Flexible tool extraction

📖 Lesson content

Summary

Writing detailed JSON schemas for structured data extraction can be a real pain point when working with AI tools. There's a clever workaround that lets you specify your desired data structure directly in your prompt instead of creating complex schemas.

The Flexible Schema Approach

Instead of writing a detailed schema for every data extraction task, you can create one generic tool called to_json that accepts any object structure. The key is setting the input schema to allow additional properties, then specifying your exact requirements in the prompt itself.

This approach removes a major pain point - constantly writing and managing large JSON schemas. The results won't be quite as good as a dedicated schema, but you'll still get high-quality JSON output with much less setup work.

How It Works

The process is straightforward:

Create a single flexible schema that accepts any object structure
In your prompt, specify exactly what data structure you want
Tell Claude to call the to_json tool with your specified structure
Use tool_choice to force Claude to use your tool

Setting Up the Prompt

When writing your prompt, be very explicit about the structure you want. Here's an example of how to structure your request:

Analyze the article below and extract key data. Then call the to_json tool.

<article_text>
{result["text"]}
</article_text>

When you call to_json, pass in the following structure:
{{
    "title": str # title of the article,
    "author": str # author of the article,
    "topics": List[str] # List of topics mentioned in the article
}}

Making the API Call

The API call uses the flexible schema and forces tool usage:

flexible_result = chat(messages, tools=[to_json_schema], tool_choice="to_json")

Easy Structure Changes

The real advantage becomes clear when you need to modify your data structure. Instead of rewriting an entire schema, you simply update your prompt. Want to add a field for the number of topics? Just add one line:

"num_topics": int # Number of topics mentioned

That's it - no schema modifications needed.

When to Use Each Approach

The flexible schema approach works great for:

Rapid prototyping and experimentation
Simple data extraction tasks
Situations where you frequently change data requirements

Stick with dedicated schemas for:

Critical production data extraction tasks
Complex nested data structures
When you need the highest possible accuracy

The flexible approach gives you about 90% of the quality with 10% of the setup work, making it perfect for most use cases where you need structured data extraction without the schema management overhead.

🔁 Related lessons

Next: The text editor tool
Previous: Structured data with tools
Same section: Overview of Claude Models · Accessing the API · Making a request
Part of paths: Path C
Reference docs: Glossary · Skills atlas · By use-case

📚 Source & attribution

Original Anthropic Academy lesson: https://anthropic.skilljar.com/claude-in-amazon-bedrock/276763