feat: PreProcessor split by token (tiktoken & Hugging Face) (#5276)

* #4983 implemented split by token for tiktoken tokenizer

* #4983 added unit test for tiktoken splitting

* #4983 implemented and added a test for splitting documents with HuggingFace tokenizer

* #4983 added support for passing HF model names (instead of objects) and added an example to the HF token splitting test

* mocked HTTP model loading in unit tests, fixed pylint error

* fix lossy tokenizers splitting, use LazyImport, ignore UnicodeEncodeError for tiktoken

* reno

* rename reno file

---------

Co-authored-by: Stefano Fiorucci <44616784+anakin87@users.noreply.github.com>
Co-authored-by: ZanSara <sara.zanzottera@deepset.ai>

haystack/nodes/preprocessor/base.py

@@ -2,6 +2,8 @@
 
 from abc import abstractmethod
 
+from transformers import PreTrainedTokenizerBase
+
 from haystack.nodes.base import BaseComponent
 from haystack.schema import Document
 
@@ -17,10 +19,11 @@ def process(
         clean_header_footer: Optional[bool] = False,
         clean_empty_lines: Optional[bool] = True,
         remove_substrings: Optional[List[str]] = None,
-        split_by: Literal["word", "sentence", "passage", None] = "word",
+        split_by: Literal["token", "word", "sentence", "passage", None] = "word",
         split_length: Optional[int] = 1000,
         split_overlap: Optional[int] = None,
         split_respect_sentence_boundary: Optional[bool] = True,
+        tokenizer: Optional[Union[str, PreTrainedTokenizerBase]] = "tiktoken",
         id_hash_keys: Optional[List[str]] = None,
     ) -> List[Document]:
         """
@@ -44,10 +47,11 @@ def clean(
     def split(
         self,
         document: Union[dict, Document],
-        split_by: Literal["word", "sentence", "passage", None],
+        split_by: Literal["token", "word", "sentence", "passage", None],
         split_length: int,
         split_overlap: int,
         split_respect_sentence_boundary: bool,
+        tokenizer: Optional[Union[str, PreTrainedTokenizerBase]] = None,
     ) -> List[Document]:
         raise NotImplementedError
 
@@ -57,10 +61,11 @@ def run(  # type: ignore
         clean_whitespace: Optional[bool] = None,
         clean_header_footer: Optional[bool] = None,
         clean_empty_lines: Optional[bool] = None,
-        split_by: Literal["word", "sentence", "passage", None] = None,
+        split_by: Literal["token", "word", "sentence", "passage", None] = None,
         split_length: Optional[int] = None,
         split_overlap: Optional[int] = None,
         split_respect_sentence_boundary: Optional[bool] = None,
+        tokenizer: Optional[Union[str, PreTrainedTokenizerBase]] = None,
         id_hash_keys: Optional[List[str]] = None,
     ):
         processed_documents = self.process(
@@ -83,10 +88,11 @@ def run_batch(  # type: ignore
         clean_whitespace: Optional[bool] = None,
         clean_header_footer: Optional[bool] = None,
         clean_empty_lines: Optional[bool] = None,
-        split_by: Literal["word", "sentence", "passage", None] = None,
+        split_by: Literal["token", "word", "sentence", "passage", None] = None,
         split_length: Optional[int] = None,
         split_overlap: Optional[int] = None,
         split_respect_sentence_boundary: Optional[bool] = None,
+        tokenizer: Optional[Union[str, PreTrainedTokenizerBase]] = None,
         id_hash_keys: Optional[List[str]] = None,
     ):
         return self.run(

haystack/nodes/preprocessor/preprocessor.py

@@ -1,4 +1,4 @@
-from typing import List, Optional, Generator, Set, Union, Tuple, Dict, Literal
+from typing import List, Optional, Generator, Set, Union, Tuple, Dict, Literal, Callable, Any
 
 import logging
 import re
@@ -17,14 +17,18 @@
 from haystack.schema import Document
 from haystack.lazy_imports import LazyImport
 
+with LazyImport("Run 'pip install transformers'") as transformers_import:
+    from transformers import PreTrainedTokenizerBase
+    from transformers import AutoTokenizer
 
-logger = logging.getLogger(__name__)
+with LazyImport("Run 'pip install tiktoken'") as tiktoken_import:
+    import tiktoken
 
+logger = logging.getLogger(__name__)
 
 with LazyImport("Run 'pip install farm-haystack[preprocessing]' or 'pip install nltk'") as nltk_import:
     import nltk
 
