Your data in.
Knowledge graphs out.
WtrDB takes the data your users put in — tables, records, text, and live operational context — and turns it into a governed, federated, queryable knowledge graph with a mathematical quality engine, a 3D workbench, and agent-facing APIs.
Vector search finds paragraphs. WtrDB connects facts.
Traditional RAG retrieves chunks of text. WtrDB ingests the data your business already runs on, extracts typed entities, maps relationships, and turns it into a knowledge graph your agents can actually reason over.
Four steps from raw data to knowledge graph
Ingest
Bring in PDFs, DOCX, TXT, URLs, tables, and operational records. WtrDB unifies unstructured content and structured data in one ingestion pipeline.
Extract
GPT-4o-mini extracts typed entities and relationships. 13-type taxonomy. Two tracks: static triples and evolutionary events with intent classification.
Adjudicate
Three sequential filters: evidence verification, logical verification, evolutionary-intent verification. Nothing deleted. Contradictions soft-deprecated with full lineage.
Quish
Natural language → Cypher → typed answers with provenance. Publish as Brain Endpoint for any AI agent.
Seven pillars of knowledge infrastructure
Evidence Substrate
Unified ingestion for records, tables, text streams, and operational inputs with chunking, embeddings, and hybrid retrieval.
Knowledge Graph Engine
Incremental, governed, dual-track extraction into typed, temporal graphs via FalkorDB.
Sheaf Quality Engine
Mathematical consistency measurement via cellular sheaf theory. Spectral gap, H⁰/H¹ cohomology, sheaf diffusion.
Federation
Union, Intersection, Differential, Sheaf-Augmented merge modes. Entity alignment, conflict workflows, schema proposals.
The Void (3D Workbench)
Continuous WebGL 2.0 environment. Type-specific wireframe entity geometries. Temporal Rewind, Provenance Thread, Ghost Overlay, Creation Theatre.
Brain Endpoints
Publish any KG as REST + MCP + SSE backend. Per-endpoint API keys, rate limits, model overrides.
Enterprise Spine
RBAC, TOTP MFA, SOC 2 / HIPAA / CMMC compliance, Stripe billing. Built in, not bolted on.
Every fact is earned, not assumed.
Evidence Verification
LLM judge confirms the source chunk actually supports the claim. Catches hallucinated extractions before graph entry.
Logical Verification
Checks candidate fact against current graph state for type violations and contradictions.
Evolutionary-Intent Verification
Classifies facts as Informational or Evolutionary. Evolutionary facts trigger soft deprecation of stale entries (status='Deprecated', _invalidAt=now()).
Conservative acceptance — nothing is ever deleted. Ambiguous facts enter with lower confidence rather than being dropped. Full audit trail of every accept / reject / deprecate decision.
Graph quality with a mathematical definition.
WtrDB uses cellular sheaf theory. Each entity gets a stalk (embedding vector in ℝᵈ); each relationship gets restriction maps measuring endpoint compatibility. The Sheaf Laplacian LF encodes global consistency. H¹ cohomology reveals conflict cycle topology — not a flat list.
Spectral Gap
Smallest nonzero eigenvalue of L_F. Higher = more globally consistent. The number you show an auditor.
Conflict Energy
E = xᵀ L_F x. Ranks which subgraphs are most conflicted.
H⁰ Cohomology
Connected, internally consistent components. How many coherent islands exist.
H¹ Cohomology
Independent conflict cycles. Resolve a cycle, clear multiple conflicts at once.
Sheaf Diffusion
Heat equation dh/dt = −L_F h. Suggests resolutions. Never auto-overwrites.
Gebhart et al. “Knowledge Sheaves” arXiv:2110.03789 · Hansen & Ghrist “Toward a Spectral Theory of Cellular Sheaves” · Bodnar et al. “Neural Sheaf Diffusion” NeurIPS 2022 · Robinson “Sheaves Are the Canonical Data Structure for Information Integration”
Merge multiple graphs with formal algebra.
Union
Keep all entities from all graphs. Maximum coverage.
Intersection
Keep only what all graphs agree on. Maximum consensus.
Differential
Keep entities unique to one graph. Reveals gaps.
Sheaf-Augmented
Resolve conflicts using sheaf cohomology and spectral gap signals. The mathematically informed merge.
One URL. Any agent.
Brain Endpoints collapse “I have a knowledge base” and “my agent can query it” into a single governed handoff.
| Transport | Endpoints |
|---|---|
| REST | POST /ask (NL question → structured answer) · POST /cypher (raw Cypher, if enabled) · GET /schema · GET /health |
| MCP | POST /mcp — full Model Context Protocol tool surface for Claude Desktop, Cursor, Codex |
| SSE | GET /sse — streaming responses |
| Self-describing | GET /llms.txt, /llms-full.txt — agent discovery |
Per-endpoint config — API key · rateLimitPerMinute · queryModel + queryReasoningEffort · traversalModel + traversalReasoningEffort · allowNlQueries · allowCypher (off by default) · expiresAt
A workbench you can actually live in.
No pages. No dashboards. A continuous 3D environment rendered in WebGL 2.0 where every knowledge graph is a star and navigation is camera movement.
The Constellation
Home. Every KG is a star; federations are lines between stars.
Graph Explorer
Full 3D graph: type-specific wireframe geometries (meshframes), edges, temporal rewind, provenance.
The Workbench
Federation merge playground: drop graphs, choose merge mode, see Venn/gap/heatmap as 3D objects.
Creation Theatre
Watch live extraction happen in 3D with commentary and slow-motion replay.
The Observatory
Quality measurement and evaluation tooling.
The Pulse Array
Real-time agent activity feed.
Meshframe taxonomy — 13 entity types, 5 geometry classes. Visual properties are data-driven: connection count → scale, freshness → opacity, confidence → edge style, status (Active / Deprecated / Contested) → edge rendering.
Enterprise-grade from day one.
RBAC
System roles (admin, manager, user, viewer) + custom org-scoped roles + team scoping + full audit trail.
MFA
TOTP with QR code. CMMC IA.2.081 compatible.
Compliance
CMMC, SOC 2, HIPAA. One-click auto-configure. Continuous evidence collection. Scheduled exports. OPA policy-as-code via Claw engine.
Multi-tenancy
Organizations, teams, per-tenant config, plan-scoped quotas.
Billing
Credit ledger, Stripe, invoices, refunds.
Deployment
Ubuntu/Debian apt package, self-host ready, systemd daemon, air-gapped enterprise option.
Not just another RAG pipeline.
| Capability | Most RAG + Graph Products | WtrDB |
|---|---|---|
| Fact validation | Trust the LLM | Three-filter governance pipeline |
| Conflict handling | Overwrite or ignore | Soft deprecation with full historical lineage |
| Quality measurement | Heuristic scores | Sheaf Laplacian spectral gap + H¹ cohomology |
| Multi-graph merge | Manual or absent | Formal merge algebra with 4 modes |
| Visualization | 2D force-directed | 3D spatial environment with type-specific meshframes |
| Agent integration | Custom glue code | Brain Endpoints: REST + MCP + SSE |
| Enterprise readiness | Bolted on | Auth, RBAC, MFA, compliance, billing in the spine |
Turn your company knowledge into an AI brain.
Join organizations already powering their AI agents with WtrDB knowledge graphs.
See how WtrDB is applied in regulated financial workflows and construction operations.