langchain/libs/text-splitters/langchain_text_splitters/markdown.py

383 lines
15 KiB
Python
Raw Normal View History

from __future__ import annotations
text-splitters: Introduce Experimental Markdown Syntax Splitter (#22257) #### Description This MR defines a `ExperimentalMarkdownSyntaxTextSplitter` class. The main goal is to replicate the functionality of the original `MarkdownHeaderTextSplitter` which extracts the header stack as metadata but with one critical difference: it keeps the whitespace of the original text intact. This draft reimplements the `MarkdownHeaderTextSplitter` with a very different algorithmic approach. Instead of marking up each line of the text individually and aggregating them back together into chunks, this method builds each chunk sequentially and applies the metadata to each chunk. This makes the implementation simpler. However, since it's designed to keep white space intact its not a full drop in replacement for the original. Since it is a radical implementation change to the original code and I would like to get feedback to see if this is a worthwhile replacement, should be it's own class, or is not a good idea at all. Note: I implemented the `return_each_line` parameter but I don't think it's a necessary feature. I'd prefer to remove it. This implementation also adds the following additional features: - Splits out code blocks and includes the language in the `"Code"` metadata key - Splits text on the horizontal rule `---` as well - The `headers_to_split_on` parameter is now optional - with sensible defaults that can be overridden. #### Issue Keeping the whitespace keeps the paragraphs structure and the formatting of the code blocks intact which allows the caller much more flexibility in how they want to further split the individuals sections of the resulting documents. This addresses the issues brought up by the community in the following issues: - https://github.com/langchain-ai/langchain/issues/20823 - https://github.com/langchain-ai/langchain/issues/19436 - https://github.com/langchain-ai/langchain/issues/22256 #### Dependencies N/A #### Twitter handle @RyanElston --------- Co-authored-by: isaac hershenson <ihershenson@hmc.edu>
2024-06-19 02:44:00 +00:00
import re
from typing import Any, Dict, List, Tuple, TypedDict, Union
from langchain_core.documents import Document
from langchain_text_splitters.base import Language
from langchain_text_splitters.character import RecursiveCharacterTextSplitter
class MarkdownTextSplitter(RecursiveCharacterTextSplitter):
"""Attempts to split the text along Markdown-formatted headings."""
def __init__(self, **kwargs: Any) -> None:
"""Initialize a MarkdownTextSplitter."""
separators = self.get_separators_for_language(Language.MARKDOWN)
super().__init__(separators=separators, **kwargs)
class MarkdownHeaderTextSplitter:
"""Splitting markdown files based on specified headers."""
def __init__(
self,
headers_to_split_on: List[Tuple[str, str]],
return_each_line: bool = False,
strip_headers: bool = True,
):
"""Create a new MarkdownHeaderTextSplitter.
Args:
headers_to_split_on: Headers we want to track
return_each_line: Return each line w/ associated headers
strip_headers: Strip split headers from the content of the chunk
"""
# Output line-by-line or aggregated into chunks w/ common headers
self.return_each_line = return_each_line
# Given the headers we want to split on,
# (e.g., "#, ##, etc") order by length
self.headers_to_split_on = sorted(
headers_to_split_on, key=lambda split: len(split[0]), reverse=True
)
# Strip headers split headers from the content of the chunk
self.strip_headers = strip_headers
def aggregate_lines_to_chunks(self, lines: List[LineType]) -> List[Document]:
"""Combine lines with common metadata into chunks
Args:
lines: Line of text / associated header metadata
"""
aggregated_chunks: List[LineType] = []
for line in lines:
if (
aggregated_chunks
and aggregated_chunks[-1]["metadata"] == line["metadata"]
):
# If the last line in the aggregated list
# has the same metadata as the current line,
# append the current content to the last lines's content
aggregated_chunks[-1]["content"] += " \n" + line["content"]
elif (
aggregated_chunks
and aggregated_chunks[-1]["metadata"] != line["metadata"]
# may be issues if other metadata is present
and len(aggregated_chunks[-1]["metadata"]) < len(line["metadata"])
and aggregated_chunks[-1]["content"].split("\n")[-1][0] == "#"
and not self.strip_headers
):
# If the last line in the aggregated list
# has different metadata as the current line,
# and has shallower header level than the current line,
# and the last line is a header,
# and we are not stripping headers,
# append the current content to the last line's content
aggregated_chunks[-1]["content"] += " \n" + line["content"]
# and update the last line's metadata
aggregated_chunks[-1]["metadata"] = line["metadata"]
else:
# Otherwise, append the current line to the aggregated list
aggregated_chunks.append(line)
return [
Document(page_content=chunk["content"], metadata=chunk["metadata"])
for chunk in aggregated_chunks
]
def split_text(self, text: str) -> List[Document]:
"""Split markdown file
Args:
text: Markdown file"""
