text-splitters: Inconsistent results with NLTKTextSplitter's add_start_index=True (#27782)

This PR closes #27781

# Problem
The current implementation of `NLTKTextSplitter` is using
`sent_tokenize`. However, this `sent_tokenize` doesn't handle chars
between 2 tokenized sentences... hence, this behavior throws errors when
we are using `add_start_index=True`, as described in issue #27781. In
particular:
```python
from nltk.tokenize import sent_tokenize

output1 = sent_tokenize("Innovation drives our success. Collaboration fosters creative solutions. Efficiency enhances data management.", language="english")
print(output1)
output2 = sent_tokenize("Innovation drives our success.        Collaboration fosters creative solutions. Efficiency enhances data management.", language="english")
print(output2)
>>> ['Innovation drives our success.', 'Collaboration fosters creative solutions.', 'Efficiency enhances data management.']
>>> ['Innovation drives our success.', 'Collaboration fosters creative solutions.', 'Efficiency enhances data management.']
```

# Solution
With this new `use_span_tokenize` parameter, we can use NLTK to create
sentences (with `span_tokenize`), but also add extra chars to be sure
that we still can map the chunks to the original text.

---------

Co-authored-by: Erick Friis <erick@langchain.dev>
Co-authored-by: Erick Friis <erickfriis@gmail.com>

libs/text-splitters/Makefile

@@ -9,6 +9,9 @@ TEST_FILE ?= tests/unit_tests/
 test tests:
 	poetry run pytest --disable-socket --allow-unix-socket $(TEST_FILE)
 
+integration_test integration_tests: 
+	poetry run pytest tests/integration_tests/
+
 test_watch:
 	poetry run ptw --snapshot-update --now . -- -vv -x tests/unit_tests
 

libs/text-splitters/langchain_text_splitters/nltk.py

@@ -9,11 +9,26 @@ class NLTKTextSplitter(TextSplitter):
     """Splitting text using NLTK package."""
 
     def __init__(
-        self, separator: str = "\n\n", language: str = "english", **kwargs: Any
+        self,
+        separator: str = "\n\n",
+        language: str = "english",
+        *,
+        use_span_tokenize: bool = False,
+        **kwargs: Any,
     ) -> None:
         """Initialize the NLTK splitter."""
         super().__init__(**kwargs)
+        self._separator = separator
+        self._language = language
+        self._use_span_tokenize = use_span_tokenize
+        if self._use_span_tokenize and self._separator != "":
+            raise ValueError("When use_span_tokenize is True, separator should be ''")
         try:
+            if self._use_span_tokenize:
+                from nltk.tokenize import _get_punkt_tokenizer
+
+                self._tokenizer = _get_punkt_tokenizer(self._language)
+            else:
                 from nltk.tokenize import sent_tokenize
 
                 self._tokenizer = sent_tokenize
@@ -21,11 +36,20 @@ class NLTKTextSplitter(TextSplitter):
             raise ImportError(
                 "NLTK is not installed, please install it with `pip install nltk`."
             )
-        self._separator = separator
-        self._language = language
 
     def split_text(self, text: str) -> List[str]:
         """Split incoming text and return chunks."""
         # First we naively split the large input into a bunch of smaller ones.
+        if self._use_span_tokenize:
+            spans = list(self._tokenizer.span_tokenize(text))
+            splits = []
+            for i, (start, end) in enumerate(spans):
+                if i > 0:
+                    prev_end = spans[i - 1][1]
+                    sentence = text[prev_end:start] + text[start:end]
+                else:
+                    sentence = text[start:end]
+                splits.append(sentence)
+        else:
             splits = self._tokenizer(text, language=self._language)
         return self._merge_splits(splits, self._separator)

libs/text-splitters/langchain_text_splitters/sentence_transformers.py

@@ -22,7 +22,7 @@ class SentenceTransformersTokenTextSplitter(TextSplitter):
             from sentence_transformers import SentenceTransformer
         except ImportError:
             raise ImportError(
-                "Could not import sentence_transformer python package. "
+                "Could not import sentence_transformers python package. "
                 "This is needed in order to for SentenceTransformersTokenTextSplitter. "
                 "Please install it with `pip install sentence-transformers`."
             )

