Store Chat Messages & State Without Managing Infrastructure.Check Out DialogueDB
llm-exe
v3.0.1-1783122442

Appearance

Back to examples

Yes/No Decisions ​

Sometimes you don't need a paragraph from an LLM — you need a boolean. "Does this comment contain a question?" "Is this ticket urgent?" "Should this PR run the expensive test suite?" This recipe builds a function that asks the LLM a question and returns an actual true or false, so the calling code can branch on it directly.

The key is the boolean parser: it converts the model's text answer ("yes", "No.", "true") into a real boolean, and throws if the response doesn't contain one — so your if statement never runs on a malformed answer.

Step 1 - Imports ​

ts
import {
  useLlm,
  createChatPrompt,
  createParser,
  createLlmExecutor,
} from "llm-exe";

Step 2 - Prepare Prompt ​

The prompt does two jobs: it asks the question, and it constrains the answer format. Telling the model it must answer yes or no (even when unsure) avoids hedged responses that can't be parsed.

ts
export const PROMPT = `Answer the following question with only "yes" or "no".

Do not explain your reasoning. Do not add punctuation. If you are unsure,
make your best judgment call - you must answer yes or no.

Question: {{question}}`;

Step 3 - Create the Function ​

The parser is boolean with match: "extract", which finds a yes/no/true/false token inside the response even if the model adds punctuation or stray words. If the response contains no boolean — or contains both a yes and a no — the parser throws instead of guessing.

ts
export async function yesNo(question: string): Promise<boolean> {
  const llm = useLlm("openai.gpt-4o-mini");
  const prompt = createChatPrompt<{ question: string }>(PROMPT);
  const parser = createParser("boolean", { match: "extract" });

  return createLlmExecutor({
    name: "yes-no-decision",
    llm,
    prompt,
    parser,
  }).execute({ question });
}

Step 4 - Use it! ​

typescript
import { yesNo } from "./somewhere";

if (await yesNo("Does this comment contain a question?")) {
  await routeToSupport(comment);
}

const shouldRunE2e = await yesNo(
  `Based on these changed files, could this PR affect the checkout flow?\n${changedFiles.join("\n")}`
);

Because the executor's return type is inferred from the parser, yesNo is a Promise<boolean> — no casting, no .toLowerCase().includes("yes") string checks.

Where this pattern fits ​

  • Pre-checks and gating — decide whether to run an expensive step (moderation review, e2e tests, human escalation)
  • Content filtering — "does this message need a disclaimer?", "is this on-topic?"
  • Input screening — a first-pass gate on untrusted input; for the full fail-closed pattern see Prompt Injection Screening
  • Workflow conditions — a typed boolean drops straight into if: conditions in CI or any orchestration code

For decisions with more than two outcomes, use the same shape with a category parser instead — see Intent Classification and Conditional Logic and Branching.