A Python library for intelligently chunking HTML documents into structured, size-limited segments based on DOM tree analysis.
This library processes HTML content to split it into semantically coherent chunks while respecting specified size constraints. It analyzes the DOM structure to identify optimal split points, preserving contextual information and document hierarchy.
DeepWiki Docs: https://deepwiki.com/carlosplanchon/betterhtmlchunking
- Custom DOM tree representation.
- Configurable chunk size limits (counting by text or HTML length).
- Intelligent region-of-interest detection.
- Dual output formats: HTML and plain text chunks.
- Preservation of structure relationships.
- Customizable tag filtering.
pip install betterhtmlchunking- Python 3.11+
- attrs
- treelib
- beautifulsoup4
- parsel-text
- lxml
- attrs-strict
from betterhtmlchunking import DomRepresentation
from betterhtmlchunking.main import ReprLengthComparisionBy
from betterhtmlchunking.main import tag_list_to_filter_out
html_content = """
<html>
<body>
<div id="content">
<h1>Document Title</h1>
<p>First paragraph...</p>
<p>Second paragraph...</p>
</div>
</body>
</html>
"""
# Create document representation with 20 character chunks.
dom_repr = DomRepresentation(
MAX_NODE_REPR_LENGTH=20,
website_code=html_content,
repr_length_compared_by=ReprLengthComparisionBy.HTML_LENGTH
# tag_list_to_filter_out=["/head", "/header", "..."] # By default tag_list_to_filter_out is used.
)
dom_repr.start()
# Render HTML:
for idx in dom_repr.tree_regions_system.sorted_roi_by_pos_xpath:
print("*" * 50)
print(f"IDX: {idx}")
roi_html_render: str =\
dom_repr.render_system.get_roi_html_render_with_pos_xpath(
roi_idx=idx
)
print(roi_html_render)
# Render text:
for idx in dom_repr.tree_regions_system.sorted_roi_by_pos_xpath:
print("*" * 50)
print(f"IDX: {idx}")
roi_text_render: str =\
dom_repr.render_system.get_roi_text_render_with_pos_xpath(
roi_idx=idx
)
print(roi_text_render)Render output (HTML):
**************************************************
IDX: 0
<h1>
Document Title
</h1>
**************************************************
IDX: 1
<p>
First paragraph...
</p>
**************************************************
IDX: 2
<p>
Second paragraph...
</p>
Render output (text):
**************************************************
IDX: 0
Document Title
**************************************************
IDX: 1
First paragraph...
**************************************************
IDX: 2
Second paragraph...
MAX_NODE_REPR_LENGTH: Maximum allowed length for each chunk (in characters)repr_length_compared_by: Length calculation method:- ReprLengthComparisionBy.HTML_LENGTH: HTML source length
- ReprLengthComparisionBy.TEXT_LENGTH: Rendered text length
website_code: Input HTML content
# Access the DOM tree structure
tree = dom_repr.tree_representation.tree
# Get node metadata:
for node in tree.all_nodes():
if node.data is not None:
print(f"XPath: {node.identifier}")
print(f"Text length: {node.data.text_length}")
print(f"HTML length: {node.data.html_length}")-
DOM Parsing
- Builds a tree representation of the HTML document.
- Calculates metadata (text length, HTML length) for each node.
-
Region Detection
- Uses Breadth First Search (BFS) to traverse the DOM tree in a level-order fashion, ensuring that each node is processed systematically.
- Combines nodes until the specified size limit is reached.
- Preserves parent-child relationships to maintain contextual integrity.
-
Chunk Generation
- Creates HTML chunks with original markup.
- Generates parallel text-only chunks.
- Maintains chunk order based on document structure.
The actual practice (Feb. 2025) is to use plain-text or token-based chunking strategies, primarily aimed at keeping prompts within certain token limits for large language models. This approach is ideal for quick semantic retrieval or QA tasks on unstructured text.
