diff --git a/quantmind/rag/document.py b/quantmind/rag/document.py
index 4f5c3b7..3c6ed2a 100644
--- a/quantmind/rag/document.py
+++ b/quantmind/rag/document.py
@@ -16,6 +16,18 @@
     ParsedPage,
 )
 
+_CJK_RANGES = "぀-ヿ㐀-䶿一-鿿豈-﫿"
+_CJK_AWARE_TOKEN_PATTERN = rf"(?u)[^\W{_CJK_RANGES}]{{2,}}|[{_CJK_RANGES}]"
+"""Tokenize CJK scripts per character, leaving other scripts untouched.
+
+The retriever default ``(?u)\\b\\w\\w+\\b`` assumes whitespace-delimited words.
+Japanese and Chinese are not, so a whole phrase becomes a single token and a
+query token can never equal a document token: every chunk scores ``0.0`` while
+the retriever still returns ``top_k`` hits. Splitting only CJK characters keeps
+the two-or-more-character rule for Latin, Cyrillic and other alphabets, so their
+scores and ranking stay identical.
+"""
+
 
 @dataclass(frozen=True)
 class SentenceSplitterConfig:
@@ -154,6 +166,7 @@ def retrieve_parsed_document(
     retriever = BM25Retriever.from_defaults(
         nodes=nodes,
         similarity_top_k=min(top_k, len(nodes)),
+        token_pattern=_CJK_AWARE_TOKEN_PATTERN,
     )
     results = retriever.retrieve(query)
     return tuple(
diff --git a/tests/rag/test_document.py b/tests/rag/test_document.py
index f80dd51..7ed95ae 100644
--- a/tests/rag/test_document.py
+++ b/tests/rag/test_document.py
@@ -5,6 +5,7 @@
 
 from quantmind.preprocess.format import parse_pdf
 from quantmind.rag import (
+    ParsedChunk,
     SentenceSplitterConfig,
     chunk_parsed_document,
     retrieve_parsed_document,
@@ -59,3 +60,63 @@ async def test_retrieval_rejects_invalid_query_arguments(self):
             retrieve_parsed_document(chunks, "   ")
         with self.assertRaisesRegex(ValueError, "top_k"):
             retrieve_parsed_document(chunks, "fixture", top_k=0)
+
+
+class ScriptAwareRetrievalTests(unittest.TestCase):
+    """Retrieval over scripts that are not whitespace-delimited."""
+
+    @staticmethod
+    def _chunks(texts: tuple[str, ...]) -> tuple[ParsedChunk, ...]:
+        return tuple(
+            ParsedChunk(
+                chunk_id=f"chunk-{index}",
+                text=text,
+                source_hash="0" * 64,
+                page_number=index + 1,
+                start_char=0,
+                end_char=len(text),
+                block_boxes=(),
+                screenshot_path=None,
+                image_paths=(),
+            )
+            for index, text in enumerate(texts)
+        )
+
+    def test_japanese_query_ranks_the_matching_chunk_first(self):
+        chunks = self._chunks(
+            (
+                "国際標準化に関する動向と規格策定プロセスの概観",
+                "人材育成の方針と研修体系の整備について",
+            )
+        )
+
+        hits = retrieve_parsed_document(chunks, "国際標準化", top_k=2)
+
+        self.assertEqual(hits[0].chunk.chunk_id, "chunk-0")
+        self.assertGreater(hits[0].score, 0.0)
+
+    def test_chinese_query_ranks_the_matching_chunk_first(self):
+        chunks = self._chunks(
+            (
+                "货币政策与利率变动对债券市场的影响",
+                "上市公司季度财报披露时间表",
+            )
+        )
+
+        hits = retrieve_parsed_document(chunks, "利率变动", top_k=2)
+
+        self.assertEqual(hits[0].chunk.chunk_id, "chunk-0")
+        self.assertGreater(hits[0].score, 0.0)
+
+    def test_latin_query_ranks_the_matching_chunk_first(self):
+        chunks = self._chunks(
+            (
+                "knowledge extraction architecture and retrieval evaluation",
+                "deterministic preprocessing keeps provenance replayable",
+            )
+        )
+
+        hits = retrieve_parsed_document(chunks, "retrieval evaluation", top_k=2)
+
+        self.assertEqual(hits[0].chunk.chunk_id, "chunk-0")
+        self.assertGreater(hits[0].score, 0.0)