fix(chunker): honor configured delimiters as soft boundaries in TokenChunker (#17721)

## Summary

`TokenChunker._invoke` discarded the user's configured `delimiters`
whenever no backtick-wrapped delimiter was present: it passed a
hardcoded `""` to `naive_merge`, so the configured delimiters (including
the default `["\n"]`) were never forwarded. `naive_merge` then ignored
the newline sentence boundary and cut chunks **mid-sentence** once the
token budget was exceeded.

## Root cause

`token_chunker.py:326` called `naive_merge(payload, chunk_token_size,
"", overlapped_percent)`. `_compile_delimiter_pattern` intentionally
returns `""` for bare (non-backtick) delimiters — that return value is a
*path selector* (empty → token-budget merge; non-empty → hard
`_split_text_by_pattern` split). The bug was not in that selector but in
the `else` branch, which threw away `self._param.delimiters` instead of
forwarding it.

## Fix

Forward the configured delimiters as a soft boundary:

```python
else naive_merge(
    payload,
    self._param.chunk_token_size,
    "".join(self._param.delimiters),
    overlapped_percent,
)
```

`naive_merge` already parses the string via the canonical
`parse_delimiter_field`, so bare and backtick-wrapped delimiters are
honored as soft boundaries while the token budget is still respected.
The path-selection role of `_compile_delimiter_pattern` is untouched:
backtick-wrapped delimiters still select the hard
`_split_text_by_pattern` path; bare delimiters still take the
token-budget merge path. No regression for any previously-working
(wrapped-delimiter) configuration.

## Test

Adds `rag/flow/tests/test_token_chunker_delimiter.py`:
- `test_token_chunker_token_size_mode_does_not_split_sentences` — fails
on the old code (3 sentences cut mid-stream), passes after the fix.
- `test_naive_merge_empty_delimiter_ignores_newline_break` — root-cause
companion asserting `naive_merge("")` cuts while `naive_merge("\n")`
preserves boundaries.

Both tests skip when the tokenizer is unavailable (dead-tokenizer
guard).

## Note

This branch contains only this one-line fix on top of `main`; it is
intentionally independent of the unrelated tokenizer/offline-BPE work.

Co-authored-by: CodeBuddy <noreply@codebuddy.ai>

rag/flow/chunker/token_chunker.py

@@ -323,7 +323,7 @@ class TokenChunker(ProcessBase):
                 else naive_merge(
                     payload,
                     self._param.chunk_token_size,
-                    "",
+                    "".join(self._param.delimiters),
                     overlapped_percent,
                 )
             )

