"""Chunking objects not specific to a particular chunking strategy.""" from __future__ import annotations import collections import copy import uuid from functools import cached_property from typing import Any, Callable, DefaultDict, Iterable, Iterator, cast import regex from lxml.etree import ParserError, tostring from lxml.html import fragment_fromstring from typing_extensions import Self, TypeAlias from unstructured.common.html_table import HtmlCell, HtmlRow, HtmlTable from unstructured.documents.elements import ( CodeSnippet, CompositeElement, ConsolidationStrategy, Element, ElementMetadata, Table, TableChunk, Title, ) from unstructured.logger import logger # ================================================================================================ # MODEL # ================================================================================================ CHUNK_MAX_CHARS_DEFAULT: int = 500 """Hard-max chunk-length when no explicit value specified in `max_characters` argument. Provided for reference only, for example so the ingest CLI can advertise the default value in its UI. External chunking-related functions (e.g. in ingest or decorators) should use `max_characters: int | None = None` and not apply this default themselves. Only `ChunkingOptions.max_characters` should apply a default value. """ CHUNK_MULTI_PAGE_DEFAULT: bool = True """When False, respect page-boundaries (no two elements from different page in same chunk). Only operative for "by_title" chunking strategy. """ BoundaryPredicate: TypeAlias = Callable[[Element], bool] """Detects when element represents crossing a semantic boundary like section or page.""" TextAndHtml: TypeAlias = tuple[str, str] class TokenCounter: """Token counting using tiktoken for token-based chunking. Lazily imports tiktoken only when token counting is first used. """ def __init__(self, tokenizer: str): self._tokenizer_name = tokenizer @cached_property def _encoder(self): """Lazily initialize the tiktoken encoder.""" import tiktoken try: # -- try as model name first (e.g., "gpt-4") -- return tiktoken.encoding_for_model(self._tokenizer_name) except KeyError: # -- fall back to encoding name (e.g., "cl100k_base") -- return tiktoken.get_encoding(self._tokenizer_name) def count(self, text: str) -> int: """Return the number of tokens in `text`.""" return len(self._encoder.encode(text)) # ================================================================================================ # CHUNKING OPTIONS # ================================================================================================ class ChunkingOptions: """Specifies parameters of optional chunking behaviors. Parameters ---------- max_characters Hard-maximum text-length of chunk. A chunk longer than this will be split mid-text and be emitted as two or more chunks. Mutually exclusive with `max_tokens`. max_tokens Hard-maximum token count of chunk. A chunk with more tokens than this will be split mid-text and be emitted as two or more chunks. Requires `tokenizer` to be specified. Mutually exclusive with `max_characters`. new_after_n_chars Preferred approximate chunk size. A chunk composed of elements totalling this size or greater is considered "full" and will not be enlarged by adding another element, even if it will fit within the remaining `max_characters` for that chunk. Defaults to `max_characters` when not specified, which effectively disables this behavior. Specifying 0 for this argument causes each element to appear in a chunk by itself (although an element with text longer than `max_characters` will be still be split into two or more chunks). new_after_n_tokens Token-based equivalent of `new_after_n_chars`. Preferred approximate chunk size in tokens. Requires `tokenizer` and `max_tokens` to be specified. combine_text_under_n_chars Provides a way to "recombine" small chunks formed by breaking on a semantic boundary. Only relevant for a chunking strategy that specifies higher-level semantic boundaries to be respected, like "section" or "page". Recursively combines two adjacent pre-chunks when the first pre-chunk is smaller than this threshold. "Recursively" here means the resulting pre-chunk can be combined with the next pre-chunk if it is still under the length threshold. Defaults to `max_characters` which combines chunks whenever space allows. Specifying 0 for this argument suppresses combining of small chunks. Note this value is "capped" at the `new_after_n_chars` value since a value higher than that would not change this parameter's effect. overlap Specifies the length of a string ("tail") to be drawn from each chunk and prefixed to the next chunk as a context-preserving mechanism. By default, this only applies to split-chunks where an oversized element is divided into multiple chunks by text-splitting. overlap_all Default: `False`. When `True`, apply overlap between "normal" chunks formed from whole elements and not subject to text-splitting. Use this with caution as it entails a certain level of "pollution" of otherwise clean semantic chunk boundaries. repeat_table_headers Default: `True`. When `True`, repeated table-header behavior is enabled for chunked table continuations. Specify `False` to opt out and preserve legacy table-chunk behavior. text_splitting_separators A sequence of strings like `("\n", " ")` to be used as target separators during text-splitting. Text-splitting only applies to splitting an oversized element into two or more chunks. These separators are tried in the specified order until one is found in the string to be split. The default separator is `""` which matches between any two characters. This separator should not be specified in this sequence because it is always the separator of last-resort. Note that because the separator is removed during text-splitting, only whitespace character sequences are suitable. tokenizer The tokenizer to use for token-based chunking. Can be either an encoding name (e.g., "cl100k_base") or a model name (e.g., "gpt-4"). Required when using `max_tokens`. """ def __init__(self, **kwargs: Any): self._kwargs = kwargs @classmethod def new(cls, **kwargs: Any) -> Self: """Return instance or raises `ValueError` on invalid arguments like overlap > max_chars.""" self = cls(**kwargs) self._validate() return self @cached_property def boundary_predicates(self) -> tuple[BoundaryPredicate, ...]: """The semantic-boundary detectors to be applied to break pre-chunks. Overridden by sub-typs to provide semantic-boundary isolation behaviors. """ return () @cached_property def combine_text_under_n_chars(self) -> int: """Combine two consecutive text pre-chunks if first is smaller than this and both will fit. Default applied here is `0` which essentially disables chunk combining. Must be overridden by subclass where combining behavior is supported. """ arg_value = self._kwargs.get("combine_text_under_n_chars") return arg_value if arg_value is not None else 0 @cached_property def hard_max(self) -> int: """The maximum size for a chunk (in characters or tokens depending on mode). A pre-chunk will only exceed this size when it contains exactly one element which by itself exceeds this size. Such a pre-chunk is subject to mid-text splitting later in the chunking process. """ if self.use_token_counting: # -- token-based chunking: max_tokens is required and validated -- return self._kwargs["max_tokens"] arg_value = self._kwargs.get("max_characters") return arg_value if arg_value is not None else CHUNK_MAX_CHARS_DEFAULT @cached_property def include_orig_elements(self) -> bool: """When True, add original elements from pre-chunk to `.metadata.orig_elements` of chunk. Default value is `True`. """ arg_value = self._kwargs.get("include_orig_elements") return True if arg_value is None else bool(arg_value) @cached_property def repeat_table_headers(self) -> bool: """When True, repeat detected table headers in continuation table chunks. Default value is `True`. """ arg_value = self._kwargs.get("repeat_table_headers") return True if arg_value is None else bool(arg_value) @cached_property def skip_table_chunking(self) -> bool: """When True, Table elements are passed through without chunking. Default value is `False`. """ arg_value = self._kwargs.get("skip_table_chunking") return False if arg_value is None else bool(arg_value) @cached_property def inter_chunk_overlap(self) -> int: """Characters of overlap to add between chunks. This applies only to boundaries between chunks formed from whole elements and not to