Skip to content

The agent engineering platform.

5.7 KB
Raw Blame
"""Test text splitters that require an integration."""

from typing import TYPE_CHECKING, cast

import pytest
from transformers.models.auto.tokenization_auto import AutoTokenizer

from langchain_text_splitters import (
    TokenTextSplitter,
)
from langchain_text_splitters.character import CharacterTextSplitter
from langchain_text_splitters.sentence_transformers import (
    SentenceTransformersTokenTextSplitter,
)

if TYPE_CHECKING:
    from transformers import PreTrainedTokenizerBase


def test_huggingface_type_check() -> None:
    """Test that type checks are done properly on input."""
    with pytest.raises(
        ValueError,
        match="Tokenizer received was not an instance of PreTrainedTokenizerBase",
    ):
        CharacterTextSplitter.from_huggingface_tokenizer("foo")  # ty: ignore[invalid-argument-type]


def test_huggingface_tokenizer() -> None:
    """Test text splitter that uses a HuggingFace tokenizer."""
    tokenizer = AutoTokenizer.from_pretrained("gpt2")
    text_splitter = CharacterTextSplitter.from_huggingface_tokenizer(
        # AutoTokenizer.from_pretrained returns a backend union
        # (TokenizersBackend | SentencePieceBackend) that ty won't narrow to
        # PreTrainedTokenizerBase, so cast to satisfy from_huggingface_tokenizer.
        cast("PreTrainedTokenizerBase", tokenizer),
        separator=" ",
        chunk_size=1,
        chunk_overlap=0,
    )
    output = text_splitter.split_text("foo bar")
    assert output == ["foo", "bar"]


def test_token_text_splitter() -> None:
    """Test no overlap."""
    splitter = TokenTextSplitter(chunk_size=5, chunk_overlap=0)
    output = splitter.split_text("abcdef" * 5)  # 10 token string
    expected_output = ["abcdefabcdefabc", "defabcdefabcdef"]
    assert output == expected_output


def test_token_text_splitter_overlap() -> None:
    """Test with overlap."""
    splitter = TokenTextSplitter(chunk_size=5, chunk_overlap=1)
    output = splitter.split_text("abcdef" * 5)  # 10 token string
    expected_output = ["abcdefabcdefabc", "abcdefabcdefabc", "abcdef"]
    assert output == expected_output


def test_token_text_splitter_from_tiktoken() -> None:
    splitter = TokenTextSplitter.from_tiktoken_encoder(model_name="gpt-4.1-mini")
    expected_tokenizer = "o200k_base"
    actual_tokenizer = splitter._tokenizer.name
    assert expected_tokenizer == actual_tokenizer


def test_character_text_splitter_from_tiktoken() -> None:
    """The base (non-`TokenTextSplitter`) `from_tiktoken_encoder` path.

    Verifies that a plain `CharacterTextSplitter` gets a token-based length
    function wired in, without the tiktoken configuration leaking into a
    constructor that does not accept it.
    """
    splitter = CharacterTextSplitter.from_tiktoken_encoder(
        encoding_name="gpt2", chunk_size=5, chunk_overlap=0
    )
    # Length is measured in tokens, not characters: "abcdef" is 2 gpt2 tokens,
    # so the 30-character string below is 10 tokens.
    assert splitter._length_function("abcdef" * 5) == 10


@pytest.mark.requires("sentence_transformers")
def test_sentence_transformers_count_tokens() -> None:
    splitter = SentenceTransformersTokenTextSplitter(
        model_name="sentence-transformers/paraphrase-albert-small-v2"
    )
    text = "Lorem ipsum"

    token_count = splitter.count_tokens(text=text)

    expected_start_stop_token_count = 2
    expected_text_token_count = 5
    expected_token_count = expected_start_stop_token_count + expected_text_token_count

    assert expected_token_count == token_count


@pytest.mark.requires("sentence_transformers")
def test_sentence_transformers_split_text() -> None:
    splitter = SentenceTransformersTokenTextSplitter(
        model_name="sentence-transformers/paraphrase-albert-small-v2"
    )
    text = "lorem ipsum"
    text_chunks = splitter.split_text(text=text)
    expected_text_chunks = [text]
    assert expected_text_chunks == text_chunks


@pytest.mark.requires("sentence_transformers")
def test_sentence_transformers_multiple_tokens() -> None:
    splitter = SentenceTransformersTokenTextSplitter(chunk_overlap=0)
    assert splitter.maximum_tokens_per_chunk is not None
    text = "Lorem "

    text_token_count_including_start_and_stop_tokens = splitter.count_tokens(text=text)
    count_start_and_end_tokens = 2
    token_multiplier = (
        count_start_and_end_tokens
        + (splitter.maximum_tokens_per_chunk - count_start_and_end_tokens)
        // (
            text_token_count_including_start_and_stop_tokens
            - count_start_and_end_tokens
        )
        + 1
    )

    # `text_to_split` does not fit in a single chunk
    text_to_embed = text * token_multiplier

    text_chunks = splitter.split_text(text=text_to_embed)

    expected_number_of_chunks = 2

    assert expected_number_of_chunks == len(text_chunks)
    actual = splitter.count_tokens(text=text_chunks[1]) - count_start_and_end_tokens
    expected = (
        token_multiplier * (text_token_count_including_start_and_stop_tokens - 2)
        - splitter.maximum_tokens_per_chunk
    )
    assert expected == actual


@pytest.mark.requires("sentence_transformers")
def test_sentence_transformers_with_additional_model_kwargs() -> None:
    """Test passing model_kwargs to SentenceTransformer."""
    # ensure model is downloaded (online)
    splitter_online = SentenceTransformersTokenTextSplitter(
        model_name="sentence-transformers/paraphrase-albert-small-v2"
    )
    text = "lorem ipsum"
    splitter_online.count_tokens(text=text)

    # test offline model loading using model_kwargs
    splitter_offline = SentenceTransformersTokenTextSplitter(
        model_name="sentence-transformers/paraphrase-albert-small-v2",
        model_kwargs={"local_files_only": True},
    )
    splitter_offline.count_tokens(text=text)
    assert splitter_offline.tokenizer is not None