Canonical schema

Pecca needs four fields. You map your columns to them in connect() or datasource.columns:

Field	Type	Notes
input	str, or dict[str, str] of fields	Multiple fields are joined by input_template, e.g. "{subject}\n\n{body}". Numeric dicts make a tabular call.
llm_output	str	The raw LLM answer.
human_label	str or null	A correction, e.g. an agent's tag.
call_name	str	Column name or {literal: route_rfi}.

Optional, passed through when present: timestamp, llm_model, language, channel, segment, request_id, latency_ms, cost_usd. Anything else is kept in extra.

Label resolution: label = human_label if not null else llm_output (outputs are normalised: stripped, lower-cased, surrounding quotes/punctuation removed). Each row records its label_source (human or llm).

Late-arriving human labels can be joined from another table with labels: {type, table/path, join_on, column}; the datasource must map request_id.