fix(text-splitters): prevent start_index=-1 in token-based splitters

[Kyle Tse (shtse8)]

Description

Fixes #29884

TextSplitter.create_documents() calculates a character offset for text.find() using self._chunk_overlap. For token-based splitters like TokenTextSplitter, chunk_overlap is expressed in tokens rather than characters, so the offset can overshoot the actual chunk position and text.find() returns -1.

Root Cause

In create_documents():

offset = index + previous_chunk_len - self._chunk_overlap
index = text.find(chunk, max(0, offset))

previous_chunk_len is in characters (from len(chunk)), but self._chunk_overlap is in tokens for TokenTextSplitter. The mismatch means the search window can start after where the chunk actually appears, causing text.find() to miss and return -1.

Fix

Add a fallback: when the initial text.find() returns -1, retry the search from the beginning of the text:

index = text.find(chunk, max(0, offset))
if index == -1:
    index = text.find(chunk)

This is safe because chunks are always substrings of the source text, and searching from position 0 guarantees we find the first occurrence.

Tests

Added test_token_text_splitter_start_index_no_negative which:

• Uses TokenTextSplitter with chunk_size=10, chunk_overlap=5, add_start_index=True
• Verifies all start_index values are >= 0
• Verifies each start_index correctly locates the chunk in the source text