novada-extract

novada_extract — Extraction Specialist Skill

When to use: You have a URL and need its content. This is the primary content extraction tool.

Decision: Which render mode?

Site type	Recommended	Reason
News, docs, blogs	render=auto (default)	Static fetch works, 112ms avg
LinkedIn, Glassdoor, React SPA	render=render	JS-rendered, needs Web Unblocker
Behind login, interactive	render=browser	Full Chromium CDP

Rule: Never use render=render for ALL pages. auto is 15-113x faster.

Batch extraction pattern

{
  "url": ["https://site.com/page1", "https://site.com/page2", "https://site.com/page3"],
  "render": "auto",
  "max_chars": 25000
}

Returns one labeled section per URL: ### [1/3] url.

Field extraction pattern

{
  "url": "https://amazon.com/product/...",
  "fields": ["price", "rating", "availability", "brand"]
}

Checks JSON-LD first, then pattern matching. Returns ## Requested Fields block.

Content too long?

Use max_chars=50000 for long docs. Default is 25000. Max is 100000.

{
  "url": "https://long-doc.com/guide",
  "max_chars": 50000
}

Bot-blocked / empty content?

Escalate: static → render → browser.

{"url": "...", "render": "render"}

If still blocked:

{"url": "...", "render": "browser"}

Common Mistakes

• ❌ Setting render=render on all calls — adds 9-16s latency on pages that don't need it
• ❌ Calling novada_extract to check if a URL exists — use novada_map for URL discovery
• ❌ Setting max_chars=100000 by default — use 25000, increase only when needed

After extraction

• Content looks JS-heavy or incomplete → retry with render=render
• Content is truncated → increase max_chars
• Need more URLs → call novada_map first, then batch extract

When to call extract after search

novada_search returns snippets only — they locate pages but do not answer questions. If accuracy matters (facts, prices, current events), call novada_extract on the top 1-3 search results and read the full content before answering. For multi-source questions, novada_research (which calls extract internally on top sources) is cheaper than doing this manually — prefer it when you need ≥2 sources.