fix: Fix LLMEvaluator serialization

haystack/components/evaluators/llm_evaluator.py

@@ -11,7 +11,7 @@
 from haystack import component, default_from_dict, default_to_dict
 from haystack.components.builders import PromptBuilder
 from haystack.components.generators import OpenAIGenerator
-from haystack.utils import Secret, deserialize_secrets_inplace
+from haystack.utils import Secret, deserialize_secrets_inplace, deserialize_type, serialize_type
 
 
 @component
@@ -266,10 +266,12 @@ def to_dict(self) -> Dict[str, Any]:
         :returns:
             The serialized component as a dictionary.
         """
+        # Since we cannot currently serialize tuples, convert the inputs to a list.
+        inputs = [[name, serialize_type(type_)] for name, type_ in self.inputs]
         return default_to_dict(
             self,
             instructions=self.instructions,
-            inputs=self.inputs,
+            inputs=inputs,
             outputs=self.outputs,
             examples=self.examples,
             api=self.api,
@@ -287,6 +289,10 @@ def from_dict(cls, data: Dict[str, Any]) -> "LLMEvaluator":
         :returns:
             The deserialized component instance.
         """
+        data["init_parameters"]["inputs"] = [
+            (name, deserialize_type(type_)) for name, type_ in data["init_parameters"]["inputs"]
+        ]
+
         deserialize_secrets_inplace(data["init_parameters"], keys=["api_key"])
         return default_from_dict(cls, data)
 

releasenotes/notes/llm-evaluator-serde-fix-aa5b27d2524db9c5.yaml

@@ -0,0 +1,4 @@
+---
+fixes:
+  - |
+    Correctly serialize tuples and types in the init parameters of the `LLMEvaluator` component and its subclasses.

test/components/evaluators/test_llm_evaluator.py

@@ -3,9 +3,9 @@
 # SPDX-License-Identifier: Apache-2.0
 from typing import List
 
-import numpy as np
 import pytest
 
+from haystack import Pipeline
 from haystack.components.evaluators import LLMEvaluator
 from haystack.utils.auth import Secret
 
@@ -205,7 +205,7 @@ def test_to_dict_default(self, monkeypatch):
                 "api_key": {"env_vars": ["OPENAI_API_KEY"], "strict": True, "type": "env_var"},
                 "api": "openai",
                 "instructions": "test-instruction",
-                "inputs": [("predicted_answers", List[str])],
+                "inputs": [["predicted_answers", "typing.List[str]"]],
                 "outputs": ["score"],
                 "progress_bar": True,
                 "examples": [
@@ -223,7 +223,7 @@ def test_from_dict(self, monkeypatch):
                 "api_key": {"env_vars": ["OPENAI_API_KEY"], "strict": True, "type": "env_var"},
                 "api": "openai",
                 "instructions": "test-instruction",
-                "inputs": [("predicted_answers", List[str])],
+                "inputs": [["predicted_answers", "typing.List[str]"]],
                 "outputs": ["score"],
                 "examples": [
                     {"inputs": {"predicted_answers": "Football is the most popular sport."}, "outputs": {"score": 0}}
@@ -266,7 +266,7 @@ def test_to_dict_with_parameters(self, monkeypatch):
                 "api_key": {"env_vars": ["ENV_VAR"], "strict": True, "type": "env_var"},
                 "api": "openai",
                 "instructions": "test-instruction",
-                "inputs": [("predicted_answers", List[str])],
+                "inputs": [["predicted_answers", "typing.List[str]"]],
                 "outputs": ["custom_score"],
                 "progress_bar": True,
                 "examples": [
@@ -282,6 +282,21 @@ def test_to_dict_with_parameters(self, monkeypatch):
             },
         }
 
+    def test_serde(self, monkeypatch):
+        monkeypatch.setenv("OPENAI_API_KEY", "test-api-key")
+        pipeline = Pipeline()
+        component = LLMEvaluator(
+            instructions="test-instruction",
+            inputs=[("questions", List[str]), ("predicted_answers", List[List[str]])],
+            outputs=["score"],
+            examples=[
+                {"inputs": {"predicted_answers": "Football is the most popular sport."}, "outputs": {"score": 0}}
+            ],
+        )
+        pipeline.add_component("evaluator", component)
+        serialized_pipeline = pipeline.dumps()
+        deserialized_pipeline = Pipeline.loads(serialized_pipeline)
+
     def test_run_with_different_lengths(self, monkeypatch):
         monkeypatch.setenv("OPENAI_API_KEY", "test-api-key")
         component = LLMEvaluator(