# Split the input text by newline character ("\n").
lines = text.split("\n")
# Final output
lines_with_metadata: List[LineType] = []
# Content and metadata of the chunk currently being processed
current_content: List[str] = []
current_metadata: Dict[str, str] = {}
# Keep track of the nested header structure
# header_stack: List[Dict[str, Union[int, str]]] = []
header_stack: List[HeaderType] = []
initial_metadata: Dict[str, str] = {}
in_code_block = False
opening_fence = ""
for line in lines:
stripped_line = line.strip()
# Remove all non-printable characters from the string, keeping only visible
# text.
stripped_line = "".join(filter(str.isprintable, stripped_line))
if not in_code_block:
# Exclude inline code spans
if stripped_line.startswith("```") and stripped_line.count("```") == 1:
in_code_block = True
opening_fence = "```"
elif stripped_line.startswith("~~~"):
in_code_block = True
opening_fence = "~~~"
else:
if stripped_line.startswith(opening_fence):
in_code_block = False
opening_fence = ""
if in_code_block:
current_content.append(stripped_line)
continue
# Check each line against each of the header types (e.g., #, ##)
for sep, name in self.headers_to_split_on:
# Check if line starts with a header that we intend to split on
if stripped_line.startswith(sep) and (
# Header with no text OR header is followed by space
# Both are valid conditions that sep is being used a header
len(stripped_line) == len(sep) or stripped_line[len(sep)] == " "
):
# Ensure we are tracking the header as metadata
if name is not None:
# Get the current header level
current_header_level = sep.count("#")
# Pop out headers of lower or same level from the stack
while (
header_stack
and header_stack[-1]["level"] >= current_header_level
):
# We have encountered a new header
# at the same or higher level
popped_header = header_stack.pop()
# Clear the metadata for the
# popped header in initial_metadata
if popped_header["name"] in initial_metadata:
initial_metadata.pop(popped_header["name"])
# Push the current header to the stack
header: HeaderType = {
"level": current_header_level,
"name": name,
"data": stripped_line[len(sep) :].strip(),
}
header_stack.append(header)
# Update initial_metadata with the current header
initial_metadata[name] = header["data"]
# Add the previous line to the lines_with_metadata
# only if current_content is not empty
if current_content:
lines_with_metadata.append(
{
"content": "\n".join(current_content),
"metadata": current_metadata.copy(),
}
)
current_content.clear()
if not self.strip_headers:
current_content.append(stripped_line)
break
else:
if stripped_line:
current_content.append(stripped_line)
elif current_content:
lines_with_metadata.append(
{
"content": "\n".join(current_content),
"metadata": current_metadata.copy(),
}
)
current_content.clear()
current_metadata = initial_metadata.copy()
if current_content:
lines_with_metadata.append(
{"content": "\n".join(current_content), "metadata": current_metadata}
)
# lines_with_metadata has each line with associated header metadata
# aggregate these into chunks based on common metadata
if not self.return_each_line:
return self.aggregate_lines_to_chunks(lines_with_metadata)
else:
return [
Document(page_content=chunk["content"], metadata=chunk["metadata"])
for chunk in lines_with_metadata
]
class LineType(TypedDict):
"""Line type as typed dict."""
metadata: Dict[str, str]
content: str
class HeaderType(TypedDict):
"""Header type as typed dict."""
level: int
name: str
data: str
text-splitters: Introduce Experimental Markdown Syntax Splitter (#22257) #### Description This MR defines a `ExperimentalMarkdownSyntaxTextSplitter` class. The main goal is to replicate the functionality of the original `MarkdownHeaderTextSplitter` which extracts the header stack as metadata but with one critical difference: it keeps the whitespace of the original text intact. This draft reimplements the `MarkdownHeaderTextSplitter` with a very different algorithmic approach. Instead of marking up each line of the text individually and aggregating them back together into chunks, this method builds each chunk sequentially and applies the metadata to each chunk. This makes the implementation simpler. However, since it's designed to keep white space intact its not a full drop in replacement for the original. Since it is a radical implementation change to the original code and I would like to get feedback to see if this is a worthwhile replacement, should be it's own class, or is not a good idea at all. Note: I implemented the `return_each_line` parameter but I don't think it's a necessary feature. I'd prefer to remove it. This implementation also adds the following additional features: - Splits out code blocks and includes the language in the `"Code"` metadata key - Splits text on the horizontal rule `---` as well - The `headers_to_split_on` parameter is now optional - with sensible defaults that can be overridden. #### Issue Keeping the whitespace keeps the paragraphs structure and the formatting of the code blocks intact which allows the caller much more flexibility in how they want to further split the individuals sections of the resulting documents. This addresses the issues brought up by the community in the following issues: - https://github.com/langchain-ai/langchain/issues/20823 - https://github.com/langchain-ai/langchain/issues/19436 - https://github.com/langchain-ai/langchain/issues/22256 #### Dependencies N/A #### Twitter handle @RyanElston --------- Co-authored-by: isaac hershenson <ihershenson@hmc.edu>
2024-06-19 02:44:00 +00:00
class ExperimentalMarkdownSyntaxTextSplitter:
"""
An experimental text splitter for handling Markdown syntax.