text-splitting boundaries that arise from splitting an oversized element. """ overlap_all_arg = self._kwargs.get("overlap_all") return self.overlap if overlap_all_arg else 0 @cached_property def overlap(self) -> int: """The number of characters to overlap text when splitting chunks mid-text. The actual overlap will not exceed this number of characters but may be less as required to respect splitting-character boundaries. """ overlap_arg = self._kwargs.get("overlap") return overlap_arg or 0 @cached_property def soft_max(self) -> int: """A pre-chunk of this size or greater is considered full. Note that while a value of `0` is valid, it essentially disables chunking by putting each element into its own chunk. """ hard_max = self.hard_max if self.use_token_counting: new_after_n_tokens_arg = self._kwargs.get("new_after_n_tokens") # -- default value is == max_tokens -- if new_after_n_tokens_arg is None: return hard_max # -- new_after_n_tokens > max_tokens behaves the same as ==max_tokens -- if new_after_n_tokens_arg > hard_max: return hard_max return new_after_n_tokens_arg new_after_n_chars_arg = self._kwargs.get("new_after_n_chars") # -- default value is == max_characters -- if new_after_n_chars_arg is None: return hard_max # -- new_after_n_chars > max_characters behaves the same as ==max_characters -- if new_after_n_chars_arg > hard_max: return hard_max # -- otherwise, give them what they asked for -- return new_after_n_chars_arg @cached_property def split(self) -> Callable[[str], tuple[str, str]]: """A text-splitting function suitable for splitting the text of an oversized pre-chunk. The function is pre-configured with the chosen chunking window size and any other applicable options specified by the caller as part of this chunking-options instance. """ return _TextSplitter(self) @cached_property def text_separator(self) -> str: """The string to insert between elements when concatenating their text for a chunk. Right now this is just "\n\n" (a blank line in plain text), but having this here rather than as a module-level constant provides a way for us to easily make it user-configurable in future if we want to. """ return "\n\n" @cached_property def text_splitting_separators(self) -> tuple[str, ...]: """Sequence of text-splitting target strings to be used in order of preference.""" text_splitting_separators_arg = self._kwargs.get("text_splitting_separators") return ( ("\n", " ") if text_splitting_separators_arg is None else tuple(text_splitting_separators_arg) ) @cached_property def token_counter(self) -> TokenCounter | None: """The token counter for token-based chunking, or None for character-based chunking.""" tokenizer = self._kwargs.get("tokenizer") return TokenCounter(tokenizer) if tokenizer else None @cached_property def use_token_counting(self) -> bool: """True when token-based chunking is configured, False for character-based.""" return self._kwargs.get("max_tokens") is not None def measure(self, text: str) -> int: """Return the size of `text` in the configured units (characters or tokens).""" if self.use_token_counting and self.token_counter: return self.token_counter.count(text) return len(text) def _validate(self) -> None: """Raise ValueError if requestion option-set is invalid.""" max_tokens = self._kwargs.get("max_tokens") max_characters = self._kwargs.get("max_characters") tokenizer = self._kwargs.get("tokenizer") # -- max_tokens and max_characters are mutually exclusive -- if max_tokens is not None and max_characters is not None: raise ValueError( "'max_tokens' and 'max_characters' are mutually exclusive;" " specify one or the other, not both" ) # -- max_tokens requires tokenizer -- if max_tokens is not None and tokenizer is None: raise ValueError("'tokenizer' is required when using 'max_tokens'") # -- max_tokens must be positive -- if max_tokens is not None and max_tokens <= 0: raise ValueError(f"'max_tokens' argument must be > 0, got {max_tokens}") # -- new_after_n_tokens requires max_tokens -- new_after_n_tokens = self._kwargs.get("new_after_n_tokens") if new_after_n_tokens is not None and max_tokens is None: raise ValueError("'new_after_n_tokens' requires 'max_tokens' to be specified") # -- new_after_n_tokens must be non-negative -- if new_after_n_tokens is not None and new_after_n_tokens < 0: raise ValueError( f"'new_after_n_tokens' argument must be >= 0, got {new_after_n_tokens}" ) # -- chunking window must have positive length -- hard_max = self.hard_max if hard_max <= 0: raise ValueError(f"'max_characters' argument must be > 0, got {hard_max}") # -- a negative value for `new_after_n_chars` is assumed to be a mistake the caller will # -- want to know about new_after_n_chars = self._kwargs.get("new_after_n_chars") if new_after_n_chars is not None and new_after_n_chars < 0: raise ValueError(f"'new_after_n_chars' argument must be >= 0, got {new_after_n_chars}") # -- overlap must be less than max-chars or the chunk text will never be consumed -- if self.overlap >= hard_max: raise ValueError( f"'overlap' argument must be less than `max_characters`," f" got {self.overlap} >= {hard_max}" ) # ================================================================================================ # TABLE ISOLATION (SHARED PRECHECKS) # ================================================================================================ # Tables are always staged alone in a pre-chunk so downstream splitting can emit `Table` / # `TableChunk` elements instead of folding them into `CompositeElement` with surrounding text. # See GitHub issue #3921 and the chunking strategy docs for rationale. def _element_is_table_family(element: Element) -> bool: """True when ``element`` is a `Table` or a concrete subtype such as `TableChunk`. Subclasses share the same isolation contract: they must not share a pre-chunk with arbitrary text elements, and two table-bearing sequences must not be merged by `PreChunkCombiner`. """ return isinstance(element, Table) def _elements_contain_table_family(elements: Iterable[Element]) -> bool: """True when ``elements`` already includes at least one table-family element.""" return any(_element_is_table_family(e) for e in elements) def _table_isolation_forbids_side_by_side_merge( left: Iterable[Element], right: Iterable[Element], ) -> bool: """True when a proposed merge of two element streams must be rejected for table isolation. If either side already contains a table, the combiner must flush before accepting the other side. This keeps `combine_text_under_n_chars` from concatenating a table pre-chunk with neighboring narrative pre-chunks. """ return _elements_contain_table_family(left) or _elements_contain_table_family(right) # ================================================================================================ # PRE-CHUNKER # ================================================================================================ class PreChunker: """Gathers sequential elements into pre-chunks as length constraints allow. The pre-chunker's responsibilities are: - **Segregate semantic units.** Identify semantic unit boundaries and segregate elements on either side of those boundaries into different sections. In this case, the primary indicator of a semantic boundary is a `Title` element. A page-break (change in page-number) is also a semantic boundary when `multipage_sections` is `False`. - **Minimize chunk count for each semantic unit.** Group the elements within a semantic unit into sections as big as possible without exceeding the chunk window size. - **Minimize chunks that must be split mid-text.