Structured Extraction

Turn read() output into validated JSON records (schema-first) so agents can extract data reliably instead of scraping brittle HTML.

This page covers:

• A schema-first extraction pattern (validated outputs)
• How to use extraction on top of read() markdown
• How to handle failures deterministically

Table of Contents

1. Concept
2. Python: typed extraction
3. Failure modes

---

Concept

The stable path to structured data is:

1. Read the page into markdown/text (read())
2. Extract a typed object from that content (extract(...))
3. Validate the object with a schema so callers don’t need prompt heuristics

---

Python: typed extraction

from pydantic import BaseModel
from predicate import read, extract

class Item(BaseModel):
    name: str
    price: str

md = read(browser, format="markdown")["content"]

result = extract(browser, llm, "Extract item name and price", schema=Item)
if result.ok:
    print(result.data.name, result.data.price)
else:
    print("extract failed:", result.error)

---

Failure modes

Extraction can fail for deterministic reasons:

• invalid JSON
• schema mismatch (missing/extra fields)
• not enough information in read() output

When extraction fails, treat it as a normal verification failure:

• retry with a clarified prompt
• narrow the read scope
• fall back to a snapshot-based workflow if the page is too dynamic

---

Related topics

• Content Reading — produce markdown/text inputs
• Tool Registry — typed tool contracts (related concept)