libs/text-splitters/pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = [ "poetry-core>=1.0.0",]
+requires = ["poetry-core>=1.0.0"]
 build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
@@ -14,7 +14,20 @@ repository = "https://github.com/langchain-ai/langchain"
 [tool.mypy]
 disallow_untyped_defs = "True"
 [[tool.mypy.overrides]]
-module = [ "transformers", "sentence_transformers", "nltk.tokenize", "konlpy.tag", "bs4", "pytest", "spacy", "spacy.lang.en", "numpy",]
+module = [
+    "transformers",
+    "sentence_transformers",
+    "nltk.tokenize",
+    "konlpy.tag",
+    "bs4",
+    "pytest",
+    "spacy",
+    "spacy.lang.en",
+    "numpy",
+    "nltk",
+    "spacy.cli",
+    "torch",
+]
 ignore_missing_imports = "True"
 
 [tool.poetry.urls]
@@ -26,15 +39,18 @@ python = ">=3.9,<4.0"
 langchain-core = "^0.3.25"
 
 [tool.ruff.lint]
-select = [ "E", "F", "I", "T201", "D",]
-ignore = [ "D100",]
+select = ["E", "F", "I", "T201", "D"]
+ignore = ["D100"]
 
 [tool.coverage.run]
-omit = [ "tests/*",]
+omit = ["tests/*"]
 
 [tool.pytest.ini_options]
 addopts = "--strict-markers --strict-config --durations=5"
-markers = [ "requires: mark tests as requiring a specific library", "compile: mark placeholder test used to compile integration tests without running them",]
+markers = [
+    "requires: mark tests as requiring a specific library",
+    "compile: mark placeholder test used to compile integration tests without running them",
+]
 asyncio_mode = "auto"
 
 [tool.poetry.group.lint]
@@ -53,19 +69,17 @@ optional = true
 convention = "google"
 
 [tool.ruff.lint.per-file-ignores]
-"tests/**" = [ "D",]
+"tests/**" = ["D"]
 
 [tool.poetry.group.lint.dependencies]
 ruff = "^0.5"
 
-
 [tool.poetry.group.typing.dependencies]
 mypy = "^1.10"
 lxml-stubs = "^0.5.1"
 types-requests = "^2.31.0.20240218"
 tiktoken = "^0.8.0"
 
-
 [tool.poetry.group.dev.dependencies]
 jupyter = "^1.0.0"
 
@@ -78,20 +92,23 @@ pytest-watcher = "^0.3.4"
 pytest-asyncio = "^0.21.1"
 pytest-socket = "^0.7.0"
 
+[tool.poetry.group.test_integration]
+optional = true
 
 [tool.poetry.group.test_integration.dependencies]
-
+spacy = { version = "*", python = "<3.13" }
+nltk = "^3.9.1"
+transformers = "^4.47.0"
+sentence-transformers = { version = ">=2.6.0", python = "<3.13" }
 
 [tool.poetry.group.lint.dependencies.langchain-core]
 path = "../core"
 develop = true
 
-
 [tool.poetry.group.dev.dependencies.langchain-core]
 path = "../core"
 develop = true
 
-
 [tool.poetry.group.test.dependencies.langchain-core]
 path = "../core"
 develop = true

libs/text-splitters/tests/integration_tests/test_nlp_text_splitters.py

@@ -1,18 +1,36 @@
 """Test text splitting functionality using NLTK and Spacy based sentence splitters."""
 
+from typing import Any
+
+import nltk
 import pytest
+from langchain_core.documents import Document
 
 from langchain_text_splitters.nltk import NLTKTextSplitter
 from langchain_text_splitters.spacy import SpacyTextSplitter
 
 
+def setup_module() -> None:
+    nltk.download("punkt_tab")
+
+
+@pytest.fixture()
+def spacy() -> Any:
+    try:
+        import spacy
+    except ImportError:
+        pytest.skip("Spacy not installed.")
+    spacy.cli.download("en_core_web_sm")  # type: ignore
+    return spacy
+
+
 def test_nltk_text_splitting_args() -> None:
     """Test invalid arguments."""
     with pytest.raises(ValueError):
         NLTKTextSplitter(chunk_size=2, chunk_overlap=4)
 
 
-def test_spacy_text_splitting_args() -> None:
+def test_spacy_text_splitting_args(spacy: Any) -> None:
     """Test invalid arguments."""
     with pytest.raises(ValueError):
         SpacyTextSplitter(chunk_size=2, chunk_overlap=4)
@@ -29,7 +47,7 @@ def test_nltk_text_splitter() -> None:
 