rag/flow/tests/test_token_chunker_delimiter.py

@@ -0,0 +1,117 @@
+"""Regression test for token_size-mode delimiter handling in TokenChunker.
+
+The default ``delimiters`` a user supplies (e.g. ``["\\n"]``) are only honoured
+when they are backtick-wrapped; a bare delimiter makes
+``_compile_delimiter_pattern`` return ``""``, so ``TokenChunker._invoke`` falls
+through to ``naive_merge(payload, chunk_token_size, "")`` (token_chunker.py:323).
+
+Passing ``""`` discards the sentence-boundary preference that ``naive_merge``
+carries by default (``"\\n。；！？"``). Without a forced ``\\n`` break, the token
+budget flush can land *inside* a sentence, so a line like
+``"Sentence number 7. alpha beta gamma delta epsilon zeta eta theta iota kappa"``
+gets split across two chunks. The Go port (``sentenceDelimiter = (\\n|[!?。；！？])``
+in token.go:340) always breaks on ``\\n``, so it never cuts a sentence mid-stream.
+
+This test exposes that divergence: with ``delimiters=["\\n"]`` in token_size mode,
+every original sentence must survive intact in a single chunk. It currently FAILS
+because of the ``""`` passed at token_chunker.py:326; forwarding
+``"".join(self._param.delimiters)`` to ``naive_merge`` (or at least
+``"\\n。；！？"``) makes it pass and aligns Python with the Go side.
+
+Note: the test needs a live tiktoken encoder. ``common.token_utils`` returns 0 on
+any encoder failure, which would make every token budget "never exceeded" and
+hide the bug behind a single-chunk baseline. We skip rather than pass under a
+dead tokenizer, mirroring ``capture_golden.py::_assert_tokenizer_alive``.
+"""
+
+import asyncio
+import types
+
+from rag.flow.chunker.token_chunker import TokenChunker, TokenChunkerParam
+from rag.nlp import naive_merge
+from common.token_utils import num_tokens_from_string
+
+
+def _build_token_chunker(param: dict) -> TokenChunker:
+    """Construct a TokenChunker without the real Graph the normal __init__ wants.
+
+    ``ComponentBase.__init__`` asserts ``canvas`` is a real ``Graph``; a unit test
+    has none. Bypass it and attach the three attributes ``_invoke`` actually reads
+    (``_canvas``, ``_param``, ``callback``) so the genuine chunking path runs with
+    the real ``naive_merge`` and the real ``num_tokens_from_string``. Mirrors
+    ``capture_golden.py::_build_component``.
+    """
+    p = TokenChunkerParam()
+    for key, value in param.items():
+        setattr(p, key, value)
+    p.check()  # __new__ bypassed the constructor's validation; reproduce it.
+
+    comp = TokenChunker.__new__(TokenChunker)
+    comp._canvas = types.SimpleNamespace(_doc_id=None, _tenant_id="t")
+    comp._param = p
+    comp.callback = lambda *_a, **_kw: None
+    return comp
+
+
+def _invoke_text(comp: TokenChunker, text: str) -> list[dict]:
+    asyncio.run(comp._invoke(name="t", output_format="text", text=text))
+    return comp._param.outputs["chunks"]["value"]
+
+
+def _make_sentences(n: int) -> list[str]:
+    # Each sentence is ~16 tokens of plain English; well under the 128 budget,
+    # so several fit per chunk and the token-budget flush lands mid-sentence.
+    return [f"Sentence number {i}. alpha beta gamma delta epsilon zeta eta theta iota kappa" for i in range(n)]
+
+
+def test_token_chunker_token_size_mode_does_not_split_sentences():
+    # Guard: a dead tokenizer would collapse everything to one chunk and hide
+    # the bug. Skip instead of recording a poisoned pass.
+    if num_tokens_from_string("alive tokenizer probe sentence") <= 0:
+        import pytest
+
+        pytest.skip("tiktoken encoder unavailable; num_tokens_from_string returned 0")
+
+    sentences = _make_sentences(30)
+    payload = "\n".join(sentences)
+
+    comp = _build_token_chunker({"chunk_token_size": 128, "delimiters": ["\n"]})
+    chunks = _invoke_text(comp, payload)
+    texts = [c["text"] for c in chunks]
+
+    # An original sentence is "intact" when its full text appears in exactly one
+    # chunk. If the budget flush fell inside it, the sentence is split across two
+    # chunks and this assertion fails.
+    split_sentences = [s for s in sentences if not any(s in t for t in texts)]
+    assert not split_sentences, (
+        f"{len(split_sentences)} sentence(s) were cut mid-stream by "
+        f"token_size mode with delimiters=['\\n'] (got {len(chunks)} chunks). "
+        "naive_merge was called with an empty delimiter (token_chunker.py:326), "
+        "dropping the '\\n' sentence-boundary preference. Example split: "
+        f"{split_sentences[0]!r}"
+    )
+
+
+def test_naive_merge_empty_delimiter_ignores_newline_break():
+    """Root-cause check: naive_merge('') can cut mid-sentence; naive_merge('\\n') cannot.
+
+    Documents why forwarding the configured delimiters fixes the TokenChunker bug.
+    """
+    if num_tokens_from_string("alive tokenizer probe sentence") <= 0:
+        import pytest
+
+        pytest.skip("tiktoken encoder unavailable; num_tokens_from_string returned 0")
+
+    sentences = _make_sentences(30)
+    payload = "\n".join(sentences)
+
+    chunks_empty = naive_merge(payload, 128, "")
+    chunks_nl = naive_merge(payload, 128, "\n")
+
+    split_empty = [s for s in sentences if not any(s in t for t in chunks_empty)]
+    split_nl = [s for s in sentences if not any(s in t for t in chunks_nl)]
+
+    # The empty-delimiter call is the one that cuts sentences; the '\\n' call
+    # pre-splits on newline and keeps each sentence whole.
+    assert split_empty, "expected naive_merge('') to cut at least one sentence mid-stream"
+    assert not split_nl, "naive_merge('\\n') should preserve every sentence boundary"