Use a browser workflow when the information depends on actions or session state. Use a single-page extractor when an approved URL already contains the required content. Rendering JavaScript can reveal a page’s content, but it does not by itself implement a login, form submission or multi-step task.
Identify the missing capability
| Requirement | Operation to verify |
|---|---|
| Text already present at a URL | Single-page extraction |
| Text appears after client rendering | Rendering plus an explicit content-ready check |
| Text appears only after a click or input | Browser action and resulting state |
| Authenticated account workflow | Authorized session, actions and output checks |
Keep these decisions separate. Reader mode 1 requests browser rendering for a URL. It should not be described as an arbitrary browser-action API. A proxy tier is an access option, not proof that an application can solve any CAPTCHA or use an account without permission.
An owned interaction fixture
The following local page has a report that is initially hidden. Clicking its button changes the visible state. The test creates the HTML itself and contacts no target site. Install Playwright and its Chromium runtime; this version uses an installed Chrome channel. The Playwright Page API documents the relevant navigation, locator and action operations.
from playwright.sync_api import sync_playwright
DEMO = """<!doctype html><title>Owned interaction fixture</title>
<button id="show" onclick="document.querySelector('#panel').hidden=false">Show report</button>
<section id="panel" hidden><h2>Fixture report</h2><p>Approved sample, not production data.</p></section>"""
def inspect_interaction():
with sync_playwright() as p:
browser = p.chromium.launch(channel="chrome", headless=True)
try:
page = browser.new_page()
page.set_content(DEMO)
before = page.locator("#panel").is_visible()
page.get_by_role("button", name="Show report").click()
page.locator("#panel").wait_for(state="visible")
return {"before_visible":before, "after_visible":True,
"text":page.locator("#panel").inner_text()}
finally:
browser.close()
What this fixture demonstrates
The expected trace is before_visible: false, after_visible: true, followed by the fixture report text. The hidden text is also present in the initial HTML, so a raw HTML or Markdown extraction may contain it even before the action. This is precisely why finding the text does not prove that the visible workflow completed.
| Evidence | What it establishes |
|---|---|
| Initial DOM and screenshot | Starting visible state |
| Action log | Which user action was attempted |
| Visible locator after the action | The expected state was reached |
| Final extracted text | The content available for downstream use |
For a real application, add failure cases: the button is absent, the action is rejected, or the expected panel never becomes visible. Check a specific ready condition instead of sleeping for an arbitrary interval. Treat cancellation and lost session state as task failures; do not convert them into an empty but successful answer.
Choose the smaller implementation when actions are unnecessary
If the page is a public article at a known URL, a browser session may add no useful capability. Evaluate the single-URL extraction contract first. the dynamic extraction workflow covers rendering and extraction failure, while the live RAG ingestion architecture covers using the resulting sources. This page owns the decision to perform browser actions and the evidence needed to accept them.