ai-output-validate Edge Function — Ai

ai-output-validate — Validate and repair LLM output format

Checks that an LLM's raw text output conforms to an expected schema type (JSON, list, boolean, number, email, URL), with optional auto-repair for common formatting mistakes.

---

API

POST /api/ai-output-validate

Request body

Field	Type	Required	Default	Description
output	string	✅	—	Raw text output from the LLM
schema	object	✅	—	Validation schema (see below)
repair	boolean	❌	false	Attempt to auto-fix minor formatting issues

Schema object

Field	Type	Description
type	"json" | "list" | "boolean" | "number" | "email" | "url"	Expected output type
required	string[]	(JSON only) Required keys in the parsed object
minItems	number	(list only) Minimum number of items
maxItems	number	(list only) Maximum number of items

Success response (200)

{
  "success": true,
  "data": {
    "valid": true,
    "parsed": {"name": "Alice", "age": 30},
    "errors": [],
    "repaired": false
  }
}

Error responses

Code	HTTP	When
INVALID_INPUT	400	Missing required field
INTERNAL_ERROR	500	Unknown schema type

---

Usage

cURL

curl -X POST "$FUNCTION_URL" \
  -H "Content-Type: application/json" \
  -d '{
    "output": "```json\n{\"name\": \"Alice\"}\n```",
    "schema": {"type": "json", "required": ["name"]},
    "repair": true
  }'

TypeScript / JavaScript (HTTP)

const response = await fetch(FUNCTION_URL, {
  method: 'POST',
  headers: { 'Content-Type': 'application/json' },
  body: JSON.stringify({
    output: llmResponse,
    schema: { type: 'json', required: ['title', 'summary'] },
    repair: true,
  }),
});
const { data } = await response.json();
if (data.valid) console.log(data.parsed);

Direct import (Node / Bun / Deno)

import { aiOutputValidate } from '@aerostack/functions/ai-output-validate';

const { valid, parsed, errors } = aiOutputValidate({
  output: rawLlmText,
  schema: { type: 'boolean' },
  repair: true,
});

---

Use Cases

• Validating that an LLM's structured output is parseable JSON before inserting it into a database.
• Checking that a classification output is a valid boolean (yes/no) before branching workflow logic.
• Extracting and verifying a URL or email from a generated text response.

---

Notes

• repair: true for JSON tries to extract content from markdown code fences (```json ... ```).
• repair: true for boolean maps common words (yes/no/yep/nope/correct/incorrect) to true/false.
• List detection: newline-separated lists take priority over comma-separated.
• repaired: true in the response means the output was modified before parsing.