Architecture

An Evidence-Backed Graph Compiler

Nexum compiles unstructured documentation, OpenAPI specifications, code repositories, and raw data records into a deterministic identity graph—the computation layer that gives autonomous models precise, traversable context.

Sources
Docs, OpenAPI, Repos, Records
Ingestion Engine
Recursive crawl & extraction
Graph Core
Identity resolution & linking
Output
Typed graph draft + apply plan

No writes to host data stores at any stage — Nexum is a pure, stateless compute layer.

01 — Deterministic Traversal

Exact authority, not fuzzy probability

Traditional vector RAG returns the "most similar" chunk of text—an approximation that breaks down the moment an agent needs to reason across systems. Nexum instead merges node identities precisely, so a customer record, an API endpoint, and a support ticket that all reference the same entity resolve to a single, unambiguous node.

Agents traverse explicit, typed edges instead of guessing at semantic proximity, which means cross-system relationships stay safe to navigate even as the graph scales.

{
  "node": "customer:8841",
  "type": "Account",
  "edges": [
    { "rel": "OWNS", "target": "subscription:2291" },
    { "rel": "FILED", "target": "ticket:5563" }
  ],
  "confidence": "exact",
  "source_hash": "sha256:9f2a..."
}
📄 Markdown documentation
🧩 Agent skill sources
🗂️ Code repositories
🔌 OpenAPI specifications
Linked, standardized graph schema
02 — Recursive Ingestion Engine

Point it at a source. It does the rest.

Nexum automatically crawls markdown documents, external skill sources, code repositories, and OpenAPI specs to extract logical schemas. It follows references recursively—so a doc that links to a spec that describes a repository all collapse into one coherent structure.

No hand-written ETL. No brittle point-to-point integrations that break on every schema change.

03 — Stateless Graph Core

A draft and a plan — never a write

Nexum never writes directly to your host data stores. Every run returns a strictly typed graph draft alongside an apply plan: a reviewable, deterministic set of changes that your own systems—or a downstream governance layer like LSCP—can choose to execute.

That statelessness is what makes Nexum safe to drop into multi-agent workflows: it computes, it never commits.

{
  "apply_plan": {
    "creates": 12,
    "merges": 4,
    "conflicts": 0,
    "requires_approval": true
  },
  "status": "draft",
  "written_to_host": false
}

Need governance on top of the graph?

LSCP enforces zero-trust boundaries on exactly what agents can extract, observe, and execute against a Nexum-compiled graph.

See LSCP's Governance Architecture →