-
 iso639_to_nltk = {
     "ru": "russian",
     "sl": "slovene",
@@ -55,11 +59,12 @@ def __init__(
         clean_header_footer: bool = False,
         clean_empty_lines: bool = True,
         remove_substrings: Optional[List[str]] = None,
-        split_by: Optional[Literal["word", "sentence", "passage"]] = "word",
+        split_by: Optional[Literal["token", "word", "sentence", "passage"]] = "word",
         split_length: int = 200,
         split_overlap: int = 0,
         split_respect_sentence_boundary: bool = True,
         tokenizer_model_folder: Optional[Union[str, Path]] = None,
+        tokenizer: Optional[Union[str, PreTrainedTokenizerBase]] = "tiktoken",
         language: str = "en",
         id_hash_keys: Optional[List[str]] = None,
         progress_bar: bool = True,
@@ -86,6 +91,9 @@ def __init__(
         :param split_respect_sentence_boundary: Whether to split in partial sentences if split_by -> `word`. If set
                                                 to True, the individual split will always have complete sentences &
                                                 the number of words will be <= split_length.
+        :param tokenizer: Specifies the tokenizer to use if split_by="token". Supported options are "tiktoken"
+                            (for OpenAI's GPT-3.5 and GPT-4) and any HuggingFace tokenizer (e.g. 'bert-base-uncased').
+                            HuggingFace tokenizers can also be passed directly as an PreTrainedTokenizerBase object.
         :param language: The language used by "nltk.tokenize.sent_tokenize" in iso639 format.
             Available options: "ru","sl","es","sv","tr","cs","da","nl","en","et","fi","fr","de","el","it","no","pl","pt","ml"
         :param tokenizer_model_folder: Path to the folder containing the NTLK PunktSentenceTokenizer models, if loading a model from a local path. Leave empty otherwise.
@@ -124,6 +132,7 @@ def __init__(
         self.split_length = split_length
         self.split_overlap = split_overlap
         self.split_respect_sentence_boundary = split_respect_sentence_boundary
+        self.tokenizer = tokenizer
         self.language = language
         self.tokenizer_model_folder = tokenizer_model_folder
         self.print_log: Set[str] = set()
@@ -139,10 +148,11 @@ def process(
         clean_header_footer: Optional[bool] = None,
         clean_empty_lines: Optional[bool] = None,
         remove_substrings: Optional[List[str]] = None,
-        split_by: Optional[Literal["word", "sentence", "passage"]] = None,
+        split_by: Optional[Literal["token", "word", "sentence", "passage"]] = None,
         split_length: Optional[int] = None,
         split_overlap: Optional[int] = None,
         split_respect_sentence_boundary: Optional[bool] = None,
+        tokenizer: Optional[Union[str, PreTrainedTokenizerBase]] = None,
         id_hash_keys: Optional[List[str]] = None,
     ) -> List[Document]:
         """
@@ -167,6 +177,7 @@ def process(
             "split_length": split_length,
             "split_overlap": split_overlap,
             "split_respect_sentence_boundary": split_respect_sentence_boundary,
+            "tokenizer": tokenizer,
         }
 
         if id_hash_keys is None:
@@ -219,10 +230,11 @@ def _process_single(
         clean_header_footer: Optional[bool] = None,
         clean_empty_lines: Optional[bool] = None,
         remove_substrings: Optional[List[str]] = None,
-        split_by: Optional[Literal["word", "sentence", "passage"]] = None,
+        split_by: Optional[Literal["token", "word", "sentence", "passage"]] = None,
         split_length: Optional[int] = None,
         split_overlap: Optional[int] = None,
         split_respect_sentence_boundary: Optional[bool] = None,
+        tokenizer: Optional[Union[str, PreTrainedTokenizerBase]] = None,
         id_hash_keys: Optional[List[str]] = None,
     ) -> List[Document]:
         if remove_substrings is None:
@@ -243,6 +255,8 @@ def _process_single(
             split_overlap = self.split_overlap
         if split_respect_sentence_boundary is None:
             split_respect_sentence_boundary = self.split_respect_sentence_boundary
+        if tokenizer is None:
+            tokenizer = self.tokenizer
 
         cleaned_document = self.clean(
             document=document,
@@ -258,6 +272,7 @@ def _process_single(
             split_length=split_length,
             split_overlap=split_overlap,
             split_respect_sentence_boundary=split_respect_sentence_boundary,
+            tokenizer=tokenizer,
             id_hash_keys=id_hash_keys,
         )
 
