Command line

pip install vegaml installs the vegaml command. State in, decide()'s JSON out, ready for jq.

vegaml "I was billed twice, refund me or I cancel." \
    --choice "team=Which team handles this?=billing,technical,sales" \
    --boolean "churn=Does the customer threaten to leave?"

Typed flags build the sheet without JSON#

Each flag repeats. The score's comma list is its ordered level criteria, lowest first. Note what the flags cannot express: per-option definitions for a choice. Those matter, so for anything you will run twice, use a file.

vegaml "..." --questions @questions.json
vegaml "..." --questions '{"team": {"type": "choice", "instructions": "...", "criteria": {...}}}'

--questions combines with the typed flags, and a duplicate id is an error rather than a silent overwrite.

Input and output#

• The state is the positional argument. -, or piping with no argument, reads stdin.
• --pretty indents. The default is one compact JSON object on one line, which is what you want in a pipeline.
• --model picks the checkpoint: nothing for the default 0.8B, 4b, a Hub id, or a path.
• --mode, --alpha, --views, --device and --max-len pass straight through to decide().

Batches#

vegaml --jsonl tickets.jsonl \
    --boolean "urgent=Does this need a reply today?" > answers.jsonl

Each input line is {"state": ..., "questions": {...}?}; a line without questions uses the flags'. The checkpoint loads once, before the first line, which is the entire point: a thousand states pay for one load. Output is one result object per line, in input order.

vegaml --jsonl tickets.jsonl --choice "team=Which team?=billing,technical,sales" \
  | jq -r '.answers.team | select(.abstain == false) | .choice' \
  | sort | uniq -c