Trafilatura and Its Forks and Ports

The Trafilatura family spans a Python toolkit, Go and Rust implementations, and an offline TypeScript engine. Compare their inputs, outputs, and runtime boundaries before selecting one. A port reimplements behavior in another language; a fork continues and changes an existing codebase.

Trafilaturacore is a direct port of the Python original rather than a fork of another port. It is also a bounded offline engine, while Markdownee adds the browser, crawler, and format-conversion layers around it.

The family at a glance

Trafilaturacore is ported from Python Trafilatura v2.2.0; go-trafilatura serves only as a DOM translation aid. Markdownee sits beyond the offline engine boundary.¹

Status and versions below were checked on July 20, 2026; the Trafilatura and go-trafilatura rows on August 31, 2026.

Trafilatura: the Python original

Adrien Barbaresi created the canonical Trafilatura project during PhD work in linguistics and natural language processing, initially to build research text databases at the Berlin-Brandenburg Academy of Sciences. The 2021 ACL system-demonstration paper describes the same corpus-building context.¹³¹⁴

The current project is broader than an extraction function. It discovers and crawls URLs, downloads pages, accepts saved HTML, extracts main text, comments, and metadata, preserves document structure, and serializes several text and markup formats.² Version 2.2.0 was released on July 31, 2026. Current packaging labels the project Production/Stable and requires Python 3.10 or newer.⁴

Trafilatura's present license is Apache-2.0. Releases before v1.8.0 used GPLv3+; that historical boundary matters when auditing copied or vendored older code.³

go-trafilatura: compatibility with substitutions

go-trafilatura ports Trafilatura into Go and deliberately mirrors much of the Python project's structure. Its goal is compatible extraction, not byte-for-byte identity. The documented departures include Go implementations of Readability and DOM Distiller for fallbacks, JSON decoding for JSON-LD instead of the Python regex path, support for custom fallback candidates, and HTML-oriented formatting.⁵

The package extracts supplied HTML. Its CLI adds downloading plus file, URL-list, feed, and sitemap modes; HTML is the default output, with text and JSON options.⁶ The v2.0.0 release tracks Python Trafilatura 2.0.0, so it does not yet claim parity with the current Python 2.2.0 release. That tag is also not installable: its go.mod keeps the unsuffixed module path, which Go's semantic import versioning rejects, so go get resolves to v1.12.2.⁷

rs-trafilatura: page-aware Rust extraction

rs-trafilatura describes itself as a Rust port of Trafilatura and go-trafilatura; its main extraction logic identifies go-trafilatura as the immediate porting source.⁸ It adds page-aware profiles for article, forum, product, collection, listing, documentation, and service pages. The active classifier combines URL heuristics, HTML signals, and machine learning, returning a page type and classification confidence.⁹

That ML role needs a narrow description. In the reviewed 0.2.2 source, the live path calculates extraction quality with a heuristic function. A separate ML-quality function remains defined but has no call site, so an extraction-quality score should not be described as a current ML prediction.¹⁵

The primary APIs accept HTML strings or bytes rather than fetching pages. An optional feature accepts already-fetched Spider pages; it is an integration point, not a built-in crawler. Results center on plain text and can also carry HTML, GitHub-flavored Markdown, comments, images, metadata, and classification details.¹⁰ The crate registry reports 0.2.2, while GitHub's newest tagged release is still v0.1.1; both facts are needed to describe its current release state accurately.¹¹

Another Rust port

The named projects are not an exhaustive list. nchapman/trafilatura-rs is a separate Apache-2.0 Rust port that follows go-trafilatura, offers library extraction and an optional network-capable CLI, and publishes bindings for several languages. Trafilaturacore's attribution record says this project was consulted but none of its code ships in Trafilaturacore.¹⁶¹

Trafilaturacore: a bounded Node engine

Trafilaturacore is an open-source pure-TypeScript port that runs anywhere Node runs, with no native module to install. Its lineage is single-sourced: the extraction core and the metadata reader are ported from Python Trafilatura v2.2.0, and go-trafilatura is consulted only as a DOM translation aid where Python's lxml idioms need a DOM equivalent.¹

Trafilaturacore adds an HTML output stage. It ships no machine-learning model, no page-type taxonomy, and no classification output; the only page type it reports is the one a page declares in its own OpenGraph or JSON-LD metadata.

This alpha engine accepts HTML plus optional URL context and returns one cleaned-HTML string with metadata. It does not fetch or render pages, and it does not itself convert content to Markdown, plain text, or JSON. A JSON CLI envelope does not change that content-output boundary.¹²

Where Markdownee begins

Markdownee depends on Trafilaturacore, not directly on rs-trafilatura. Its Crawlee and Playwright layer fetches and, when necessary, renders pages. The extraction package sends that HTML through Trafilaturacore once; a separate conversion package turns the cleaned HTML into txt, markdown, readable html, or minified-html.¹⁷

This separates code lineage from runtime dependency: Trafilaturacore ports the Python original, and Markdownee consumes Trafilaturacore as its engine.

One benchmark, one dataset

The archived ScrapingHub article-extraction benchmark is the narrow comparison retained here. This vendor-authored test uses 181 stored HTML pages, article-body ground truth, and a four-token-shingle evaluator. Precision and recall are normalized per page before averaging; ± is a bootstrap-estimated standard deviation from resampled pages, not a confidence interval.¹⁸

Results below are third-party implementation results retrieved July 20, 2026. They measure only the listed revisions on this article set.¹⁸

The rs-trafilatura README separately reports F1 0.966 from its project run. The table uses the benchmark archive's committed 0.970 row so all three values share one provenance and method.¹⁹ These small differences do not establish general superiority: the set covers article-body extraction, not metadata, crawling, other page types, or current releases.

Trafilatura's ACL and documentation evaluations use different document sets and segment-based labeling; WCXB spans seven page types and uses word-overlap scoring. Figures from those datasets cannot be ranked against the ScrapingHub table.²⁰²¹ Trafilaturacore is not a row in this benchmark, and no Python, Go, or rs-trafilatura result transfers to Trafilaturacore or Markdownee.

Choosing by boundary

• Python Trafilatura provides discovery, downloading, extraction, metadata, and several serialized formats in one toolkit.
• go-trafilatura provides a Go package and network-capable CLI, with the fallback and formatting differences documented above.
• rs-trafilatura provides native Rust extraction with page-type classification and optional Spider integration.
• Trafilaturacore serves Node applications that supply HTML and need cleaned HTML from an alpha engine.
• Markdownee adds fetching, browser rendering, crawling, storage, and format conversion around that engine.

Updated: August 31, 2026