How We Easily Extract Information from AI

• 25 March 2025
• ai, code, simplification, parsing

When we first started using AI to build internal business applications at Contember, we quickly faced a tricky problem: How do we easily extract clear, useful information from the AI’s output?

Initially, we tried JSON because it’s a common format. But generating JSON with AI was slow and complicated. It required strict rules with lots of brackets, quotes, and commas. Any small mistake would break the entire output.

{
	entities: [
		{
			name: "Author",
			columns: [
				{
					name: "name",
					type: "string"
				},
				{
					name: "website",
					type: "string",
					required: true
				}
			]
		}
	]
}

When AI forgot a comma, it broke the entire output. We tried asking AI to fix its own mistakes by showing the incorrect JSON and previous context, hoping for a corrected version. But this was only successful about half the time.

{
	entities: [
		{
			name: "Author",
			columns: [
				{
-  				name: "name"
+  				name: "name",
					type: "string"
				},
				{
					name: "website",
					type: "string",
					required: true
				}
			]
		}
	]
}

Next, we switched to YAML. YAML was quicker and simpler, but still required a specific structure. Even with better instructions, we couldn't reliably achieve more than 70% accuracy. Any small mistake still meant regenerating the entire output.

entities:
  - name: Author
    columns:
      - name: name
        type: string
      - name: website
        type: string
        required: true

YAML was easier to generate, but missing just one line broke everything, forcing us to ask the AI for fixes again.

entities:
  - name: Author
    columns:
      - name: name
-        type: string
      - name: website
        type: string
        required: true

Finally, we created our own simple format inspired by SQL. It was easy for AI to generate and easy for us to read and process. If the AI made a mistake, we could easily skip that one line and continue without issues.

Here's our custom format:

create entity Author;
add column Author.name string;
add column Author.website string required;
default column Article.createdAt now;

If AI made a mistake, like missing a column name, everything else was still okay, and the fix was quick—just append a new correct line.

create entity Author;
- add column Author.name string;
add column Author.website string required;
default column Article.createdAt now;

add column Author.email string;

This simple solution dramatically improved our success rate, bringing it close to 100%.

I hope this inspires you to simplify your AI workflows too!

• ← Previous
  Making Shadcn UI Components Simple with UIC