[Bug]: PruningContentFilter cuts inline spans out of sentences it keeps (e.g. "renamed Tokyo (lit. '')")

Aperta
#2,417 2 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

@siva23367 ci sta già lavorando.

Dal 19/10/2026.

Valutazione

Difficoltà
3/5
Tempo stimato
1-2 giorni
Idoneità per principianti
75/100
Tipo di issue
Bug
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
python
Ambito
backend, data

Direzione di ricerca

Start by reading PruningContentFilter and _prune_tree in crawl4ai/content_filter_strategy.py, then run the provided reproducer with nested spans inside kept paragraphs. Check how the default filter handles those spans compared with preserve_tags=["span"]. Done means the default output retains the text inside nested inline spans without dropping otherwise kept paragraph content.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

⚙ Done 🐞 Bug 📌 Root caused
crawl4ai version

0.9.4 (reproduced). PruningContentFilter (crawl4ai/content_filter_strategy.py) is unchanged in v0.9.5 and on main (checked 2026-10-19).

Expected Behavior

When PruningContentFilter keeps a paragraph, the words inside that paragraph stay. A filter that removes a whole block is expected; a kept sentence with words missing from its middle is not.

Current Behavior

A <span> whose content is another element, such as a span around a link or a span around a span, is removed from a paragraph that is otherwise kept. Wikipedia marks up glosses and native-script names this way, so on https://en.wikipedia.org/wiki/Tokyo fit_markdown reads:

the city was renamed Tokyo (lit. '')

The source HTML is (<abbr>lit.</abbr> ... <span class="gloss-text"><a href="...">Eastern Capital</a></span> ...).

In the snippet below, plain spans, <strong>, <em> and links placed directly in the paragraph are kept; only the nested spans are lost.

My reading of the cause: _prune_tree scores every descendant element as if it were a block, and the score includes text density (text length over markup length) and link density. A short span whose only content is another element has little text against its markup, because the inner element's tags count as markup too. When the inner element is a link, the link density is also 1. Either way the span scores under the threshold and is removed.

Is this reproducible?

Yes

Inputs Causing the Bug

HTML with short nested inline spans inside a paragraph, for example https://en.wikipedia.org/wiki/Tokyo.

Steps to Reproduce
  1. Run the snippet below.
  2. Compare the default output with the paragraphs in HTML.
Code snippets
import asyncio
from crawl4ai import AsyncWebCrawler, CrawlerRunConfig, CacheMode, DefaultMarkdownGenerator, PruningContentFilter

HTML = """<html><body><article>
<p>After the <a href="/wiki/Meiji_Restoration">Meiji Restoration</a> (1868), the imperial capital in <a href="/wiki/Kyoto">Kyoto</a> was moved to Edo, and the city was renamed Tokyo (<abbr title="literal translation">lit.</abbr> <span class="gloss-quot">'</span><span class="gloss-text"><a href="/wiki/Capital_of_Japan">Eastern Capital</a></span><span class="gloss-quot">'</span>). The city was greatly damaged by the 1923 earthquake and rebuilt in the following decades.</p>
<p>Tokyo (<span title="Japanese-language text"><span lang="ja">東京</span></span>) is the capital of Japan and one of the most populous cities in the world, with a population of about fourteen million.</p>
<p>The museum near the harbour holds <span>old maps</span> of the bay, <strong>ship models</strong> and <em>letters</em> written by the first merchants who settled on the northern quay.</p>
</article></body></html>"""

async def main():
    async with AsyncWebCrawler() as crawler:
        for name, flt in [("default", PruningContentFilter()), ("preserve_tags=['span']", PruningContentFilter(preserve_tags=["span"]))]:
            config = CrawlerRunConfig(cache_mode=CacheMode.BYPASS, markdown_generator=DefaultMarkdownGenerator(content_filter=flt))
            result = await crawler.arun("raw:" + HTML, config=config)
            print(f"--- {name}")
            print(result.markdown.fit_markdown)

asyncio.run(main())

Output on 0.9.4:

--- default
After the [Meiji Restoration](/wiki/Meiji_Restoration) (1868), the imperial capital in [Kyoto](/wiki/Kyoto) was moved to Edo, and the city was renamed Tokyo (lit. ''). The city was greatly damaged by the 1923 earthquake and rebuilt in the following decades.
Tokyo () is the capital of Japan and one of the most populous cities in the world, with a population of about fourteen million.
The museum near the harbour holds old maps of the bay, **ship models** and _letters_ written by the first merchants who settled on the northern quay.
  *[lit.]: literal translation

--- preserve_tags=['span']
After the [Meiji Restoration](/wiki/Meiji_Restoration) (1868), the imperial capital in [Kyoto](/wiki/Kyoto) was moved to Edo, and the city was renamed Tokyo (lit. '[Eastern Capital](/wiki/Capital_of_Japan)'). The city was greatly damaged by the 1923 earthquake and rebuilt in the following decades.
Tokyo (東京) is the capital of Japan and one of the most populous cities in the world, with a population of about fourteen million.
The museum near the harbour holds old maps of the bay, **ship models** and _letters_ written by the first merchants who settled on the northern quay.
  *[lit.]: literal translation
Workaround and suggestion

preserve_tags=["span"] fixes it: a preserved element is kept whole and not scored. On the Tokyo page it grows fit_markdown by about 2% (139,263 to 141,985 characters), so it does not let noticeably more noise through. Lowering the threshold, the workaround discussed in #582, applies to the whole page instead.

Since a filter that keeps a block should keep the block's words, a possible fix is to stop scoring inline (phrasing) elements on their own and let them follow their nearest block ancestor. Making span part of the default preserve_tags would cover the cases above with a smaller change.

#582 reported a similar symptom for <a> and <strong> and was closed after tuning the threshold. This report is narrower: the kept sentence itself is left with holes.

OS

Debian GNU/Linux 12 (bookworm), in the official unclecode/crawl4ai:0.9.4 Docker image

Python version

3.12.14

Error logs & Screenshots (if applicable)

None.

Lingua principale
Python
Stelle
87.2k
Fork
9.1k
Merge medio
1g 6h
PR unite (30g)
25

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di unclecode/crawl4ai

Tutte le issue di unclecode/crawl4ai

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.