@@ -332,10 +347,11 @@ def clean(
     def split(
         self,
         document: Union[dict, Document],
-        split_by: Optional[Literal["word", "sentence", "passage"]],
+        split_by: Optional[Literal["token", "word", "sentence", "passage"]],
         split_length: int,
         split_overlap: int,
         split_respect_sentence_boundary: bool,
+        tokenizer: Optional[Union[str, PreTrainedTokenizerBase]] = None,
         id_hash_keys: Optional[List[str]] = None,
     ) -> List[Document]:
         """Perform document splitting on a single document. This method can split on different units, at different lengths,
@@ -359,8 +375,10 @@ def split(
         if not split_length:
             raise Exception("split_length needs be set when using split_by.")
 
-        if split_respect_sentence_boundary and split_by != "word":
-            raise NotImplementedError("'split_respect_sentence_boundary=True' is only compatible with split_by='word'.")
+        if split_respect_sentence_boundary and split_by not in ["word", "token"]:
+            raise NotImplementedError(
+                "'split_respect_sentence_boundary=True' is only compatible with split_by='word' or 'token'."
+            )
 
         if type(document.content) is not str:
             logger.error("Document content is not of type str. Nothing to split.")
@@ -369,13 +387,17 @@ def split(
         text = document.content
         headlines = document.meta["headlines"] if "headlines" in document.meta else []
 
-        if split_respect_sentence_boundary and split_by == "word":
-            text_splits, splits_pages, splits_start_idxs = self._split_by_word_respecting_sent_boundary(
-                text=text, split_length=split_length, split_overlap=split_overlap
+        if split_respect_sentence_boundary and split_by in ["word", "token"]:
+
+            def split_function(text):
+                return self._split_tokens(text, tokenizer=tokenizer) if split_by == "token" else text.split()
+
+            text_splits, splits_pages, splits_start_idxs = self._split_into_units_respecting_sent_boundary(
+                text=text, split_length=split_length, split_overlap=split_overlap, split_function=split_function
             )
         else:
             # create individual "elements" of passage, sentence, or word
-            elements, split_at = self._split_into_units(text=text, split_by=split_by)
+            elements, split_at = self._split_into_units(text=text, split_by=split_by, tokenizer=tokenizer)
 
             # concatenate individual elements based on split_length & split_stride
             text_splits, splits_pages, splits_start_idxs = self._concatenate_units(
@@ -467,47 +489,47 @@ def _remove_substring(text: str, substring: str, headlines: List[Dict]) -> Tuple
         cleaned_text = text.replace(substring, "")
         return cleaned_text, headlines
 
-    def _split_by_word_respecting_sent_boundary(
-        self, text: str, split_length: int, split_overlap: int
+    def _split_into_units_respecting_sent_boundary(
+        self, text: str, split_length: int, split_overlap: int, split_function: Callable
     ) -> Tuple[List[str], List[int], List[int]]:
         """
         Splits the text into parts of split_length words while respecting sentence boundaries.
         """
         sentences = self._split_sentences(text)
 
-        word_count_slice = 0
+        unit_count_slice = 0
         cur_page = 1
         cur_start_idx = 0
         splits_pages = []
         list_splits = []
         splits_start_idxs = []
         current_slice: List[str] = []
         for sen in sentences:
-            word_count_sen = len(sen.split())
+            unit_count_sen = len(split_function(sen))
 
-            if word_count_sen > split_length:
+            if unit_count_sen > split_length:
                 long_sentence_message = (
-                    "We found one or more sentences whose word count is higher than the split length."
+                    "We found one or more sentences whose split count is higher than the split length."
                 )
                 if long_sentence_message not in self.print_log:
                     self.print_log.add(long_sentence_message)
                     logger.warning(long_sentence_message)
 
-            if word_count_slice + word_count_sen > split_length:
+            if unit_count_slice + unit_count_sen > split_length:
                 # Number of words exceeds split_length -> save current slice and start a new one
                 if current_slice:
                     list_splits.append(current_slice)
                     splits_pages.append(cur_page)
                     splits_start_idxs.append(cur_start_idx)
 
                 if split_overlap:
-                    processed_sents, current_slice, word_count_slice = self._get_overlap_from_slice(
-                        current_slice, split_length, split_overlap
+                    processed_sents, current_slice, unit_count_slice = self._get_overlap_from_slice(
+                        current_slice, split_length, split_overlap, split_function
                     )
                 else:
                     processed_sents = current_slice
                     current_slice = []
-                    word_count_slice = 0
+                    unit_count_slice = 0
 
                 cur_start_idx += len("".join(processed_sents))
 
