mirror of
https://github.com/0xMassi/webclaw.git
synced 2026-06-10 22:45:13 +02:00
df8bdc96db
This PR fixes three independent issues that surface when running
`webclaw --format llm` against modern news index pages. They were
all reproducible against bbc.com/news/world and reuters.com/world/middle-east
during a real briefing-generation run.
### 1. Framework hydration blobs no longer dump into the output
`to_llm_text` was unconditionally appending every parsed structured-data
item as a `## Structured Data` JSON fence. On Next.js sites, that means
the entire `__NEXT_DATA__` `pageProps` object — ad-targeting flags,
build IDs, schedule paths, feature toggles — gets serialized straight
into the LLM context. On bbc.com/news/world it was about 140 KB of
pure framework noise drowning the actual page content.
The fix layers three filters:
- Items with a Schema.org `@type` of `WebSite`, `WebPage`, or
`SiteNavigationElement` are dropped as chrome.
- Items without an `@type` (typical of `pageProps` or SvelteKit
data) are kept only if their serialized size stays under 4 KB —
small parsed records with real content survive, hydration blobs
do not.
- The whole section is suppressed if the total serialized size
exceeds 16 KB, regardless of type. Past that threshold it is
almost never useful to a downstream LLM.
JSON-LD records with content-bearing `@type` values (`Article`,
`NewsArticle`, `Product`, `Recipe`, `FAQPage`, `Event`, etc.) are
preserved.
### 2. Element → Text node smashing
`children_to_md` and `inline_text` only ran the `needs_separator`
check on `Element → Element` transitions. When an element rendered
text with no trailing whitespace and was followed by a sibling
text node that started with a non-whitespace character, the two
got concatenated with no separator. The same check now applies to
the `Text` branch in both functions.
### 3. Accessibility link chrome no longer leaks into prose
Sites like Reuters wrap external/new-window links with
screen-reader-only spans (e.g. `, opens new tab`, `external link`).
These have no consistent class hook, so the structural noise filter
cannot reliably catch them and they bleed into the rendered text —
sometimes dozens of times per page.
A targeted regex scrub now runs in two places: in the body cleanup
pipeline (`strip_a11y_link_chrome`, called early after `strip_leaked_js`)
and in the link-label cleaner (`clean_link_label`) so the deduplicated
`## Links` section is also clean.
### Tests
All 286 existing unit tests pass. 8 new tests cover:
- structured-data filter: chrome-type drop, oversized untyped drop,
small untyped keep, `NewsArticle` keep
- markdown separator: `Element → Text → Element` no longer smashes
- a11y stripper: common phrasings, variant phrasings ("opens in a
new window", "external link"), and code-fence preservation
|
||
|---|---|---|
| .. | ||
| webclaw-cli | ||
| webclaw-core | ||
| webclaw-fetch | ||
| webclaw-llm | ||
| webclaw-mcp | ||
| webclaw-pdf | ||
| webclaw-server | ||