Extract

Extract turns documents into atoms. It runs whenever a document is

uploaded or re-uploaded, and re-running on the same document

produces zero new atoms (idempotency by identityKey).

How it works

1. Document is uploaded; documentParse row is created
2. The cron worker /api/cron/process-extractions picks it up
3. extractAgreementGraph calls the LLM with prompts from

src/lib/substrate/extraction/prompts/

1. Each returned atom is validated against the kind's Zod schema in

src/lib/substrate/schemas/ then written via createAtom

1. Edges are written via createEdge to record the typed relations

(e.g. defines, references)

Source quotes are mandatory

Every atom carries sourceQuote, sourcePage, sourceCharStart,

sourceCharEnd. The validator rejects atoms without; "I think this

is in there" is not a Nebula atom.

Domain pack effects

Active packs (defence, commonwealth_compliance, infrastructure)

contribute extra prompt instructions, custom edge types, and

post-extraction validators. See /docs/security/residency for the

defence pack's classification atom emission rules.

Cross links

• /docs/concepts/atoms: what an atom is
• /docs/api/atoms: API endpoints over atoms