** Precompute the text length of each section and only produce a section that exceeds the chunk window size when there is a single element with text longer than that window. A Table element is placed into a section by itself. CheckBox elements are dropped. The "by-title" strategy specifies breaking on section boundaries; a `Title` element indicates a new "section", hence the "by-title" designation. """ def __init__(self, elements: Iterable[Element], opts: ChunkingOptions): self._elements = elements self._opts = opts @classmethod def iter_pre_chunks( cls, elements: Iterable[Element], opts: ChunkingOptions ) -> Iterator[PreChunk]: """Generate pre-chunks from the element-stream provided on construction.""" return cls(elements, opts)._iter_pre_chunks() def _iter_pre_chunks(self) -> Iterator[PreChunk]: """Generate pre-chunks from the element-stream provided on construction. A *pre-chunk* is the largest sub-sequence of elements that will both fit within the chunking window and respects the semantic boundary rules of the chunking strategy. When a single element exceeds the chunking window size it is placed in a pre-chunk by itself and is subject to mid-text splitting in the second phase of the chunking process. """ pre_chunk_builder = PreChunkBuilder(self._opts) for element in self._elements: # -- start new pre-chunk when necessary to uphold segregation guarantees -- if ( # -- start new pre-chunk when necessary to uphold segregation guarantees -- self._is_in_new_semantic_unit(element) # -- or when next element won't fit -- or not pre_chunk_builder.will_fit(element) ): yield from pre_chunk_builder.flush() # -- add this element to the work-in-progress (WIP) pre-chunk -- pre_chunk_builder.add_element(element) # -- flush "tail" pre-chunk, any partially-filled pre-chunk after last element is # -- processed yield from pre_chunk_builder.flush() @cached_property def _boundary_predicates(self) -> tuple[BoundaryPredicate, ...]: """The semantic-boundary detectors to be applied to break pre-chunks.""" return self._opts.boundary_predicates def _is_in_new_semantic_unit(self, element: Element) -> bool: """True when `element` begins a new semantic unit such as a section or page.""" # -- all detectors need to be called to update state and avoid double counting # -- boundaries that happen to coincide, like Table and new section on same element. # -- Using `any()` would short-circuit on first True. semantic_boundaries = [pred(element) for pred in self._boundary_predicates] return any(semantic_boundaries) class PreChunkBuilder: """An element accumulator suitable for incrementally forming a pre-chunk. Provides the trial method `.will_fit()` a pre-chunker can use to determine whether it should add the next element in the element stream. `.flush()` is used to build a PreChunk object from the accumulated elements. This method returns an iterator that generates zero-or-one `PreChunk` object and is used like so: yield from builder.flush() If no elements have been accumulated, no `PreChunk` instance is generated. Flushing the builder clears the elements it contains so it is ready to build the next pre-chunk. """ def __init__(self, opts: ChunkingOptions) -> None: self._opts = opts self._separator_len = len(opts.text_separator) self._elements: list[Element] = [] # -- overlap is only between pre-chunks so starts empty -- self._overlap_prefix: str = "" # -- only includes non-empty element text, e.g. PageBreak.text=="" is not included -- self._text_segments: list[str] = [] # -- combined length of text-segments, not including separators -- self._text_len: int = 0 def add_element(self, element: Element) -> None: """Add `element` to this section.""" # -- do not prefix a table-only pre-chunk with narrative overlap from the prior chunk -- if len(self._elements) == 0 and _element_is_table_family(element): self._overlap_prefix = "" self._text_segments = [] self._text_len = 0 self._elements.append(element) if element.text: self._text_segments.append(element.text) # -- only track char-based length; token-based length computed on demand -- self._text_len += len(element.text) def flush(self) -> Iterator[PreChunk]: """Generate zero-or-one `PreChunk` object and clear the accumulator. Suitable for use to emit a PreChunk when the maximum size has been reached or a semantic boundary has been reached. Also to clear out a terminal pre-chunk at the end of an element stream. """ elements = self._elements if not elements: return # -- copy element list, don't use original or it may change contents as builder proceeds -- pre_chunk = PreChunk(elements, self._overlap_prefix, self._opts) # -- clear builder before yield so we're not sensitive to the timing of how/when this # -- iterator is exhausted and can add elements for the next pre-chunk immediately. overlap_for_next = pre_chunk.overlap_tail # -- table tails must not prefix the following narrative pre-chunk (overlap_all) -- if len(elements) == 1 and _element_is_table_family(elements[0]): overlap_for_next = "" self._reset_state(overlap_for_next) yield pre_chunk def will_fit(self, element: Element) -> bool: """True when `element` can be added to this prechunk without violating its limits. There are several limits: - A `Table` element will never fit with any other element. It will only fit in an empty pre-chunk. - No element will fit in a pre-chunk that already contains a `Table` element. - A text-element will not fit in a pre-chunk that already exceeds the soft-max (aka. new_after_n_chars/new_after_n_tokens). - A text-element will not fit when together with the elements already present it would exceed the hard-max (aka. max_characters/max_tokens). """ # -- a `Table` can only start a pre-chunk; it is never appended to a non-empty pre-chunk -- if _element_is_table_family(element): return len(self._elements) == 0 # -- no non-table element may share a pre-chunk with a `Table` -- if _elements_contain_table_family(self._elements): return False # -- an empty pre-chunk will accept any element (including an oversized-element) -- if len(self._elements) == 0: return True # -- a pre-chunk that already exceeds the soft-max is considered "full" -- if self._text_length > self._opts.soft_max: return False # -- don't add an element if it would increase total size beyond the hard-max -- # -- for token counting, compute what the new total would be -- if self._opts.use_token_counting: new_text = self._opts.text_separator.join( self._text_segments + ([element.text] if element.text else []) ) return self._opts.measure(new_text) <= self._opts.hard_max # -- for character counting, use the efficient incremental approach -- return not self._remaining_space < len(element.text or "") @property def _remaining_space(self) -> int: """Maximum text-length of an element that can be added without exceeding maxlen.""" # -- include length of trailing separator that will go before next element text -- separators_len = self._separator_len * len(self._text_segments) return self._opts.hard_max - self._text_len - separators_len def _reset_state(self, overlap_prefix: str) -> None: """Set working-state values back to "empty", ready to accumulate next pre-chunk.""" self._overlap_prefix = overlap_prefix self._elements.clear() self._text_segments = [overlap_prefix] if overlap_prefix else [] self._text_len = len(overlap_prefix) @property def _text_length(self) -> int: """Size of the text in this pre-chunk (in characters or tokens depending on mode). This value represents the chunk-size that would result if this pre-chunk was flushed in its current state. In particular, it does not include the length of a trailing separator (since that would only appear if an additional element was added). Not suitable for judging remaining space, use `.remaining_space` for that value. """ # -- for token counting, compute the actual token count of the joined text -- if self._opts.use_token_counting: if not self._text_segments: return 0 text = self._opts.text_separator.join(self._text_segments) return self._opts.measure(text) # -- for character counting, use the efficient incremental approach -- # -- number of text separators present in joined text of elements. This includes only # -- separators *between* text segments, not one at the end. Note there are zero separators # -- for both 0 and 1 text-segments. n = len(self._text_segments) separator_count = n - 1 if n else 0 return self._text_len + (separator_count * self._separator_len) # ================================================================================================ # PRE-CHUNK # ================================================================================================ class PreChunk: """Sequence of elements staged to form a single chunk. This object is purposely immutable. """ def __init__( self, elements: Iterable[Element], overlap_prefix: str, opts: ChunkingOptions ) -> None: self._elements = list(elements) self._overlap_prefix = overlap_prefix self._opts = opts def __eq__(self, other: Any) -> bool: if not isinstance(other, PreChunk): return False return self._overlap_prefix == other._overlap_prefix and self._elements == other._elements def can_combine(self, pre_chunk: PreChunk) -> bool: """True when `pre_chunk` can be combined with this one without exceeding size limits.""" if _table_isolation_forbids_side_by_side_merge(self._elements, pre_chunk._elements): return False if len(self._text) >= self._opts.combine_text_under_n_chars: return False # -- avoid duplicating length