Knowledge

Knowledge is one graph per environment that your people, apps and agents write as they work. You write episodes (what someone said or did, with its source) into spaces; Knowledge keeps them as primary records and answers recall with the facts the caller may read, each citing the episodes it came from. Nothing needs to be set up first: the graph exists as soon as you create its first space.

Everything is under https://api.sylphx.com/v1/orgs/{org}/projects/{project}/envs/{env}/, and an environment outside the key's scope answers 404.

#1. Create a space

curl -X POST "https://api.sylphx.com/v1/orgs/acme/projects/shop/envs/production/spaces?space_id=support" \
  -H "Authorization: Bearer $SYLPHX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"spec":{"kind":"workspace"}}'

A space is the unit of permission. Its creator may read and write it; grant others through the space's members. A sensitive space is read only by keys it names directly.

#2. Write an episode

curl -X POST "https://api.sylphx.com/v1/orgs/acme/projects/shop/envs/production/spaces/support/episodes" \
  -H "Authorization: Bearer $SYLPHX_API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{"kind":"note","text":"Ada moved to London in March.","provenance":{"references":[{"uri":"urn:support:ticket:4812"}]}}'

Every episode names its source in provenance.references; Knowledge records who wrote it from your key. The episode is stored at once, indexed for recall shortly after, and its facts are extracted in the background.

#3. Recall with citations

const { context } = await sylphx.knowledge.entities.recall({
  parent: 'orgs/acme/projects/shop/envs/production',
  query: 'Where does Ada live?',
  spaces: ['orgs/acme/projects/shop/envs/production/spaces/support'],
})

Recall searches only the spaces you name and the caller may read, expands up to two hops through the graph, and returns facts with the episodes that support them. A claim whose source changed since it was pinned is marked stale and must be re-read before use.

#Erase

Deleting an episode erases it and everything derived only from it; it is never answered again, even from a replayed request.