SchemaCast¶
SchemaCast reads short texts — messages, tickets, notes — and writes what they state as a labeled property graph, typed by a GraFlo schema you govern.
Each text gets one model call, checked against the schema. What fails a check is corrected at most twice, then quarantined, never guessed. Every vertex, edge and property keeps the quote it was read from.
It is for teams that already have a schema and want text to land in it.
Early development
SchemaCast is on PyPI (pip install schemacast) and its API may still change; see
Installation.
What it does¶
One call, checked¶
The model sees only the slice of the schema the text is about, and replies in a flat format of vertices, edges and quotes. Types, relations, properties, values, identities and quotes are checked; a reply that fails gets a typed correction, at most twice.
Identities, not guesses¶
A text's mentions are resolved by their keys, never guessed: a complete key identifies an entity, an alternate key looks one up, and anything less waits, staged with its reason. Writing a cast twice leaves a database that upserts unchanged.
Grounded and replayable¶
Every assertion carries its source, character span and quote. Replies are cached by request, so a replay makes no request and reproduces the graph byte for byte.
What to read next¶
-
Install the package and point it at a model.
-
Cast five support tickets into a schema.
-
The pipeline from text to graph, step by step.
-
Score a model against texts whose graphs are known.