computations by doing a trial-combine which is just as # -- efficient and definitely more robust than hoping two different computations of combined # -- length continue to get the same answer as the code evolves. Only possible because # -- `.combine()` is non-mutating. combined_len = len(self.combine(pre_chunk)._text) return combined_len <= self._opts.hard_max def combine(self, other_pre_chunk: PreChunk) -> PreChunk: """Return new `PreChunk` that combines this and `other_pre_chunk`.""" # -- combined pre-chunk gets the overlap-prefix of the first pre-chunk. The second overlap # -- is automatically incorporated at the end of the first chunk, where it originated. return PreChunk( self._elements + other_pre_chunk._elements, overlap_prefix=self._overlap_prefix, opts=self._opts, ) def iter_chunks(self) -> Iterator[CompositeElement | Table | TableChunk]: """Form this pre-chunk into one or more chunk elements maxlen or smaller. When the total size of the pre-chunk will fit in the chunking window, a single chunk it emitted. When this prechunk contains an oversized element (always isolated), it is split into two or more chunks that each fit the chunking window. """ # -- a one-table-only pre-chunk is handled specially, by `TablePreChunk`, mainly because # -- it may need to be split into multiple `TableChunk` elements and that operation is # -- quite specialized. if len(self._elements) == 1 and isinstance(self._elements[0], Table): if self._opts.skip_table_chunking: yield self._elements[0] else: yield from _TableChunker.iter_chunks( self._elements[0], self._overlap_prefix, self._opts ) else: yield from _Chunker.iter_chunks(self._elements, self._text, self._opts) @cached_property def overlap_tail(self) -> str: """The portion of this chunk's text to be repeated as a prefix in the next chunk. This value is the empty-string ("") when either the `.overlap` length option is `0` or `.overlap_all` is `False`. When there is a text value, it is stripped of both leading and trailing whitespace. """ overlap = self._opts.inter_chunk_overlap return self._text[-overlap:].strip() if overlap else "" def _iter_text_segments(self) -> Iterator[str]: """Generate overlap text and each element text segment in order. Empty text segments are not included. CodeSnippet elements preserve their original whitespace (including newlines) to maintain code formatting. """ if self._overlap_prefix: yield self._overlap_prefix for e in self._elements: if e.text and len(e.text): # -- preserve all whitespace for code snippets to maintain formatting -- if isinstance(e, CodeSnippet): yield e.text else: text = " ".join(e.text.strip().split()) if text: yield text @cached_property def _text(self) -> str: """The concatenated text of all elements in this pre-chunk, including any overlap. Whitespace is normalized to a single space. The text of each element is separated from that of the next by a blank line ("\n\n"). """ return self._opts.text_separator.join(self._iter_text_segments()) # ================================================================================================ # CHUNKING HELPER/SPLITTERS # ================================================================================================ class _Chunker: """Forms chunks from a pre-chunk other than one containing only a `Table`. Produces zero-or-more `CompositeElement` objects. """ def __init__(self, elements: Iterable[Element], text: str, opts: ChunkingOptions) -> None: self._elements = list(elements) self._text = text self._opts = opts @classmethod def iter_chunks( cls, elements: Iterable[Element], text: str, opts: ChunkingOptions ) -> Iterator[CompositeElement]: """Form zero or more chunks from `elements`. One `CompositeElement` is produced when all `elements` will fit. Otherwise there is a single `Text`-subtype element and chunks are formed by splitting. """ return cls(elements, text, opts)._iter_chunks() def _iter_chunks(self) -> Iterator[CompositeElement]: """Form zero or more chunks from `elements`.""" # -- a pre-chunk containing no text (maybe only a PageBreak element for example) does not # -- generate any chunks. if not self._text: return # -- `split()` is the text-splitting function used to split an oversized element -- split = self._opts.split # -- emit first chunk -- s, remainder = split(self._text) yield CompositeElement(text=s, metadata=self._consolidated_metadata) # -- an oversized pre-chunk will have a remainder, split that up into additional chunks. # -- Note these get continuation_metadata which includes is_continuation=True. while remainder: s, remainder = split(remainder) yield CompositeElement(text=s, metadata=self._continuation_metadata) @cached_property def _all_metadata_values(self) -> dict[str, list[Any]]: """Collection of all populated metadata values across elements. The resulting dict has one key for each `ElementMetadata` field that had a non-None value in at least one of the elements in this pre-chunk. The value of that key is a list of all those populated values, in element order, for example: { "filename": ["sample.docx", "sample.docx"], "languages": [["lat"], ["lat", "eng"]] ... } This preprocessing step provides the input for a specified consolidation strategy that will resolve the list of values for each field to a single consolidated value. """ def iter_populated_fields(metadata: ElementMetadata) -> Iterator[tuple[str, Any]]: """(field_name, value) pair for each non-None field in single `ElementMetadata`.""" return ( (field_name, value) for field_name, value in metadata.known_fields.items() if value is not None ) field_values: DefaultDict[str, list[Any]] = collections.defaultdict(list) # -- collect all non-None field values in a list for each field, in element-order -- for e in self._elements: for field_name, value in iter_populated_fields(e.metadata): field_values[field_name].append(value) return dict(field_values) @cached_property def _consolidated_metadata(self) -> ElementMetadata: """Metadata applicable to this pre-chunk as a single chunk. Formed by applying consolidation rules to all metadata fields across the elements of this pre-chunk. For the sake of consistency, the same rules are applied (for example, for dropping values) to a single-element pre-chunk too, even though metadata for such a pre-chunk is already "consolidated". """ consolidated_metadata = ElementMetadata(**self._meta_kwargs) if self._opts.include_orig_elements: consolidated_metadata.orig_elements = self._orig_elements return consolidated_metadata @cached_property def _continuation_metadata(self) -> ElementMetadata: """Metadata applicable to the second and later text-split chunks of the pre-chunk. The same metadata as the first text-split chunk but includes `.is_continuation = True`. Unused for non-oversized pre-chunks since those are not subject to text-splitting. """ # -- we need to make a copy, otherwise adding a field would also change metadata value # -- already assigned to another chunk (e.g. the first text-split chunk). Deep-copy is not # -- required though since we're not changing any collection fields. continuation_metadata = copy.copy(self._consolidated_metadata) continuation_metadata.is_continuation = True return continuation_metadata @cached_property def _meta_kwargs(self) -> dict[str, Any]: """The consolidated metadata values as a dict suitable for constructing ElementMetadata. This is where consolidation strategies are actually applied. The output is suitable for use in constructing an `ElementMetadata` object like `ElementMetadata(**self._meta_kwargs)`. """ CS = ConsolidationStrategy field_consolidation_strategies = ConsolidationStrategy.field_consolidation_strategies() def iter_kwarg_pairs() -> Iterator[tuple[str, Any]]: """Generate (field-name, value) pairs for each field in consolidated metadata.""" for field_name, values in self._all_metadata_values.items(): strategy = field_consolidation_strategies.get(field_name) if strategy is CS.FIRST: yield field_name, values[0] # -- concatenate lists from each element that had one, in order -- elif strategy is CS.LIST_CONCATENATE: yield field_name, sum(values, cast("list[Any]", [])) # -- union lists from each element, preserving order of appearance -- elif strategy is CS.LIST_UNIQUE: # -- Python 3.7+ maintains dict insertion order -- ordered_unique_keys = {key: None for val_list in values for key in val_list} yield field_name, list(ordered_unique_keys.keys()) elif strategy is CS.STRING_CONCATENATE: yield field_name, " ".join(val.strip() for val in values) elif strategy is