CLI

The package installs two console scripts: narratoflow and the shorter alias narrato. Both expose the same commands.

compress

Compress a text file and emit a JSON envelope.

narratoflow compress input.txt \
  --lang no \
  --provider anthropic \
  --extractor-model claude-haiku-4-5-20251001 \
  --target-model claude-opus-4-7 \
  --schema narrative \
  --out compressed.json

flag	default	meaning
--lang	no	source language ISO code
--provider	anthropic	anthropic or openai
--extractor-model	claude-haiku-4-5-20251001	cheap model used for L3
--target-model	claude-opus-4-7	model token-counter calibrates against
--schema	narrative	preset name or import path
--layer	(all)	repeat to pick a subset, e.g. --layer preprocess --layer codebook
--out	(stdout)	output file

eval

End-to-end benchmark: tokens, cost savings, LLM-judge quality.

narratoflow eval input.txt \
  --target-task "Skriv en kort fortelling (200 ord) på norsk." \
  --provider openai \
  --extractor-model gpt-4o-mini \
  --target-model gpt-4o

Add --skip-quality for token-only runs (cheaper).

prompt

Build a ready-to-send prompt from an existing compress JSON output.

narratoflow prompt compressed.json \
  --instruction "Skriv en kort fortelling basert på fakta." \
  --target-lang no

Environment

Provider keys are read from:

• ANTHROPIC_API_KEY
• OPENAI_API_KEY

If you keep them in .env, source it in your shell or load it via python-dotenv in your own wrapper.