, URLs, or raw text. Content passes through chunking, normalization, and persistence with provenance metadata attached to every record.
- ▸POST /api/import/file, /text, /url
- ▸Chunker + normalizer + persister pipeline
- ▸Document, chunk, and metadata models in Prisma
knowledge graph
Knowledge Graph
Claims, entities, relationships, and evidence form a directed graph. React Flow renders the overlay; graph statistics and node detail APIs expose structure for inspection.
- ▸GET /api/graph, /api/graph/node/:id
- ▸Entity and relationship extraction with provenance
- ▸Interactive visualization at /app/knowledge/graph
knowledge evolution
Knowledge Evolution
Snapshots detect changes across the knowledge base. Evolution runs produce append-only change events so canonical mutations remain traceable over time.
- ▸POST /api/evolution/run
- ▸Detector, snapshot, and filter modules
- ▸Run history at /app/knowledge/evolution/runs
semantic search
Semantic Search
Hybrid retrieval combines BM25 lexical scoring with pgvector embeddings. Provider abstraction supports OpenAI and Voyage AI embedding models.
- ▸POST /api/search/semantic
- ▸Configurable embedding providers
- ▸Precision@k, recall@k, and nDCG in benchmark suite
verification
Verification Engine
Claims receive verification labels through deterministic runs. Results persist per claim and feed governance policy checks before reasoning uses contradicted evidence.
- ▸POST /api/verification/run
- ▸Per-claim results at /api/verification/results/:claimId
- ▸verification_accuracy metric in evaluation
ai reasoning
AI Reasoning
Reasoning runs retrieve grounded context, invoke OpenAI, Anthropic, or local providers, and emit suggestions for human review. Accepted suggestions require explicit approval.
- ▸POST /api/reasoning/run
- ▸Suggestion accept/reject with audit trail
- ▸Citation coverage and verification compliance metrics
governance
Governance (AIGov)
Deterministic policy engine — no AI in policy decisions. Governance checks gate import, verification, reasoning, and suggestion acceptance. Human approval records are append-only.
- ▸POST /api/governance/check
- ▸Policy decisions, violations, and approval flows
- ▸governance_policy_compliance benchmark metric
benchmark suite
Benchmark Suite
Gold-standard datasets in atlas-benchmark/ drive offline evaluation. Thirteen metrics cover extraction, retrieval, reasoning, and governance. Regression detection gates CI.
- ▸pnpm run evaluate:benchmark
- ▸Suite: atlas-v1 across 4 domains
- ▸Exports: JSON, CSV, Markdown reports
Benchmark suite
Atlas ships a deterministic evaluation engine with gold-standard datasets across four domains. Metrics are computed offline; CI fails when regression exceeds a configurable threshold.
Domains: ai-governance, medical, legal, software
Deterministic evaluation — no LLM calls in the metric computation path. Regression testing via EVALUATION_REGRESSION_THRESHOLD.
Platform preview
Knowledge workspace and pipeline dashboards in light and dark themes.
Technology stack
Production dependencies aligned with the actual implementation in apps/atlas-frontend.
Build on deterministic knowledge
Install locally, run migrations, and execute the benchmark gate before deploying.