Module chart_spec_generation

Structs§

DATA_SAMPLE_HEADING: Introduces the sample of the teacher’s data file. Also how the test-mode mock Azure API finds the sample, which it needs to answer with the field names the teacher’s file actually has.
EXISTING_SPEC_HEADING: Introduces the specification being edited, and so also ends DATA_SAMPLE_HEADING’s section.
MAX_GENERATION_ATTEMPTS 🔒: How many validation attempts the model gets in total (first answer + repair rounds).
MAX_REPORTED_VALIDATION_ERRORS 🔒: Caps for how much validation detail is echoed back to the model and the teacher; union schemas like Vega-Lite’s can produce very many, very long errors.
MAX_UNION_DEPTH 🔒: How many levels of union to follow when explaining why a specification was rejected.
MAX_VALIDATION_ERROR_CHARS 🔒
PLACEHOLDER_DATA_FILE 🔒: Stands in for the teacher’s file while a spec that has none is validated. The Vega-Lite schema requires a data source on a top-level view, so a chart still waiting for its file could not be checked at all otherwise. Never leaves parse_and_validate_spec.
SYSTEM_PROMPT 🔒: System prompt for generating a Vega-Lite chart specification for the CMS chart block.
USER_PROMPT_PREFIX: User prompt prefix. Also how the test-mode mock Azure API recognises a request from this feature, since a JSON-mode request carries no schema name to identify it by.

VEGA_LITE_SCHEMA_JSON 🔒: Vendored copy of https://vega.github.io/schema/vega-lite/v6.json, used to validate generated specs without network access. All of its $refs are internal.

attach_data_file 🔒: Points the spec at the data file the request named.
describe_validation_error 🔒: Describes a validation failure in terms a model can act on, into out.
find_data_source 🔒: Names the first data source anywhere in the spec, or None when it carries none.
generate_chart_spec: Generate a Vega-Lite chart specification from a teacher’s prompt using an LLM with structured JSON output. The result is validated against the Vega-Lite v6 JSON Schema; on failure the validation errors are sent back to the model for one repair round. Returns the specification pretty-printed as a JSON string.
parse_and_validate_spec 🔒: Parses one LLM answer, rejects any data the model supplied, attaches data_file and validates the result against the Vega-Lite schema. The error string describes the failure in a form suitable for feeding back to the model.
remove_data_sources 🔒
response_format 🔒: The format the chart spec LLM is asked to answer in.
sanitize_refs 🔒: Percent-encodes every $ref string in the schema.
text_message 🔒
union_branches 🔒: The branch errors of a union keyword, if this failure is one.
vega_lite_validator 🔒: The Vega-Lite validator built from the vendored schema; built once per process.
without_data 🔒: The spec with every data source taken out, or None when it isn’t parseable JSON.