@@ -522,7 +544,7 @@ def _split_by_word_respecting_sent_boundary(
                     cur_page += num_page_breaks
 
             current_slice.append(sen)
-            word_count_slice += word_count_sen
+            unit_count_slice += unit_count_sen
 
         if current_slice:
             list_splits.append(current_slice)
@@ -539,7 +561,7 @@ def _split_by_word_respecting_sent_boundary(
 
     @staticmethod
     def _get_overlap_from_slice(
-        current_slice: List[str], split_length: int, split_overlap: int
+        current_slice: List[str], split_length: int, split_overlap: int, split_function: Callable
     ) -> Tuple[List[str], List[str], int]:
         """
         Returns a tuple with the following elements:
@@ -553,7 +575,7 @@ def _get_overlap_from_slice(
         current_slice_copy = deepcopy(current_slice)
         # Next overlapping Document should not start exactly the same as the previous one, so we skip the first sentence
         for idx, s in reversed(list(enumerate(current_slice))[1:]):
-            sen_len = len(s.split())
+            sen_len = len(split_function(s))
             if word_count_overlap < split_overlap and sen_len < split_length:
                 overlap.append(s)
                 word_count_overlap += sen_len
@@ -566,7 +588,7 @@ def _get_overlap_from_slice(
 
         return processed_sents, next_slice, word_count_slice
 
-    def _split_into_units(self, text: str, split_by: str) -> Tuple[List[str], str]:
+    def _split_into_units(self, text: str, split_by: str, tokenizer: Any) -> Tuple[List[str], str]:
         if split_by == "passage":
             elements = text.split("\n\n")
             split_at = "\n\n"
@@ -576,8 +598,13 @@ def _split_into_units(self, text: str, split_by: str) -> Tuple[List[str], str]:
         elif split_by == "word":
             elements = text.split(" ")
             split_at = " "
+        elif split_by == "token":
+            elements = self._split_tokens(text, tokenizer)
+            split_at = ""
         else:
-            raise NotImplementedError("PreProcessor only supports 'passage', 'sentence' or 'word' split_by options.")
+            raise NotImplementedError(
+                "PreProcessor only supports 'passage', 'sentence', 'word' or 'token' split_by options."
+            )
 
         return elements, split_at
 
@@ -823,6 +850,35 @@ def _split_sentences(self, text: str) -> List[str]:
         sentences = sentence_tokenizer.tokenize(text)
         return sentences
 
+    def _split_tokens(self, text: str, tokenizer: Any) -> List[str]:
+        if tokenizer == "tiktoken":
+            tiktoken_import.check()
+            enc = tiktoken.get_encoding("cl100k_base")  # tiktoken is reversible and lossless
+            integer_tokens = enc.encode(text, disallowed_special=())
+            elements = [enc.decode_single_token_bytes(token).decode(errors="ignore") for token in integer_tokens]
+            return elements
+        if isinstance(tokenizer, str):
+            transformers_import.check()
+            try:
+                tokenizer = AutoTokenizer.from_pretrained(tokenizer)
+            except Exception:
+                raise ValueError(
+                    f"Could not load tokenizer '{tokenizer}' from HuggingFace model hub. "
+                    f"Please make sure that the tokenizer is correct and exists."
+                )
+        if isinstance(tokenizer, PreTrainedTokenizerBase):
+            encoded = tokenizer.encode_plus(text, return_offsets_mapping=True, add_special_tokens=False)
+            elements = []
+            for i in range(l := len(encoded.offset_mapping)):
+                start_current = encoded.offset_mapping[i][0]
+                start_next = encoded.offset_mapping[i + 1][0] if i < l - 1 else len(text)
+                elements.append(text[start_current:start_next])
+            return elements
+        raise ValueError(
+            f"Unsupported tokenizer specification {tokenizer}. "
+            f"Please provide either the string 'tiktoken' or a HuggingFace tokenizer (PreTrainedTokenizerBase)."
+        )
+
     def _load_sentence_tokenizer(self, language_name: Optional[str]) -> "nltk.tokenize.punkt.PunktSentenceTokenizer":
         # Try to load a custom model from 'tokenizer_model_path'
         if self.tokenizer_model_folder is not None:

releasenotes/notes/split-by-token-b9a4f954d4077ecc.yaml

@@ -0,0 +1,2 @@
+features:
+  - Add `split_length` by token in PreProcessor