diff --git a/CHANGELOG.md b/CHANGELOG.md
index 8142bc16f..facd8f7f5 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -11,6 +11,9 @@ This project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 ### Added
 - `pw.xpacks.llm.rerankers.LLMReranker` now accepts `call_kwargs`, the kwargs passed to each call of the LLM. The default is still `{"temperature": 0}`; pass `call_kwargs={}` to use the reranker with models that accept only the default `temperature`, such as Claude Opus 4.7 and newer (including Claude Opus 5.5) on Bedrock.
 
+### Fixed
+- `pathway.xpacks.llm.splitters.TokenCountSplitter` no longer skips source text when a punctuation cut falls inside a token. It locates the cut using token byte lengths and advances by exactly the whole source tokens emitted in the chunk.
+
 ## [0.33.0] - 2026-09-18
 
 ### Changed
diff --git a/python/pathway/xpacks/llm/splitters.py b/python/pathway/xpacks/llm/splitters.py
index ea8457af7..3d94f70f9 100644
--- a/python/pathway/xpacks/llm/splitters.py
+++ b/python/pathway/xpacks/llm/splitters.py
@@ -259,6 +259,7 @@ def chunk(self, text: str, metadata: dict = {}, **kwargs) -> list[tuple[str, dic
         while i < len(tokens):
             chunk_tokens = tokens[i : i + max_tokens]
             chunk = tokenizer.decode(chunk_tokens)
+            consumed = len(chunk_tokens)
             last_punctuation = max(
                 [chunk.rfind(p) for p in self.PUNCTUATION], default=-1
             )
@@ -266,8 +267,24 @@ def chunk(self, text: str, metadata: dict = {}, **kwargs) -> list[tuple[str, dic
                 last_punctuation != -1
                 and last_punctuation > self.CHARS_PER_TOKEN * min_tokens
             ):
-                chunk = chunk[: last_punctuation + 1]
-            i += len(tokenizer.encode_ordinary(chunk))
+                cut_bytes = len(chunk[: last_punctuation + 1].encode("utf-8"))
+                token_bytes = 0
+                cut_tokens = 0
+                # Token boundaries can split a UTF-8 character. Count bytes rather
+                # than decoding every growing prefix to locate the punctuation cut.
+                for token in chunk_tokens:
+                    token_bytes += len(tokenizer.decode_single_token_bytes(token))
+                    if token_bytes > cut_bytes:
+                        break
+                    cut_tokens += 1
+                # If the first token straddles the cut, keep the full window so
+                # that the emitted text still matches the tokens we consume.
+                if cut_tokens:
+                    consumed = cut_tokens
+                    chunk = tokenizer.decode(chunk_tokens[:consumed])
+            # Guard against a chunk that would not advance the cursor, which would
+            # otherwise stall the loop forever.
+            i += max(consumed, 1)
             output.append((chunk, metadata))
 
         return output
diff --git a/python/pathway/xpacks/llm/tests/test_splitters.py b/python/pathway/xpacks/llm/tests/test_splitters.py
index 88bf4b1de..afaa24ebc 100644
--- a/python/pathway/xpacks/llm/tests/test_splitters.py
+++ b/python/pathway/xpacks/llm/tests/test_splitters.py
@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import pandas as pd
+import pytest
 
 import pathway as pw
 from pathway.tests.utils import assert_table_equality
@@ -31,6 +32,49 @@ def test_tokencount():
     assert_table_equality(result, input_table)
 
 
+def test_tokencount_does_not_drop_characters():
+    # Regression: when a chunk was cut at a punctuation mark, the cursor advanced
+    # by the re-encoded kept text, which could skip the tokens straddling the cut
+    # and silently drop characters. Concatenating the chunks must reconstruct the
+    # (unicode-normalized) input exactly.
+    import unicodedata
+
+    splitter = TokenCountSplitter(min_tokens=1, max_tokens=3)
+    txt = "a.b.c.d.e.f.g.h.i.j.k.l."
+    chunks = [chunk for chunk, _ in splitter.chunk(txt)]
+
+    assert "".join(chunks) == unicodedata.normalize("NFKC", txt)
+
+
+@pytest.mark.parametrize(
+    "txt",
+    [
+        "Привет, мир. Это проверка разбиения текста! Работает ли оно? Да. " * 5,
+        "你好,世界。这是一个测试!它有效吗?是的. " * 10,
+    ],
+    ids=["russian", "chinese"],
+)
+def test_tokencount_does_not_duplicate_non_ascii(txt):
+    import unicodedata
+
+    splitter = TokenCountSplitter()
+    chunks = [chunk for chunk, _ in splitter.chunk(txt)]
+
+    assert "".join(chunks) == unicodedata.normalize("NFKC", txt)
+
+
+def test_tokencount_preserves_token_straddling_first_punctuation_cut():
+    # cl100k_base encodes "...)" as a single token. If the
+    # punctuation prefix has no whole token, emit the window intact.
+    splitter = TokenCountSplitter(min_tokens=0, max_tokens=1)
+    txt = "...) tail"
+    metadata = {"source": "example"}
+    chunks = splitter.chunk(txt, metadata)
+
+    assert "".join(chunk for chunk, _ in chunks) == txt
+    assert all(chunk and meta == metadata for chunk, meta in chunks)
+
+
 def test_recursive_from_encoding():
     splitter = RecursiveSplitter(
         encoding_name="cl100k_base", chunk_size=30, chunk_overlap=0