Fix knowledge loss before oversized sentences

services/kb_service.py

@@ -56,7 +56,8 @@ def _chunk_text(text: str) -> list[str]:
             current = f"{current}\n{piece}"
         else:
             chunks.append(current)
-            tail = current[-CHUNK_OVERLAP:] if CHUNK_OVERLAP else ""
+            overlap = min(CHUNK_OVERLAP, max(0, CHUNK_SIZE - len(piece) - 1))
+            tail = current[-overlap:] if overlap else ""
             current = (tail + "\n" + piece).strip() if tail else piece
     if current:
         chunks.append(current)
@@ -69,6 +70,8 @@ def _split_long(text: str) -> list[str]:
     current = ""
     for s in sentences:
         if len(s) > CHUNK_SIZE:
+            if current:
+                out.append(current)
             # Hard-split an oversized sentence on word boundaries.
             for i in range(0, len(s), CHUNK_SIZE):
                 out.append(s[i : i + CHUNK_SIZE])

tests/unit/services/test_kb_service.py

@@ -38,6 +38,21 @@ def test_split_long_hard_splits_oversized_sentence():
     assert "".join(pieces) == "X" * 1000
 
 
+def test_split_long_preserves_buffer_before_oversized_sentence():
+    prefix = "Warranty lasts two years."
+    oversized = "X" * (kb.CHUNK_SIZE + 25)
+    text = f"{prefix} {oversized}"
+
+    pieces = kb._split_long(text)
+    chunks = kb._chunk_text(text)
+
+    assert pieces[0] == prefix
+    assert "".join(pieces[1:]) == oversized
+    assert all(len(piece) <= kb.CHUNK_SIZE for piece in pieces)
+    assert any(prefix in chunk for chunk in chunks)
+    assert all(len(chunk) <= kb.CHUNK_SIZE for chunk in chunks)
+
+
 # --- build_context_block -----------------------------------------------------
 def test_context_block_empty():
     assert kb.build_context_block([]) == ""