Small things, worth knowing.

For English text, one token is roughly four characters, or about three quarters of a word. That is close enough to answer the only question that usually matters before you send a request: will this fit?

const estTokens = (s) => Math.ceil(s.length / 4);

Use it to guard a request, trim context, or warn a user before they paste a novel. When you need the exact number, run the model's real tokenizer, which our Token Counter does in your browser.

Telling a model to "reply in JSON" is a hope, not a guarantee. The reliable approach is structured outputs: you hand the API a JSON Schema and the model is constrained to match it, field for field.

response_format: {
  type: "json_schema",
  json_schema: { name: "extract", schema: { /* your schema */ } }
}

That is the difference between parsing whatever came back and getting exactly the shape you asked for. Build a schema for OpenAI, Anthropic, or Gemini with the Structured Output Builder.

The pattern is a fetch whose body you read in a loop, decoding chunks as they land:

const res = await fetch(url, { method: "POST", body });
const reader = res.body.getReader();
const dec = new TextDecoder();
while (true) {
  const { value, done } = await reader.read();
  if (done) break;
  render(dec.decode(value));
}

Most LLM APIs stream as Server-Sent Events, so each chunk is one or more data: lines you split on newlines and parse. The user sees the first words in well under a second, which is most of why streaming feels fast.

An API key in client-side JavaScript is visible to anyone who opens dev tools, so calls that need a secret key belong on a server. For everything else, two habits keep keys out of git:

# .gitignore
.env
.env.local

Keep keys in a .env file your code reads at runtime, and make sure .env is ignored before the first commit. If a key ever reaches a public repo, rotate it right away. Already committed? .gitignore will not untrack it; run git rm --cached .env first.