 
 @pytest.mark.parametrize("pipeline", ["sentencizer", "en_core_web_sm"])
-def test_spacy_text_splitter(pipeline: str) -> None:
+def test_spacy_text_splitter(pipeline: str, spacy: Any) -> None:
     """Test splitting by sentence using Spacy."""
     text = "This is sentence one. And this is sentence two."
     separator = "|||"
@@ -40,7 +58,7 @@ def test_spacy_text_splitter(pipeline: str) -> None:
 
 
 @pytest.mark.parametrize("pipeline", ["sentencizer", "en_core_web_sm"])
-def test_spacy_text_splitter_strip_whitespace(pipeline: str) -> None:
+def test_spacy_text_splitter_strip_whitespace(pipeline: str, spacy: Any) -> None:
     """Test splitting by sentence using Spacy."""
     text = "This is sentence one. And this is sentence two."
     separator = "|||"
@@ -50,3 +68,35 @@ def test_spacy_text_splitter_strip_whitespace(pipeline: str) -> None:
     output = splitter.split_text(text)
     expected_output = [f"This is sentence one. {separator}And this is sentence two."]
     assert output == expected_output
+
+
+def test_nltk_text_splitter_args() -> None:
+    """Test invalid arguments for NLTKTextSplitter."""
+    with pytest.raises(ValueError):
+        NLTKTextSplitter(
+            chunk_size=80,
+            chunk_overlap=0,
+            separator="\n\n",
+            use_span_tokenize=True,
+        )
+
+
+def test_nltk_text_splitter_with_add_start_index() -> None:
+    splitter = NLTKTextSplitter(
+        chunk_size=80,
+        chunk_overlap=0,
+        separator="",
+        use_span_tokenize=True,
+        add_start_index=True,
+    )
+    txt = (
+        "Innovation drives our success.        "
+        "Collaboration fosters creative solutions. "
+        "Efficiency enhances data management."
+    )
+    docs = [Document(txt)]
+    chunks = splitter.split_documents(docs)
+    assert len(chunks) == 2
+    for chunk in chunks:
+        s_i = chunk.metadata["start_index"]
+        assert chunk.page_content == txt[s_i : s_i + len(chunk.page_content)]

libs/text-splitters/tests/integration_tests/test_text_splitter.py

@@ -1,5 +1,7 @@
 """Test text splitters that require an integration."""
 
+from typing import Any
+
 import pytest
 
 from langchain_text_splitters import (
@@ -11,6 +13,15 @@ from langchain_text_splitters.sentence_transformers import (
 )
 
 
+@pytest.fixture()
+def sentence_transformers() -> Any:
+    try:
+        import sentence_transformers
+    except ImportError:
+        pytest.skip("SentenceTransformers not installed.")
+    return sentence_transformers
+
+
 def test_huggingface_type_check() -> None:
     """Test that type checks are done properly on input."""
     with pytest.raises(ValueError):
@@ -52,7 +63,7 @@ def test_token_text_splitter_from_tiktoken() -> None:
     assert expected_tokenizer == actual_tokenizer
 
 
-def test_sentence_transformers_count_tokens() -> None:
+def test_sentence_transformers_count_tokens(sentence_transformers: Any) -> None:
     splitter = SentenceTransformersTokenTextSplitter(
         model_name="sentence-transformers/paraphrase-albert-small-v2"
     )
@@ -67,7 +78,7 @@ def test_sentence_transformers_count_tokens() -> None:
     assert expected_token_count == token_count
 
 
-def test_sentence_transformers_split_text() -> None:
+def test_sentence_transformers_split_text(sentence_transformers: Any) -> None:
     splitter = SentenceTransformersTokenTextSplitter(
         model_name="sentence-transformers/paraphrase-albert-small-v2"
     )
@@ -77,7 +88,7 @@ def test_sentence_transformers_split_text() -> None:
     assert expected_text_chunks == text_chunks
 
 
-def test_sentence_transformers_multiple_tokens() -> None:
+def test_sentence_transformers_multiple_tokens(sentence_transformers: Any) -> None:
     splitter = SentenceTransformersTokenTextSplitter(chunk_overlap=0)
     text = "Lorem "