ThinkThen

rank options

The command takes QUESTION or @FILE. It reads one record per line. A pointer from --field or a question file makes it read JSON Lines. Start with its value and exit code.

OptionWhat it does
--top NPrints the first N records of the order. Every record is still judged.
--true TEXTWhat a yes means, in the words the model reads.
--false TEXTWhat a no means.
--input FILEReads the evidence from a file instead of standard input.
--planPrints the plan and sends nothing. It needs no key.
--lines, --jsonl, --csv, --tsvSays how a stream of records is framed. Pick at most one.
--field POINTERNames the part of each record to judge, as a JSON Pointer. It may repeat.
--jobs NHow many requests run at once, from 1 to 32. The default is 4. It works on a stream of records, annotate on one document, and relate on one entity set.
--model NAMEThe model the request carries. Name a version to pin a run. The default is jev-1.13.0.
--url BASEThe server's base address. It outranks THINKTHEN_BASE_URL.
--timeout SECONDSHow long one attempt may take. The default is 30.
--max-retries NHow many times a retried status is sent again. A transport failure is never sent again. The default is 3.
--record DIRCalls the backend and saves each exchange in DIR.
--replay DIRAnswers from DIR alone, with no key and no network.
--cache DIRAnswers from DIR when it can and saves new exchanges there.
--no-cacheTurns off the answer cache for one run.
--profile FILEApplies local backend limits and names the calibration profile in use.

More examples

The checkout crash stops customers from buying. --top 1 prints it alone, id and all.
question="Does this stop customers from buying?"

cat <<'EOF' |
{"id": "B-1", "title": "Typo on the About page"}
{"id": "B-2", "title": "Checkout crashes for every customer"}
{"id": "B-3", "title": "Logo looks blurry on large screens"}
EOF
thinkthen rank "$question" \
  --batch 1 \
  --jsonl \
  --field /title \
  --top 1 |
jq .
Output
{
  "id": "B-2",
  "title": "Checkout crashes for every customer"
}
exit 0