support predicted output

Author: chenmoneygithub

dspy/clients/lm.py

@@ -83,9 +83,9 @@ def __init__(
 
         if model_pattern:
             # Handle OpenAI reasoning models (o1, o3)
-            assert (
-                max_tokens >= 20_000 and temperature == 1.0
-            ), "OpenAI's reasoning models require passing temperature=1.0 and max_tokens >= 20_000 to `dspy.LM(...)`"
+            assert max_tokens >= 20_000 and temperature == 1.0, (
+                "OpenAI's reasoning models require passing temperature=1.0 and max_tokens >= 20_000 to `dspy.LM(...)`"
+            )
             self.kwargs = dict(temperature=temperature, max_completion_tokens=max_tokens, **kwargs)
         else:
             self.kwargs = dict(temperature=temperature, max_tokens=max_tokens, **kwargs)

dspy/predict/predict.py

@@ -108,6 +108,17 @@ def _forward_preprocess(self, **kwargs):
         if (temperature is None or temperature <= 0.15) and num_generations > 1:
             config["temperature"] = 0.7
 
+        if "prediction" in kwargs:
+            if (
+                isinstance(kwargs["prediction"], dict)
+                and kwargs["prediction"].get("type") == "content"
+                and "content" in kwargs["prediction"]
+            ):
+                # If the `prediction` is the standard predicted outputs format
+                # (https://platform.openai.com/docs/guides/predicted-outputs), we remvoe it from input kwargs and add it
+                # to the lm kwargs.
+                config["prediction"] = kwargs.pop("prediction")
+
         if not all(k in kwargs for k in signature.input_fields):
             present = [k for k in signature.input_fields if k in kwargs]
             missing = [k for k in signature.input_fields if k not in kwargs]

tests/predict/test_predict.py

@@ -572,3 +572,27 @@ async def test_async_predict():
     dspy.settings.configure(lm=DummyLM([{"answer": "Paris"}]))
     result = await program.acall(question="What is the capital of France?")
     assert result.answer == "Paris"
+
+
+def test_predicted_outputs_piped_from_predict_to_lm_call():
+    program = Predict("question -> answer")
+    dspy.settings.configure(lm=dspy.LM("openai/gpt-4o-mini"))
+
+    with patch("litellm.completion") as mock_completion:
+        program(
+            question="Why did a chicken cross the kitchen?",
+            prediction={"type": "content", "content": "A chicken crossing the kitchen"},
+        )
+
+        assert mock_completion.call_args[1]["prediction"] == {
+            "type": "content",
+            "content": "A chicken crossing the kitchen",
+        }
+
+    # If the signature has prediction as an input field, and the prediction is not set as the standard predicted output
+    # format, it should not be passed to the LM.
+    program = Predict("question, prediction -> judgement")
+    with patch("litellm.completion") as mock_completion:
+        program(question="Why did a chicken cross the kitchen?", prediction="To get to the other side!")
+
+    assert "prediction" not in mock_completion.call_args[1]