AI & Agent Dev Bug Sandbox Back to Radar

HTMLSemanticPreservingSplitter Drops Video/Audio URLs From Nested Source Tags

The splitter fails to extract media URLs from nested <source> elements inside <video> and <audio> tags, resulting in empty Markdown links when preserve_videos or preserve_audio is enabled. This causes data loss for common HTML media patterns.

mediumConfidence 85%Langchain

Origin Analysis

The media preservation logic only reads the direct `src` attribute of the <video>/<audio> tag and does not fall back to the first child <source> element. According to HTML spec, media sources can be provided either via a direct src attribute or via nested source tags, but the implementation ignores the latter.

Run the provided Python code with HTML containing <video><source src='...'></video> and <audio><source src='...'></audio>, set preserve_videos=True and preserve_audio=True, and observe empty Markdown links like ![video:]() instead of URLs.

Fixing Code Block

--- a/libs/text-splitters/langchain_text_splitters/html.py
+++ b/libs/text-splitters/langchain_text_splitters/html.py
@@ -100,10 +100,14 @@
         if element.name in ["video", "audio"]:
             if (element.name == "video" and self.preserve_videos) or (element.name == "audio" and self.preserve_audio):
                 src = element.get("src")
+                if not src:
+                    source_element = element.find("source")
+                    if source_element:
+                        src = source_element.get("src")
                 if src:
                     media_text = f"![{element.name}:{src}]({src})"
                 else:
                     media_text = f"![{element.name}:]()"

Add a fallback to extract the src from the first nested <source> element when the direct src attribute is absent. This aligns with common HTML authoring patterns and ensures the URL is preserved.

Edge Case Audit

This patch picks the first <source> element regardless of its type or order. If multiple <source> tags exist (e.g., different resolutions), it may select a suboptimal URL. In concurrent environments, do not monkey-patch globally; apply the change directly to the library source or use a subclass. Rollback: revert to the original single-line src extraction if unexpected media links appear. The patch does not handle srcset or data-src attributes.