CS.DROP: continue else: # pragma: no cover # -- not likely to hit this since we have a test in `text_elements.py` that # -- ensures every ElementMetadata fields has an assigned strategy. raise NotImplementedError( f"metadata field {repr(field_name)} has no defined consolidation strategy" ) return dict(iter_kwarg_pairs()) @cached_property def _orig_elements(self) -> list[Element]: """The `.metadata.orig_elements` value for chunks formed from this pre-chunk.""" def iter_orig_elements(): for e in self._elements: if e.metadata.orig_elements is None: yield e continue # -- make copy of any element we're going to mutate because these elements don't # -- belong to us (the user may have downstream purposes for them). orig_element = copy.copy(e) # -- prevent recursive .orig_elements when element is a chunk (has orig-elements of # -- its own) orig_element.metadata.orig_elements = None yield orig_element return list(iter_orig_elements()) class _TableChunker: """Responsible for forming chunks, especially splits, from a single-table pre-chunk. Table splitting is specialized because we recursively split on an even row, cell, text boundary. This object encapsulate those details. """ def __init__(self, table: Table, overlap_prefix: str, opts: ChunkingOptions) -> None: self._table = table self._overlap_prefix = overlap_prefix self._opts = opts @classmethod def iter_chunks( cls, table: Table, overlap_prefix: str, opts: ChunkingOptions ) -> Iterator[Table | TableChunk]: """Split this pre-chunk into `Table` or `TableChunk` objects maxlen or smaller.""" return cls(table, overlap_prefix, opts)._iter_chunks() def _iter_chunks(self) -> Iterator[Table | TableChunk]: """Split this pre-chunk into `Table` or `TableChunk` objects maxlen or smaller.""" # -- A table with no non-whitespace text produces no chunks -- if not self._table_text: return # -- only text-split a table when it's longer than the chunking window -- maxlen = self._opts.hard_max measure = self._opts.measure text_size = measure(self._text_with_overlap) html_size = measure(self._html) if self._html else 0 if text_size <= maxlen and html_size <= maxlen: # -- use the compactified html for .text_as_html, even though we're not splitting -- metadata = self._metadata metadata.text_as_html = self._html or None # -- note the overlap-prefix is prepended to its text -- yield Table(text=self._text_with_overlap, metadata=metadata) return # -- When there's no HTML, split it like a normal element. Also fall back to text-only # -- chunks when `max_characters` is less than 50 (or in token mode, less than 15 tokens). # -- `.text_as_html` metadata is impractical for a chunking window that small because the # -- 33 characters of HTML overhead for each chunk would produce many very small chunks. min_html_threshold = 15 if self._opts.use_token_counting else 50 if not self._html or self._opts.hard_max < min_html_threshold: yield from self._iter_text_only_table_chunks() return # -- otherwise, form splits with "synchronized" text and html -- yield from self._iter_text_and_html_table_chunks() @cached_property def _html(self) -> str: """The compactified HTML for this table when it has text-as-HTML. The empty string when table-structure has not been captured, perhaps because `infer_table_structure` was set `False` in the partitioning call. """ if not (html_table := self._html_table): return "" return html_table.html @cached_property def _html_table(self) -> HtmlTable | None: """The `lxml` HTML element object for this table. `None` when the `Table` element has no `.metadata.text_as_html`. """ if (text_as_html := self._table.metadata.text_as_html) is None: return None text_as_html = text_as_html.strip() if not text_as_html: # pragma: no cover return None try: return HtmlTable.from_html_text(text_as_html) except (ParserError, ValueError): logger.warning( "Could not parse text_as_html for table element; skipping HTML-based chunking." " text_as_html: %s", text_as_html[:100] + "..." if len(text_as_html) > 100 else text_as_html, ) return None @cached_property def _leading_header_row_count(self) -> int: """Number of contiguous leading rows that should be treated as table headers.""" html_table = self._html_table if html_table is None: return 0 count = 0 for row in html_table.iter_rows(): if not row.is_header: break count += 1 return count def _iter_text_and_html_table_chunks(self) -> Iterator[TableChunk]: """Split table into chunks where HTML corresponds exactly to text. `.metadata.text_as_html` for each chunk is a parsable `` HTML fragment. """ if (html_table := self._html_table) is None: # pragma: no cover raise ValueError("this method is undefined for a table having no .text_as_html") header_row_count = self._leading_header_row_count if self._opts.repeat_table_headers else 0 splitter = _HtmlTableSplitter( html_table, self._opts, header_row_count=header_row_count, ) yield from self._make_table_chunks( _HtmlTableSplitter.iter_subtables( html_table, self._opts, header_row_count=header_row_count, ), num_carried_over_header_rows=splitter.carried_over_header_row_count, ) def _iter_text_only_table_chunks(self) -> Iterator[TableChunk]: """Split oversized text-only table (no text-as-html) into chunks. `.metadata.text_as_html` is optional, not included when `infer_table_structure` is `False`. """ def _iter_text_splits() -> Iterator[tuple[str, None]]: text_remainder = self._text_with_overlap split = self._opts.split while text_remainder: # -- split off the next chunk-worth of characters into a TableChunk -- chunk_text, text_remainder = split(text_remainder) yield chunk_text, None yield from self._make_table_chunks(_iter_text_splits()) def _make_table_chunks( self, text_html_pairs: Iterator[tuple[str, str | None]], num_carried_over_header_rows: int = 0, ) -> Iterator[TableChunk]: """Form `TableChunk` objects from (text, html) pairs. Handles `is_continuation` and chunk sequencing metadata (`table_id`, `chunk_index`) so the original table can be reconstructed from its chunks. Carries `num_carried_over_header_rows` so synthetic repeated header rows can be removed. """ table_id = str(uuid.uuid4()) carried_header_row_count = max(0, num_carried_over_header_rows) for chunk_index, (text, html) in enumerate(text_html_pairs): metadata = self._metadata if html is not None: metadata.text_as_html = html else: metadata.text_as_html = None # -- second and later chunks get `.metadata.is_continuation = True` -- metadata.is_continuation = (chunk_index > 0) or None metadata.num_carried_over_header_rows = ( carried_header_row_count if chunk_index > 0 else 0 ) chunk = TableChunk(text=text, metadata=metadata) chunk.metadata.table_id = table_id chunk.metadata.chunk_index = chunk_index yield chunk @property def _metadata(self) -> ElementMetadata: """The base `.metadata` value for chunks formed from this pre-chunk. The term "base" here means that other metadata fields will be added, depending on the chunk. In particular, `.metadata.text_as_html` will be different for each text-split chunk and `.metadata.is_continuation` must be added for second-and-later text-split chunks. Note this is a fresh copy of the metadata on each call since it will need to be mutated differently for each chunk formed from this pre-chunk. """ CS = ConsolidationStrategy metadata = copy.deepcopy(self._table.metadata) # -- drop metadata fields not appropriate for chunks, in particular # -- parent_id's will not reliably point to an existing element drop_field_names = [ field_name for field_name, strategy in CS.field_consolidation_strategies().items() if strategy is CS.DROP ] for field_name in drop_field_names: setattr(metadata, field_name, None) if self._opts.include_orig_elements: metadata.orig_elements = self._orig_elements return metadata @cached_property def _orig_elements(self) -> list[Element]: """The `.metadata.orig_elements` value for chunks formed from this pre-chunk. Note this is not just the `Table` element, it must be adjusted to strip out any `.metadata.orig_elements` value it may have when it is itself a chunk and not a direct product of partitioning. """ # -- make a copy because we're going to mutate the `Table` element and it doesn't belong to # -- us (the user may have downstream purposes for it). orig_table = copy.deepcopy(self._table) # -- prevent recursive .orig_elements when `Table` element is a chunk -- orig_table.metadata.orig_elements = None return [orig_table] @cached_property def _table_text(self) -> str: """The text in this table, not including any overlap-prefix or extra whitespace.""" if not self._table.text: return "" return " ".join(self._table.text.split()) @cached_property def _text_with_overlap(self) -> str: """The text for this chunk, including the overlap-prefix when present.""" overlap_prefix = self._overlap_prefix table_text = "" if not self._table.text else self._table.text.strip() # -- use row-separator between overlap and table-text -- return overlap_prefix + "\n" + table_text if overlap_prefix else table_text # ================================================================================================ # HTML SPLITTERS # ================================================================================================ class _HtmlTableSplitter: """Produces (text, html) pairs for a `