This splitter aims to retain the exact whitespace of the original text while
extracting structured metadata, such as headers. It is a re-implementation of the
MarkdownHeaderTextSplitter with notable changes to the approach and
additional features.
Key Features:
- Retains the original whitespace and formatting of the Markdown text.
- Extracts headers, code blocks, and horizontal rules as metadata.
- Splits out code blocks and includes the language in the "Code" metadata key.
- Splits text on horizontal rules (`---`) as well.
- Defaults to sensible splitting behavior, which can be overridden using the
`headers_to_split_on` parameter.
Parameters:
----------
headers_to_split_on : List[Tuple[str, str]], optional
Headers to split on, defaulting to common Markdown headers if not specified.
return_each_line : bool, optional
When set to True, returns each line as a separate chunk. Default is False.
Usage example:
--------------
>>> headers_to_split_on = [
>>> ("#", "Header 1"),
>>> ("##", "Header 2"),
>>> ]
>>> splitter = ExperimentalMarkdownSyntaxTextSplitter(
>>> headers_to_split_on=headers_to_split_on
>>> )
>>> chunks = splitter.split(text)
>>> for chunk in chunks:
>>> print(chunk)
This class is currently experimental and subject to change based on feedback and
further development.
"""
DEFAULT_HEADER_KEYS = {
"#": "Header 1",
"##": "Header 2",
"###": "Header 3",
"####": "Header 4",
"#####": "Header 5",
"######": "Header 6",
}
def __init__(
self,
headers_to_split_on: Union[List[Tuple[str, str]], None] = None,
return_each_line: bool = False,
strip_headers: bool = True,
):
self.chunks: List[Document] = []
self.current_chunk = Document(page_content="")
self.current_header_stack: List[Tuple[int, str]] = []
self.strip_headers = strip_headers
if headers_to_split_on:
self.splittable_headers = dict(headers_to_split_on)
else:
self.splittable_headers = self.DEFAULT_HEADER_KEYS
self.return_each_line = return_each_line
def split_text(self, text: str) -> List[Document]:
raw_lines = text.splitlines(keepends=True)
while raw_lines:
raw_line = raw_lines.pop(0)
header_match = self._match_header(raw_line)
code_match = self._match_code(raw_line)
horz_match = self._match_horz(raw_line)
if header_match:
self._complete_chunk_doc()
if not self.strip_headers:
self.current_chunk.page_content += raw_line
# add the header to the stack
header_depth = len(header_match.group(1))
header_text = header_match.group(2)
self._resolve_header_stack(header_depth, header_text)
elif code_match:
self._complete_chunk_doc()
self.current_chunk.page_content = self._resolve_code_chunk(
raw_line, raw_lines
)
self.current_chunk.metadata["Code"] = code_match.group(1)
self._complete_chunk_doc()
elif horz_match:
self._complete_chunk_doc()
else:
self.current_chunk.page_content += raw_line
self._complete_chunk_doc()
# I don't see why `return_each_line` is a necessary feature of this splitter.
# It's easy enough to to do outside of the class and the caller can have more
# control over it.
if self.return_each_line:
return [
Document(page_content=line, metadata=chunk.metadata)
for chunk in self.chunks
for line in chunk.page_content.splitlines()
if line and not line.isspace()
]
return self.chunks
def _resolve_header_stack(self, header_depth: int, header_text: str) -> None:
for i, (depth, _) in enumerate(self.current_header_stack):
if depth == header_depth:
self.current_header_stack[i] = (header_depth, header_text)
self.current_header_stack = self.current_header_stack[: i + 1]
return
self.current_header_stack.append((header_depth, header_text))
def _resolve_code_chunk(self, current_line: str, raw_lines: List[str]) -> str:
chunk = current_line
while raw_lines:
raw_line = raw_lines.pop(0)
chunk += raw_line
if self._match_code(raw_line):
return chunk
return ""
def _complete_chunk_doc(self) -> None:
chunk_content = self.current_chunk.page_content
# Discard any empty documents
if chunk_content and not chunk_content.isspace():
# Apply the header stack as metadata
for depth, value in self.current_header_stack:
header_key = self.splittable_headers.get("#" * depth)
self.current_chunk.metadata[header_key] = value
self.chunks.append(self.current_chunk)
# Reset the current chunk
self.current_chunk = Document(page_content="")
# Match methods
def _match_header(self, line: str) -> Union[re.Match, None]:
match = re.match(r"^(#{1,6}) (.*)", line)
# Only matches on the configured headers
if match and match.group(1) in self.splittable_headers:
return match
return None
def _match_code(self, line: str) -> Union[re.Match, None]:
matches = [re.match(rule, line) for rule in [r"^```(.*)", r"^~~~(.*)"]]
return next((match for match in matches if match), None)
def _match_horz(self, line: str) -> Union[re.Match, None]:
matches = [
re.match(rule, line) for rule in [r"^\*\*\*+\n", r"^---+\n", r"^___+\n"]
]
return next((match for match in matches if match), None)