RFC: optional actionable-snapshot ExtractionStrategy — elements + selectors + state, ~68% lighter than raw_markdown for action-selection

Ouverte
#2,463 1 commentaire 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

Personne n'a encore pris cette issue.

Évaluation

Difficulté
4/5
Temps estimé
3-5 jours
Accessibilité débutants
35/100
Type d'issue
Fonctionnalité
Clarté
Clairement spécifiée
Activité
Active
Stack technique
python
Domaine
ai, backend

Piste de recherche

Read Crawl4AI's ExtractionStrategy implementation and the existing optional LLM strategies to understand how strategies are integrated and dependencies handled. The proposed adapter uses ZeroDOM's parse_html, to_compact_text, and ref_map APIs; the issue says it has been verified against zerodom 0.1.4. If maintainers invite a PR, add the strategy module, a test, and a short docs example; done means it runs through CrawlerRunConfig with HTML input and stays opt-in.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

✨ Enhancement
Problem

Agents built on top of Crawl4AI that need to act on a page (click/fill) mostly need
one thing from a crawl: the actionable elements and a way to target each. Today the
lightest thing available is cleaned content / markdown — great for reading and
extraction, but it carries the page's prose and doesn't hand back stable selectors to act
with. There's no actionable, selector-backed representation to reach for.

Proposed solution

An opt-in ExtractionStrategy (ZeroDOMExtractionStrategy) that returns a flat,
uniquely-addressable actionable snapshot as extracted_content:

[01] Log in
[02] input Password filled
[03] button Sign in

with the CSS selectors kept out of the model context in a ref_map
({"01": "body > a", ...}) so a downstream agent can act on what it read. State markers
are inline (x checked, empty, filled, active tab…), and sensitive values are
redacted
(a filled password/card/CVV/SSN shows as filled, never the value).

It's thin glue over ZeroDOM (zerodom>=0.1.4
on PyPI), already verified to conform and run through Crawl4AI's own run() path with
input_format="html":

from crawl4ai.extraction_strategy import ExtractionStrategy

class ZeroDOMExtractionStrategy(ExtractionStrategy):
    def __init__(self, viewport_only=False, with_selectors=True, **kwargs):
        kwargs["input_format"] = "html"
        super().__init__(**kwargs)
        self.viewport_only = viewport_only; self.with_selectors = with_selectors
        import zerodom  # optional dep; imported only when this strategy is used
    def extract(self, url, html, *q, **kwargs):
        from zerodom import parse_html
        g = parse_html(html or "", url=url, viewport_only=self.viewport_only)
        b = {"index": 0, "tags": ["zerodom.snapshot"], "content": g.to_compact_text(terse=True)}
        if self.with_selectors: b["ref_map"] = g.ref_map()
        return [b]
CrawlerRunConfig(extraction_strategy=ZeroDOMExtractionStrategy())

Fully opt-in: no change to markdown/scraping/other extraction, and no new hard
dependency
— zerodom is imported only if you instantiate the strategy (same pattern
as the LLM strategies' optional deps).

Measured — against Crawl4AI's own raw_markdown

Same pages, run through Crawl4AI via raw:// (its real pipeline), tokens via tiktoken
cl100k_base. Snapshot vs the page's own raw_markdown:

page raw_markdown snapshot fewer
en.wikipedia.org (Web scraping) 16,717 5,178 69%
docs.python.org/3 13,654 1,016 93%
flask.palletsprojects.com 16,289 2,462 85%
developer.mozilla.org 3,977 1,649 59%
go.dev 3,975 1,440 64%
news.ycombinator.com 4,304 2,652 38%
www.python.org 4,125 1,300 68%
gov.uk 2,506 693 72%
www.rust-lang.org 1,314 262 80%
httpbin.org/forms/post 51 103 −102%

Median ~68% fewer tokens across 15 pages. Honest edges: on a trivial form-only page
(httpbin.org/forms/post) the snapshot is larger — there's almost no prose to drop and
the element list has its own floor. And this is deliberately not like-for-like:
markdown is prose for reading/extraction; this is actionable elements for acting. It's
leaner because it drops prose — but it's also the only representation here that carries
selectors + element state, which markdown doesn't. So the win is "a purpose-built, lighter
action representation Crawl4AI doesn't have today," not "a better markdown."

For reference vs other tools (not Crawl4AI): ~1/5 the tokens of Stagehand's
observe() a11y tree and ~49% leaner than agent-browser snapshot -i at matching
coverage — methodology/raw runs in the ZeroDOM repo (benchmarks/). Parse is
deterministic, no LLM, network-free.

Honest scope / caveats
  • For the single-shot action-perception use case — not a markdown or data-extraction
    replacement (it drops prose by design).
  • Crawl4AI hands the strategy already-rendered HTML, so state comes from the static
    DOM that HTML exposes (aria/native attributes, serialized values). Live-only DOM
    properties set purely in JS after render aren't visible to an HTML-only parse.
Alternatives considered
  • Trimming markdown/content in place — smaller win, and it couples the action and
    extraction paths that want different things.
  • Keeping it fully external — works today, but a first-class opt-in strategy is where
    it actually saves users tokens on the action path without plumbing.
Are you open to an invited PR?

The adapter is written and verified to conform to ExtractionStrategy and run through
Crawl4AI's pipeline against the published zerodom 0.1.4. Happy to open a PR with the
strategy module + a test + a short docs example on your word — wanted to check interest
and get a pointer to where you'd like it to live first.

Langage dominant
Python
Étoiles
87.2k
Forks
9.1k
Merge moyen
1 j 6 h
PR mergées (30 j)
25

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de unclecode/crawl4ai

Toutes les issues de unclecode/crawl4ai

Issues similaires

Plus d'issues Python

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.