` HtmlElement. Each chunk contains a whole number of rows whenever possible. An oversized row is split on an even cell boundary and a single cell that is by itself too big to fit in the chunking window is divided by text-splitting. The returned `html` value is always a parseable HTML `
` subtree. """ def __init__(self, table_element: HtmlTable, opts: ChunkingOptions, header_row_count: int = 0): self._table_element = table_element self._opts = opts self._header_row_count = max(0, header_row_count) @classmethod def iter_subtables( cls, table_element: HtmlTable, opts: ChunkingOptions, header_row_count: int = 0 ) -> Iterator[TextAndHtml]: """Generate (text, html) pair for each split of this table pre-chunk. Each split is on an even row boundary whenever possible, falling back to even cell and even word boundaries when a row or cell is by itself oversized, respectively. """ return cls(table_element, opts, header_row_count=header_row_count)._iter_subtables() def _iter_subtables(self) -> Iterator[TextAndHtml]: """Generate (text, html) pairs containing as many whole rows as will fit in window. Falls back to splitting rows into whole cells when a single row is by itself too big to fit in the chunking window. """ is_first_chunk = True accum = _RowAccumulator(maxlen=self._maxlen(is_first_chunk), measure=self._opts.measure) for row in self._table_element.iter_rows(): # -- if row won't fit, any WIP chunk is done, send it on its way -- if not accum.will_fit(row): for text, html in accum.flush(): yield self._prepend_repeated_headers(text, html, is_first_chunk) is_first_chunk = False accum = _RowAccumulator( maxlen=self._maxlen(is_first_chunk), measure=self._opts.measure ) # -- if row fits, add it to accumulator -- if accum.will_fit(row): accum.add_row(row) else: # -- otherwise, single row is bigger than chunking window -- for text, html in self._iter_row_splits(row, maxlen=self._maxlen(is_first_chunk)): yield self._prepend_repeated_headers(text, html, is_first_chunk) is_first_chunk = False accum = _RowAccumulator( maxlen=self._maxlen(is_first_chunk), measure=self._opts.measure ) for text, html in accum.flush(): yield self._prepend_repeated_headers(text, html, is_first_chunk) is_first_chunk = False def _iter_row_splits(self, row: HtmlRow, maxlen: int) -> Iterator[TextAndHtml]: """Split oversized row into (text, html) pairs containing as many cells as will fit.""" accum = _CellAccumulator(maxlen=maxlen) for cell in row.iter_cells(): # -- if cell won't fit, flush and check again -- if not accum.will_fit(cell): yield from accum.flush() # -- if cell fits, add it to accumulator -- if accum.will_fit(cell): accum.add_cell(cell) else: # -- otherwise, single cell is bigger than chunking window -- yield from self._iter_cell_splits(cell, maxlen=maxlen) yield from accum.flush() def _iter_cell_splits(self, cell: HtmlCell, maxlen: int) -> Iterator[TextAndHtml]: """Split a single oversized cell into sub-sub-sub-table HTML fragments.""" # -- 33 is len("
"), HTML overhead beyond text content -- # -- For token-based chunking, we subtract 33 chars worth of overhead but still use tokens # -- for the actual content limit. For character-based, we use the reduced character limit. if self._opts.use_token_counting: # -- In token mode, keep token limit but account for HTML overhead in char terms -- # -- The HTML tags themselves are usually ~10-15 tokens, so we reduce by a small amount opts = ChunkingOptions( max_tokens=max(1, maxlen - 10), tokenizer=self._opts._kwargs.get("tokenizer"), ) else: opts = ChunkingOptions(max_characters=max(1, maxlen - 33)) split = _TextSplitter(opts) text, remainder = split(cell.text) yield text, f"
{text}
" # -- an oversized cell will have a remainder, split that up into additional chunks. while remainder: text, remainder = split(remainder) yield text, f"
{text}
" @cached_property def _header_text(self) -> str: """Concatenated text for leading header rows identified by caller.""" return " ".join(text for row in self._header_rows for text in row.iter_cell_texts()) @cached_property def _header_rows(self) -> tuple[HtmlRow, ...]: """Leading rows that should be repeated on continuation chunks, if any.""" if self._header_row_count <= 0: return () rows: list[HtmlRow] = [] for idx, row in enumerate(self._table_element.iter_rows()): if idx >= self._header_row_count: break rows.append(row) return tuple(rows) @cached_property def _header_rows_html(self) -> str: """HTML for repeated header rows, preserving header semantics.""" if not self._header_rows: return "" rows_html = "".join(self._as_header_row_html(row) for row in self._header_rows) return f"{rows_html}" @cached_property def carried_over_header_row_count(self) -> int: """Header-row count prepended to each continuation chunk, or 0 when disabled.""" return len(self._header_rows) if self._should_repeat_headers else 0 @cached_property def _should_repeat_headers(self) -> bool: """True when header repetition is enabled and not pathologically expensive.""" if not self._header_rows: return False # -- guard against pathological headers where a single repeated header row would consume # -- more than half the chunking window. return self._max_header_row_len <= (self._opts.hard_max + 1) // 2 @cached_property def _max_header_row_len(self) -> int: """Largest leading-header row text length.""" if not self._header_rows: return 0 return max(self._opts.measure(" ".join(row.iter_cell_texts())) for row in self._header_rows) @cached_property def _header_text_len(self) -> int: """Size of repeated header text in chunking units.""" return self._opts.measure(self._header_text) def _maxlen(self, is_first_chunk: bool) -> int: """Available size for non-header row content of the next chunk.""" if is_first_chunk or not self._should_repeat_headers: return self._opts.hard_max # -- reserve one separator between repeated header text and chunk body text -- return max(1, self._opts.hard_max - self._header_text_len - 1) def _prepend_repeated_headers(self, text: str, html: str, is_first_chunk: bool) -> TextAndHtml: """Prepend repeated header rows to continuation chunk when enabled.""" if is_first_chunk or not self._should_repeat_headers: return text, html header_text = self._header_text chunk_text = f"{header_text} {text}" if header_text and text else (header_text or text) html_inner = html.removeprefix("").removesuffix("
") chunk_html = f"{self._header_rows_html}{html_inner}
" return chunk_text, chunk_html @staticmethod def _as_header_row_html(row: HtmlRow) -> str: """Serialize `row` preserving source HTML while converting direct-child `` to ``.""" row_html = row.source_html or row.html tr = _HtmlTableSplitter._parse_row_fragment(row_html) if tr is None and row.source_html: tr = _HtmlTableSplitter._parse_row_fragment(row.html) if tr is None: return row.html for cell in tr: if getattr(cell, "tag", None) == "td": cell.tag = "th" return tostring(tr, encoding=str) @staticmethod def _parse_row_fragment(row_html: str): """Parse `row_html` and return a `` element when recoverable.""" try: parsed = fragment_fromstring(row_html) except (ParserError, ValueError): return None if parsed.tag == "tr": return parsed rows = parsed.xpath(".