The pipeline
EDN turns raw content into governed context in a fixed sequence of steps.1
Content enters
Your application sends content to EDN: a conversation round, a document to
upload, or a correction. Each item arrives through a typed endpoint in the
API.
2
Classification
EDN labels each record with a provenance class that records its origin, for
example whether a user stated it, a user confirmed it, it came from a
document, or a model inferred or generated it.
3
Governed storage
Records are stored with their type, provenance and domain. Replaced records
are marked rather than silently overwritten, so history stays traceable.
4
Retrieval
When your application asks for context, EDN finds the stored records that
are relevant to the query, considering only the provenance classes that are
allowed as context. No model chooses the records.
5
Governed context
The selected records are assembled into a compact package, the SIP
(Session Initialisation Package). Unconfirmed AI output never enters it.
Your application places this package in your model’s prompt.
API call flow
Each turn of a conversation takes two calls to EDN: one before your model replies, and one after.
Want users to confirm the assistant’s answers? See
Classify an assistant reply for the optional
confirm flow.
Where a model is used
EDN never answers questions for you. Your model reads the SIP and produces the answer. EDN does use a language model internally to classify some of what it stores: facts stated in a user’s message, corrections, and, on some turns, whether an assistant reply is supported by the context it was given. Building the SIP uses no model.See Provenance for which provenance classes are
used as context, and The SIP for how citations work.