By contrast, betterhtmlchunking preserves the HTML DOM structure, calculating chunk boundaries based on each node’s text or HTML length. This approach is especially useful when you want to:
- Retain or leverage the hierarchical relationships in the HTML (e.g., headings, nested divs)
- Filter out undesired tags or sections (like
<script>or<style>) - Pinpoint exactly where each chunk originated in the document (via positional XPaths)
You can even combine the two techniques if you need both structured extraction (via betterhtmlchunking) and LLM-friendly text chunking (via LangChain) for advanced tasks such as summarization, semantic search, or large-scale QA pipelines.
The package ships with a command line interface built with Typer. You can pipe HTML to the tool and work with chunks in multiple ways.
Get a specific chunk as HTML:
cat input.html | betterhtmlchunking --max-length 32768 --chunk-index 0 > chunk.htmlBy default the command reads from stdin, processes chunks up to a maximum length of 32,768 characters, and prints the HTML corresponding to chunk index 0 to stdout.
See how many chunks are created and their sizes without processing the full content:
cat input.html | betterhtmlchunking --max-length 32768 --list-chunksOutput:
Total chunks: 5
Chunk 0: 28543 chars HTML, 12834 chars text
Chunk 1: 31245 chars HTML, 15234 chars text
Chunk 2: 19823 chars HTML, 9234 chars text
...
Extract all chunks to separate files:
cat input.html | betterhtmlchunking --max-length 32768 --all-chunks --output-dir ./chunksThis creates chunk_0.html, chunk_1.html, etc. in the specified directory.
Get plain text content without HTML markup:
# Single chunk as text
cat input.html | betterhtmlchunking --max-length 32768 --chunk-index 0 --text-only > chunk.txt
# All chunks as text files
cat input.html | betterhtmlchunking --max-length 32768 --all-chunks --text-only --output-dir ./chunksThis creates chunk_0.txt, chunk_1.txt, etc. with plain text content.
Get structured JSON output with all chunks and metadata:
cat input.html | betterhtmlchunking --max-length 32768 --format json > output.jsonOutput format:
{
"total_chunks": 5,
"max_length": 32768,
"compared_by": "html",
"chunks": [
{
"index": 0,
"html": "<h1>...</h1>",
"text": "...",
"html_length": 28543,
"text_length": 12834
},
...
]
}This is useful for programmatic processing. You can pipe the output to tools like jq or process it with scripts:
# Extract only text from all chunks
cat input.html | betterhtmlchunking --format json | jq -r '.chunks[].text'
# Get chunk count
cat input.html | betterhtmlchunking --format json | jq '.total_chunks'
# Filter chunks by size
cat input.html | betterhtmlchunking --format json | jq '.chunks[] | select(.text_length > 1000)'Enable progress logging with --verbose. Logs are written to stderr so they don't interfere with chunk output:
cat input.html | betterhtmlchunking --max-length 32768 --all-chunks --output-dir ./chunks --verboseFor detailed inspection of the DOM, nodes, ROIs, and chunk lengths, use --maximal-verbose. This logs:
- Total DOM nodes and their HTML/text lengths
- Each ROI (region of interest) with constituent node XPaths and lengths
- Final chunk HTML and text sizes
cat input.html | betterhtmlchunking --max-length 32768 --chunk-index 0 --maximal-verbose > chunk.html 2> logs.txtchunk.htmlcontains the selected chunk HTML.logs.txtcaptures all detailed logging information.
This mode is useful for debugging, testing, or analyzing how the document is split into chunks.
| Option | Short | Description |
|---|---|---|
--max-length |
-l |
Maximum length for each chunk (default: 32768) |
--chunk-index |
-c |
Index of specific chunk to output (default: 0) |
--list-chunks |
List all chunks with their sizes | |
--all-chunks |
Output all chunks (requires --output-dir) |
|
--output-dir |
-o |
Directory to save chunks when using --all-chunks |
--text-only |
Output plain text instead of HTML | |
--format |
-f |
Output format: json for structured JSON output |
--text |
Compare length using text instead of HTML | |
--verbose |
-v |
Enable progress logging |
--maximal-verbose |
Enable detailed debug logging |
MIT License
Feel free to open issues or submit pull requests if you have suggestions or improvements.