Firecrawl Community Reports Image-Cleanup Bug That May Delete Text and Misalign Table Columns

A community report has flagged a potential data-correctness issue in Firecrawl’s web data-cleaning process. An issue filed on October 8 says that when the same line of Markdown contains a non-Base64 data:image image followed by a Base64 image, the default cleanup may also delete the text between them. The report provides HTML and API requests for a self-hosted environment, allowing the issue to be checked with a small test case. Issue report

The official API documentation confirms that removeBase64Images defaults to true. It is intended to prevent lengthy embedded images from taking up space in Markdown; the expected behavior is to preserve alt text and replace only the image URL with a placeholder, without affecting html or rawHtml. This gives the bug practical significance: the cleanup may run even when users have not configured image filtering themselves. API documentation

The report attributes the cause to the regular expression’s matching boundary: the URL portion can cross the closing parenthesis of the first image and continue searching for a later Base64 marker. In the supplied comparison table, a data row under a three-column header consequently loses a cell: the negating text disappears and the following value shifts left. The report also includes an example of text being swallowed in a regular paragraph. These results are reproductions supplied by the reporter and have not been independently verified in this article. Reproduction and cause analysis

As an engineering inference, if this kind of Markdown is sent to a RAG index or an agent research workflow, misaligned fields could lead a model to attribute one option’s capabilities to another. The data may remain readable and embeddings may still be generated, so checking only whether a request succeeded is not enough to verify content integrity. Firecrawl also provides HTML and raw HTML output, which teams can use to compare text and table structure on representative pages. Output format documentation

The reporter says disabling removeBase64Images preserves the cells in the test case and recommends constraining the regular expression so it cannot cross a closing parenthesis. Teams using this workaround still need to manage the input length caused by embedded images. Follow-up should track an upstream fix, regression tests, and its inclusion in a release. Current evidence is limited to a self-hosted environment on the main branch and is insufficient to determine whether all released versions or cloud deployments are affected. Workaround and suggested fix