DSPy Conversation History

Experiment with `dspy-ai` conversation history helpers—messages persist in memory and each response is generated with the full transcript as context.

How it works

A DSPy `Predict` module runs with a signature that includes the full conversation transcript before asking for the next reply.

Execution flow

• Append each user turn to in-memory history before calling the model.
• Expose `history` + `user_query` in the signature for structured context.
• Fallback to the OpenAI helper if DSPy fails, mirroring the tutorial's reliability pattern.
• Caching disabled here to align with stateless tutorial behavior.

import dspy

class ConversationSignature(dspy.Signature):
    history = dspy.InputField()
    user_query = dspy.InputField()
    response = dspy.OutputField()

lm = dspy.LM("openai/gpt-4o-mini", temperature=0.6)
dspy.configure(lm=lm)

predict = dspy.Predict(ConversationSignature)
prediction = predict(
    history="Assistant: Hello!",
    user_query="How does DSPy track history?",
)
print(prediction.response)

Further reading

• DSPy Conversation History tutorial
• DSPy Signatures guide
• DSPy Predict module reference
• Handling caches in DSPy