Ask OpenAI about your data

Instead of writing a template, describe the result in plain language and let an OpenAI model produce it: a summary, an answer, or a conversion that is hard to express as a template. Experimental.

export OPENAI_API_KEY=sk-...
curl -fs https://www.ecb.europa.eu/stats/eurofxref/eurofxref-daily.xml \
  | bafi -f xml -gq "What is the current CZK rate? Answer with the number only."
  • -gq is the question; it switches bafi into OpenAI mode, no -t needed.
  • The key comes from the OPENAI_API_KEY environment variable. -gk also works but puts the key in your shell history and the process list.
  • -gm picks the model: gpt35 (default), gpt4, gpt4o, gpt4o-mini, gpt5, gpt5mini, gpt5nano.
  • -o answer.txt saves the answer to a file.

What is sent

bafi sends the question and the whole parsed input, encoded as TOON to save tokens. Do not use it with data that must not leave your network. Preview exactly what goes out:

{"invoice": "2024-17", "customer": "Alice Smith", "lines": [{"item": "Coffee", "qty": 2, "price": 12.9}]}

bafi -i invoice.json -t '?{{toTOON .}}'

customer: Alice Smith
invoice: 2024-17
lines[1]{item,price,qty}:
  Coffee,12.9,2

Examples

bafi -i invoice.json -gq "Create a UBL 2.1 XML invoice from this data" -gm gpt4o -o invoice.xml
curl -fs "https://api.open-meteo.com/v1/forecast?latitude=50.08&longitude=14.42&hourly=temperature_2m" \
  | bafi -f json -gm gpt4o-mini -gq "Summarise tomorrow's temperatures in two sentences"

Model output is not deterministic; for repeatable conversions write a template instead and keep OpenAI for exploration.

Related: Convert between JSON, YAML, TOON and BSON ยท Security notes