Preparing a schema

LPGCast casts into a graflo Schema, or the schema block of a GraphManifest, in YAML or JSON. Anything graflo accepts works; a few things decide how well a model reads it. Check the result with lpgcast sheet --schema schema.yaml: that is everything the model sees.

Describe types in the reader's words

• description on a vertex or edge is printed in the type sheet. One sentence saying what an instance is beats a definition.
• semantics.synonyms are other names the text may use (press, station for a Machine). They are printed in the sheet and they seed selection: a text that says "press" pulls in Machine.

Make identities extractable

• A natural identity (identity: [serial]) must be something texts state; a vertex whose identity the text does not give is left out.
• Use hash_identity_properties when the key is a combination of properties the text gives.
• Use assigned: true or blank: true for things a text creates without naming: a note, a work order, a comment.
• Declare secondary_identities for keys texts use to refer to an entity they do not otherwise identify (an email, a phone number, a badge). Such a mention is looked up at write time; see Identities and mentions.

Type the properties

Every property should have a type. Values are coerced to it before writing:

Type	Accepted from the model
INT, UINT	integers, integral floats, digit strings; not 1,000
FLOAT, DOUBLE	numbers and numeric strings
BOOL	true/false, yes/no, 1/0
DATETIME	ISO 8601, kept as a string; 2024 and 2024-03 stay partial
STRING	anything, stripped
LIST	a list, or a single value, each item coerced

A numeric field with semantics.unit (unit: kW) accepts a value written with that unit and strips it. A value in any other unit is refused rather than converted, and units are case-sensitive (mW is not MW). The sheet shows the unit, so the model knows what to expect. The model writes numbers plain (1250) and dates as ISO 8601; the quote keeps the text's own form.

Large schemas

A schema over the slice budget (LPGCAST_SLICE_MAX_ELEMENTS, LPGCAST_SLICE_MAX_TOKENS) is cut down per text: the types the text matches are seeds, and graflo's subschema() closes them into a valid slice, never dropping an identity field. lpgcast select --episodes episodes.jsonl shows each slice and its seeds without calling a model. A text that matches nothing in a large schema is quarantined as selection_no_match rather than cast against the whole schema.