compute cache key once (#8394)

* compute cache key once

* compute cache key once in async wrapper

* use a copy of request instead of changing function signatures

* add unit test and fix comments

* bring back deepcopy

---------

Co-authored-by: Antoine Sinton <antoine@cnty.ai>
Co-authored-by: chenmoneygithub <chen.qian@databricks.com>

Author: 3 people

dspy/clients/cache.py

@@ -30,7 +30,6 @@ def __init__(
         disk_cache_dir: str,
         disk_size_limit_bytes: Optional[int] = 1024 * 1024 * 10,
         memory_max_entries: Optional[int] = 1000000,
-
     ):
         """
         Args:
@@ -230,9 +229,11 @@ def sync_wrapper(*args, **kwargs):
                 return cached_result
 
             # Otherwise, compute and store the result
+            # Make a copy of the original request in case it's modified in place, e.g., deleting some fields
+            original_request = copy.deepcopy(modified_request)
             result = fn(*args, **kwargs)
             # `enable_memory_cache` can be provided at call time to avoid indefinite growth.
-            cache.put(modified_request, result, ignored_args_for_cache_key, enable_memory_cache)
+            cache.put(original_request, result, ignored_args_for_cache_key, enable_memory_cache)
 
             return result
 
@@ -249,8 +250,10 @@ async def async_wrapper(*args, **kwargs):
                 return cached_result
 
             # Otherwise, compute and store the result
+            # Make a copy of the original request in case it's modified in place, e.g., deleting some fields
+            original_request = copy.deepcopy(modified_request)
             result = await fn(*args, **kwargs)
-            cache.put(modified_request, result, ignored_args_for_cache_key, enable_memory_cache)
+            cache.put(original_request, result, ignored_args_for_cache_key, enable_memory_cache)
 
             return result
 

tests/clients/test_cache.py

@@ -280,3 +280,40 @@ async def test_function(prompt, model):
         # Call with different arguments should compute again
         result3 = await test_function(prompt="Different", model="openai/gpt-4o-mini")
         assert result3 == "Response for Different with openai/gpt-4o-mini"
+
+
+def test_cache_consistency_with_lm_call_modifies_the_request(cache):
+    """Test that the cache is consistent with the LM call that modifies the request."""
+    from dspy.clients.cache import request_cache
+
+    # Mock the dspy.cache attribute
+    with patch("dspy.cache", cache):
+        # Define a test function
+        @request_cache()
+        def test_function(**kwargs):
+            del kwargs["field_to_delete"]
+            return kwargs
+
+        # First call should compute the result
+        test_function(field_to_delete="delete", field_to_keep="keep")
+
+        # The cache key should use the original request, not the modified one
+        assert (
+            cache.get(
+                {
+                    "field_to_keep": "keep",
+                    "_fn_identifier": f"{test_function.__module__}.{test_function.__qualname__}",
+                }
+            )
+            is None
+        )
+        assert (
+            cache.get(
+                {
+                    "field_to_keep": "keep",
+                    "field_to_delete": "delete",
+                    "_fn_identifier": f"{test_function.__module__}.{test_function.__qualname__}",
+                }
+            )
+            is not None
+        )