fix: fix dedup logic

src/rag/retrievers/schema.py

@@ -178,6 +178,22 @@ class SchemaRetriever(BaseRetriever):
             for row in rows
         ]
 
+    @staticmethod
+    def _chunk_key(r: RetrievalResult) -> tuple:
+        """Stable identity for dedup/RRF.
+
+        Includes filename and sheet_name so that tabular column chunks with
+        the same column name across different files (e.g. `id` in two CSVs)
+        and future sheet-level chunks across XLSX sheets don't collide.
+        """
+        d = r.metadata.get("data", {})
+        return (
+            d.get("table_name"),
+            d.get("column_name"),
+            d.get("filename"),
+            d.get("sheet_name"),
+        )
+
     def _rrf_merge(
         self,
         *ranked_lists: list[RetrievalResult],
@@ -190,25 +206,19 @@ class SchemaRetriever(BaseRetriever):
 
         for ranked in ranked_lists:
             for rank, result in enumerate(ranked):
-                data = result.metadata.get("data", {})
-                key = (data.get("table_name"), data.get("column_name") or data.get("filename"))
+                key = self._chunk_key(result)
                 scores[key] = scores.get(key, 0.0) + 1.0 / (k_rrf + rank + 1)
                 if key not in index or result.score > index[key].score:
                     index[key] = result
 
-        def _key(r: RetrievalResult) -> tuple:
-            d = r.metadata.get("data", {})
-            return (d.get("table_name"), d.get("column_name") or d.get("filename"))
-
-        merged = sorted(index.values(), key=lambda r: scores[_key(r)], reverse=True)
+        merged = sorted(index.values(), key=lambda r: scores[self._chunk_key(r)], reverse=True)
         return merged[:top_k]
 
     def _dedup(self, results: list[RetrievalResult]) -> list[RetrievalResult]:
-        """Deduplicate by (table_name, column_name), keeping highest score per unique column."""
+        """Deduplicate by chunk identity, keeping highest score per unique key."""
         seen: dict[tuple, RetrievalResult] = {}
         for r in results:
-            data = r.metadata.get("data", {})
-            key = (data.get("table_name"), data.get("column_name") or data.get("filename"))
+            key = self._chunk_key(r)
             if key not in seen or r.score > seen[key].score:
                 seen[key] = r
         return sorted(seen.values(), key=lambda r: r.score, reverse=True)