RecursiveCharacterTextSplitter uses regex value instead of original separator when merging and keep_separator is false about langchain

[wulifu2hao]

I drafted a fix for it and will open an pull request once I have the time to add some test cases tomorrow

[loperamos]

Adding some post clarification, I believe that the expected behaviour should be as the following code snippet, since I would expect we want to keep the chunk exactly as the original text. The keep separator flag should only affect when the separator is at the begining of a chunk.

if __name__ == "__main__":
    # Wrong behaviour, using \s instead of regular space
    splitter_keep = RecursiveCharacterTextSplitter(
        separators=[r"\s"],
        keep_separator=False,
        is_separator_regex=True,
        chunk_size=15,
        chunk_overlap=0,
        strip_whitespace=False)
    assert splitter_keep.split_text("Hello world")[0] == r"Hello world"