//tr") return rows[0] if rows else None class _TextSplitter: """Provides a text-splitting function configured on construction. Text is split on the best-available separator, falling-back from the preferred separator through a sequence of alternate separators. - The separator is removed by splitting so only whitespace strings are suitable separators. - A "blank-line" ("\n\n") is unlikely to occur in an element as it would have been used as an element boundary during partitioning. This is a *callable* object. Constructing it essentially produces a function: split = _TextSplitter(opts) fragment, remainder = split(s) This allows it to be configured with length-options etc. on construction and used throughout a chunking operation on a given element-stream. """ def __init__(self, opts: ChunkingOptions): self._opts = opts def __call__(self, s: str) -> tuple[str, str]: """Return pair of strings split from `s` on the best match of configured patterns. The first string is the split, the second is the remainder of the string. The split string will never be longer than `maxlen` (in characters or tokens depending on mode). The separators are tried in order until a match is found. The last separator is "" which matches between any two characters so there will always be a split. The separator is removed and does not appear in the split or remainder. An `s` that is already less than the maximum length is returned unchanged with no remainder. This allows this function to be called repeatedly with the remainder until it is consumed and returns a remainder of "". """ maxlen = self._opts.hard_max # -- for token counting, use the measurement abstraction for size check -- if self._opts.use_token_counting: if self._opts.measure(s) <= maxlen: return s, "" return self._split_by_tokens(s) # -- character-based splitting (original logic) -- if len(s) <= maxlen: return s, "" for p, sep_len in self._patterns: # -- length of separator must be added to include that separator when it happens to be # -- located exactly at maxlen. Otherwise the search-from-end regex won't find it. fragment, remainder = self._split_from_maxlen(p, sep_len, s) if ( # -- no available split with this separator -- not fragment # -- split did not progress, consuming part of the string -- or len(remainder) >= len(s) ): continue return fragment.rstrip(), remainder.lstrip() # -- the terminal "" pattern is not actually executed via regex since its implementation is # -- trivial and provides a hard back-stop here in this method. No separator is used between # -- tail and remainder on arb-char split. return s[:maxlen].rstrip(), s[maxlen - self._opts.overlap :].lstrip() def _split_by_tokens(self, s: str) -> tuple[str, str]: """Split text `s` on a separator boundary while respecting token limits. Tries each separator in order of preference, looking for the rightmost split position that keeps the fragment under the token limit. Falls back to splitting on whitespace boundaries if no separator works. """ maxlen = self._opts.hard_max overlap = self._opts.overlap measure = self._opts.measure # -- try each separator in order of preference -- for pattern, _ in self._patterns: # -- find all matches of this separator in the string -- # -- note: (?r) flag makes finditer return matches right-to-left -- matches = list(pattern.finditer(s)) # -- iterate through matches (already right-to-left due to (?r) flag) -- for match in matches: match_start, match_end = match.span() fragment = s[:match_start].rstrip() # -- check if fragment fits within token limit -- if measure(fragment) <= maxlen: # -- skip if fragment is too short (needs at least some content) -- if measure(fragment) == 0: continue raw_remainder = s[match_end:].lstrip() # -- add overlap if configured -- if overlap > 0: # -- token-based overlap: find tail with ~overlap tokens -- tail = self._get_token_overlap_tail(fragment, overlap) overlapped_remainder = tail + " " + raw_remainder return fragment, overlapped_remainder return fragment, raw_remainder # -- fallback: split on whitespace boundary using binary search to find token limit -- # -- find the approximate character position that corresponds to maxlen tokens -- low, high = 0, len(s) best_pos = max(overlap + 1, 1) # -- minimum viable position -- while low <= high: mid = (low + high) // 2 if measure(s[:mid]) <= maxlen: best_pos = mid low = mid + 1 else: high = mid - 1 # -- try to find a whitespace boundary near best_pos, searching backwards -- split_pos = best_pos for i in range(best_pos, max(overlap, 0), -1): if i < len(s) and s[i].isspace(): split_pos = i break # -- ensure the fragment still fits after whitespace adjustment -- fragment = s[:split_pos].rstrip() if measure(fragment) > maxlen and split_pos > overlap + 1: # -- whitespace boundary pushed us over; use the binary search result directly -- fragment = s[:best_pos].rstrip() split_pos = best_pos raw_remainder = s[split_pos:].lstrip() if overlap > 0 and fragment: tail = self._get_token_overlap_tail(fragment, overlap) overlapped_remainder = tail + " " + raw_remainder return fragment, overlapped_remainder return fragment, raw_remainder def _get_token_overlap_tail(self, text: str, target_tokens: int) -> str: """Extract tail of text containing approximately `target_tokens` tokens. Uses binary search to find the character position from which the tail contains approximately the specified number of tokens. Adjusts to word boundaries to avoid splitting words. """ measure = self._opts.measure # -- if the entire text has fewer tokens than target, return all of it -- if measure(text) <= target_tokens: return text.strip() # -- binary search to find the character position that yields ~target_tokens -- low, high = 0, len(text) while low < high: mid = (low + high) // 2 tail = text[mid:] token_count = measure(tail) if token_count > target_tokens: low = mid + 1 else: high = mid # -- adjust to word boundary: search forward for whitespace then skip it -- pos = low while pos < len(text) and not text[pos].isspace(): pos += 1 while pos < len(text) and text[pos].isspace(): pos += 1 # -- if we've moved too far, fall back to just stripping leading whitespace -- if pos >= len(text): return text[low:].lstrip() return text[pos:] @cached_property def _patterns(self) -> tuple[tuple[regex.Pattern[str], int], ...]: """Sequence of (pattern, len) pairs to match against. Patterns appear in order of preference, those following are "fall-back" patterns to be used if no match of a prior pattern is found. NOTE these regexes search *from the end of the string*, which is what the "(?r)" bit specifies. This is much more efficient than starting at the beginning of the string which could result in hundreds of matches before the desired one. """ separators = self._opts.text_splitting_separators return tuple((regex.compile(f"(?r){sep}"), len(sep)) for sep in separators) def _split_from_maxlen( self, pattern: regex.Pattern[str], sep_len: int, s: str ) -> tuple[str, str]: """Return (split, remainder) pair split from `s` on the right-most match before `maxlen`. Returns `"", s` if no suitable match was found. Also returns `"", s` if splitting on this separator produces a split shorter than the required overlap (which would produce an infinite loop). `split` will never be longer than `maxlen` and there is no longer split available using `pattern`. The separator is removed and does not appear in either the split or remainder. """ maxlen, overlap = self._opts.hard_max, self._opts.overlap # -- A split not longer than overlap will not progress (infinite loop). On the right side, # -- need to extend search range to include a separator located exactly at maxlen. match = pattern.search(s, pos=overlap + 1, endpos=maxlen + sep_len) if match is None: return "", s # -- characterize match location match_start, match_end = match.span() # -- matched separator is replaced by single-space in overlap string -- separator = " " # -- in multi-space situation, fragment may have trailing whitespace because match is from # -- right to left fragment = s[:match_start].rstrip() # -- remainder can have leading space when match is on "\n" followed by spaces -- raw_remainder = s[match_end:].lstrip() if overlap <= len(separator): return fragment, raw_remainder # -- compute overlap -- tail_len = overlap - len(separator) tail = fragment[-tail_len:].lstrip() overlapped_remainder = tail + separator + raw_remainder return fragment, overlapped_remainder class _CellAccumulator: """Incrementally build `` fragment cell-by-cell to maximally fill chunking window. Accumulate cells until chunking window is filled, then generate the text and HTML for the subtable composed of all those rows that fit in the window. """ def __init__(self, maxlen: int): self._maxlen = maxlen self._cells: list[HtmlCell] = [] def add_cell(self, cell: HtmlCell) -> None: """Add `cell` to this accumulation. Caller is responsible for ensuring it will fit.""" self._cells.append(cell) def flush(self) -> Iterator[TextAndHtml]: """Generate zero-or-one (text, html) pairs for accumulated sub-sub-table.""" if not self._cells: return text = " ".join(self._iter_cell_texts()) tds_str = "".join(c.html for c in self._cells) html = f"
