Reject NaN/inf numeric results instead of returning them as valid

_validate_response checks isinstance(value, (int, float, np.int64)) for type
'number', but NaN and inf are floats and pass that check, so they are wrapped in
a NumberResponse and returned as the answer. These almost always come from an
aggregation over an empty result (e.g. df['sales'].mean() when a filter matched
zero rows) - a silent wrong answer.

This adds a finite-number check that raises InvalidOutputValueMismatch for NaN/inf.
Adds a regression test (no LLM) that fails on main and passes with the fix.

Author: aaravanmay

pandasai/core/response/parser.py

@@ -43,6 +43,14 @@ def _validate_response(self, result: dict):
                 raise InvalidOutputValueMismatch(
                     "Invalid output: Expected a numeric value for result type 'number', but received a non-numeric value."
                 )
+            # NaN / inf are floats, so they pass the isinstance check above and would be
+            # returned as a valid number. They almost always come from an aggregation over
+            # empty data (e.g. df["x"].mean() on a zero-row result) - reject instead of
+            # silently returning NaN as the answer.
+            if isinstance(result["value"], float) and not np.isfinite(result["value"]):
+                raise InvalidOutputValueMismatch(
+                    "Invalid output: Numeric result is NaN or infinite (likely an aggregation over empty data)."
+                )
         elif result["type"] == "string":
             if not isinstance(result["value"], str):
                 raise InvalidOutputValueMismatch(

tests/test_nan_number_rejected.py

@@ -0,0 +1,33 @@
+"""Regression test: a NaN/inf numeric result must be rejected, not returned as valid.
+
+NaN and inf are floats, so they pass the isinstance(value, (int, float, np.int64))
+check in _validate_response and would be wrapped in a NumberResponse and returned as
+the answer. They almost always come from an aggregation over an empty result
+(e.g. df["sales"].mean() when a WHERE clause matched zero rows).
+
+  with the fix -> PASS (raises InvalidOutputValueMismatch)
+  without it   -> FAIL (returns NumberResponse(nan) silently)
+"""
+import numpy as np
+import pytest
+
+from pandasai.core.response.parser import ResponseParser
+from pandasai.exceptions import InvalidOutputValueMismatch
+
+
+def test_nan_number_rejected():
+    parser = ResponseParser()
+    with pytest.raises(InvalidOutputValueMismatch, match="NaN"):
+        parser.parse({"type": "number", "value": float("nan")})
+
+
+def test_inf_number_rejected():
+    parser = ResponseParser()
+    with pytest.raises(InvalidOutputValueMismatch, match="NaN"):
+        parser.parse({"type": "number", "value": float("inf")})
+
+
+def test_normal_number_still_ok():
+    parser = ResponseParser()
+    resp = parser.parse({"type": "number", "value": 42})
+    assert resp.value == 42