Architecture
See how itworks.
No code required. Press play and watch a document turn into a searchable knowledge graph.
1. Ingest
A PDF - the engine extracts its text, page by page.
What's inside
Nodes and edges.
Every chunk is a node that carries meaning; the links between them are edges. That is the whole data model.
Node
- domain
- Pdf · Email · Codebase · Web · Custom
- kind
- Entity · Topic · Event · Capability
- source_text
- the canonical text that gets embedded
- embeddings
- 384-dim BGE-small vector
- metadata
- tags incl. workspace_id
- zone
- 0 = own · 1 = shared
Edge
- from → to
- a directed link between two nodes
- label
- the relationship (mentions, explains…)
- relationship_probability
- confidence 0.0 - 1.0
Stored in SurrealDB - one graph + vector store.
Library-first
Embed it, or run it.
The logic lives in crates you can link in-process; the servers are thin shells around the same core. Same engine, your choice of surface.
Library core
The engine as linkable crates behind one facade - storage, query, embeddings.
Transport shells
Thin binaries that expose the core as GraphQL subgraphs.
Gateway
Apollo Router composes them into one API on :4001.
Ready to go deeper?
The docs walk through building a real app step by step - or see what others are building.