Pulling fields out of text

Requests arrive as prose: “Hello, my name is Igor, we need a wholesale batch, please call me.” To put that into a table, send it to a CRM or count it, you need fields, not a paragraph. This scenario turns one into the other.

What is inside

1. Request text — the entry: a chat message, the body of an email, a form comment.
2. Break into fields — the structured_output_llm node. It differs from an ordinary agent in one way: the model must return exactly the fields its schema describes and nothing else. Not a description of the request, but name, contact, topic.
3. Request fields — the exit. From here they can be stored, compared or routed.

What to set up for yourself

• The schema is the whole configuration. Describe the fields you need: amount, deadline, city, contract number. Write the field descriptions in your own language — they are what tells the model what you mean.
• Write the schema title in Latin characters. It travels to the model provider as the name of the response format, and the provider accepts only Latin letters, digits, _ and -. A title in Cyrillic is rejected with an error that names none of your fields.
• Required fields. Mark as required only what is always present. Not everyone leaves a phone number, and demanding one makes the model invent it.

What to connect

Nothing. From here the fields usually go somewhere: a table, a CRM, an email to a manager. “A form request, parsed and filed” shows the whole path: the same extraction carried through to a stored row and a reply to the sender.