The LLM API cheat sheet

Quick reference

Chat

$curl -s $BASE/v1/chat/completions ...
$curl -s $BASE/v1/models | jq '.data[].id'
$jq -r '.choices[0].message.content'
$jq -r '.choices[0].finish_reason'
$export BASE=https://api.example.com

Tokens

$jq '.usage.prompt_tokens'
$jq '.usage.completion_tokens'
$jq '.usage.total_tokens'
$wc -c prompt.txt  # ~4 chars per token

Tools

$jq '.tools[].function.name' req.json
$jq '.choices[0].message.tool_calls'
$jq '.choices[0].message.tool_calls[0]'
$jq -r '.choices[0].finish_reason'

Evals

$python3 -m pytest evals -q
$jq -r .score out/*.json | sort -n
$jq '[.[] | select(.pass)] | length'
$git diff --stat evals/baseline.json

Explained

Chat

The call is a POST to $BASE/v1/chat/completions with a model and a messages array. curl -s proves the endpoint answers, and a GET to $BASE/v1/models lists what you can call. Read choices[0].message.content for the answer and finish_reason for why it stopped: stop, length or tool_calls.

Tokens

Every response carries a usage object. jq '.usage.prompt_tokens' and .completion_tokens give the two numbers that bill differently, and total_tokens is their sum. For a rough pre-flight estimate, run wc -c on the prompt and divide by four.

Tools

You describe tools in the request as JSON schemas, each with a name and parameters. The model does not call them; it returns tool_calls with a name and an arguments string. Parse it, validate it against the schema, run your function, and send the result back as a tool message.

Evals

Keep a folder of cases with expected answers and a score per case. Run them on every prompt change with python3 -m pytest evals -q, diff the scores against the baseline in git, and count the passes rather than reading a few outputs. An eval that is not in CI is an opinion.

In one line

Save the shape

$curl -s $BASE/v1/models