{tds_str}
" self._cells.clear() yield text, html def will_fit(self, cell: HtmlCell) -> bool: """True when `cell` will fit within remaining space left by accummulated cells.""" return self._remaining_space >= len(cell.text) def _iter_cell_texts(self) -> Iterator[str]: """Generate contents of each accumulated cell as a separate string. A cell that is empty or contains only whitespace does not generate a string. """ for cell in self._cells: if not (text := cell.text): continue yield text @property def _remaining_space(self) -> int: """Number of characters remaining when text of accumulated cells is joined.""" # -- separators are one space (" ") at the end of each cell's text, including last one to # -- account for space before prospective next cell. separators_len = len(self._cells) return self._maxlen - separators_len - sum(len(c.text) for c in self._cells) class _RowAccumulator: """Maybe `SubtableAccumulator`. Accumulate rows until chunking window is filled, then generate the text and HTML for the subtable composed of all those rows that fit in the window. """ def __init__(self, maxlen: int, measure: Callable[[str], int] = len): self._maxlen = maxlen self._measure = measure self._rows: list[HtmlRow] = [] self._row_text_len = 0 def add_row(self, row: HtmlRow) -> None: """Add `row` to this accumulation. Caller is responsible for ensuring it will fit.""" self._rows.append(row) self._row_text_len += self._measured_row_text_len(row) def flush(self) -> Iterator[TextAndHtml]: """Generate zero-or-one (text, html) pairs for accumulated sub-table.""" if not self._rows: return text = " ".join(self._iter_cell_texts()) trs_str = "".join(r.html for r in self._rows) html = f"{trs_str}
" self._rows.clear() self._row_text_len = 0 yield text, html def will_fit(self, row: HtmlRow) -> bool: """True when `row` will fit within remaining space left by accummulated rows.""" return self._remaining_space >= self._measured_row_text_len(row) def _iter_cell_texts(self) -> Iterator[str]: """Generate contents of each row cell as a separate string. A cell that is empty or contains only whitespace does not generate a string. """ for r in self._rows: yield from r.iter_cell_texts() @property def _remaining_space(self) -> int: """Number of chunk-size units remaining for accumulated row text.""" # -- separators are one space (" ") at the end of each row's text, including last one to # -- account for space before prospective next row. separators_len = len(self._rows) return self._maxlen - separators_len - self._row_text_len def _measured_row_text_len(self, row: HtmlRow) -> int: """Length of `row` text in configured chunk-size units.""" return self._measure(" ".join(row.iter_cell_texts())) # ================================================================================================ # PRE-CHUNK COMBINER # ================================================================================================ class PreChunkCombiner: """Filters pre-chunk stream to combine small pre-chunks where possible.""" def __init__(self, pre_chunks: Iterable[PreChunk], opts: ChunkingOptions): self._pre_chunks = pre_chunks self._opts = opts def iter_combined_pre_chunks(self) -> Iterator[PreChunk]: """Generate pre-chunk objects, combining `PreChunk` objects when they'll fit in window.""" accum = _PreChunkAccumulator(self._opts) for pre_chunk in self._pre_chunks: # -- finish accumulating pre-chunk when it's full -- if not accum.will_fit(pre_chunk): yield from accum.flush() accum.add_pre_chunk(pre_chunk) yield from accum.flush() class _PreChunkAccumulator: """Accumulates, measures, and combines pre-chunks. Used for combining pre-chunks for chunking strategies like "by-title" that can potentially produce undersized chunks and offer the `combine_text_under_n_chars` option. Provides `.add_pre_chunk()` allowing a pre-chunk to be added to the chunk and provides monitoring properties `.remaining_space` and `.text_length` suitable for deciding whether to add another pre-chunk. `.flush()` is used to combine the accumulated pre-chunks into a single `PreChunk` object. This method returns an interator that generates zero-or-one `PreChunk` objects and is used like so: yield from accum.flush() If no pre-chunks have been accumulated, no `PreChunk` is generated. Flushing the builder clears the pre-chunks it contains so it is ready to accept the next pre-chunk. """ def __init__(self, opts: ChunkingOptions) -> None: self._opts = opts self._pre_chunk: PreChunk | None = None def add_pre_chunk(self, pre_chunk: PreChunk) -> None: """Add a pre-chunk to the accumulator for possible combination with next pre-chunk.""" self._pre_chunk = ( pre_chunk if self._pre_chunk is None else self._pre_chunk.combine(pre_chunk) ) def flush(self) -> Iterator[PreChunk]: """Generate accumulated pre-chunk as a single combined pre-chunk. Does not generate a pre-chunk when none has been accumulated. """ # -- nothing to do if no pre-chunk has been accumulated -- if not self._pre_chunk: return # -- otherwise generate the combined pre-chunk -- yield self._pre_chunk # -- and reset the accumulator (to empty) -- self._pre_chunk = None def will_fit(self, pre_chunk: PreChunk) -> bool: """True when there is room for `pre_chunk` in accumulator. An empty accumulator always has room. Otherwise there is only room when `pre_chunk` can be combined with any other pre-chunks in the accumulator without exceeding the combination limits specified for the chunking run. """ # -- an empty accumulator always has room -- if self._pre_chunk is None: return True return self._pre_chunk.can_combine(pre_chunk) # ================================================================================================ # CHUNK BOUNDARY PREDICATES # ------------------------------------------------------------------------------------------------ # A *boundary predicate* is a function that takes an element and returns True when the element # represents the start of a new semantic boundary (such as section or page) to be respected in # chunking. # # Some of the functions below *are* a boundary predicate and others *construct* a boundary # predicate. # # These can be mixed and matched to produce different chunking behaviors like "by_title" or left # out altogether to produce "basic-chunking" behavior. # # The effective lifetime of the function that produce a predicate (rather than directly being one) # is limited to a single element-stream because these retain state (e.g. current page number) to # determine when a semantic boundary has been crossed. # ================================================================================================ def is_on_next_page() -> BoundaryPredicate: """Not a predicate itself, calling this returns a predicate that triggers on each new page. The lifetime of the returned callable cannot extend beyond a single element-stream because it stores current state (current page-number) that is particular to that element stream. The returned predicate tracks the "current" page-number, starting at 1. An element with a greater page number returns True, indicating the element starts a new page boundary, and updates the enclosed page-number ready for the next transition. An element with `page_number == None` or a page-number lower than the stored value is ignored and returns False. """ current_page_number: int = 1 is_first: bool = True def page_number_incremented(element: Element) -> bool: nonlocal current_page_number, is_first page_number = element.metadata.page_number # -- The first element never reports a page break, it starts the first page of the # -- document. That page could be numbered (page_number is non-None) or not. If it is not # -- numbered we assign it page-number 1. if is_first: current_page_number = page_number or 1 is_first = False return False # -- An element with a `None` page-number is assumed to continue the current page. It never # -- updates the current-page-number because once set, the current-page-number is "sticky" # -- until replaced by a different explicit page-number. if page_number is None: return False if page_number == current_page_number: return False # -- it's possible for a page-number to decrease. We don't expect that, but if it happens # -- we consider it a page-break. current_page_number = page_number return True return page_number_incremented def is_title(element: Element) -> bool: """True when `element` is a `Title` element, False otherwise.""" return isinstance(element, Title)