DSPy integration

DSPy programs accumulate dspy.History over many turns. Plug convopack in front of the LM call to keep that history under the model's context window.

import dspy
from convopack import Packer, SummaryEvict

dspy.configure(lm=dspy.LM("openai/gpt-4o"))

packer = Packer(
    budget=8000,
    tokenizer="tiktoken:gpt-4o",
    strategy=SummaryEvict(summarise),
    pin=("system", "last_user"),
    cache=True,
)

class Chat(dspy.Signature):
    """Be a helpful assistant."""
    history: dspy.History = dspy.InputField()
    user_input: str = dspy.InputField()
    response: str = dspy.OutputField()

predict = dspy.Predict(Chat)

history = dspy.History(messages=[])

def turn(user_input: str) -> str:
    # convopack reads the dspy.History and writes a fresh, budget-trimmed one.
    trimmed = packer.pack_dspy(history, as_history=True)
    out = predict(history=trimmed, user_input=user_input)
    # Append the new exchange to the long history we keep on the side.
    history.messages.append({"role": "user", "content": user_input})
    history.messages.append({"role": "assistant", "content": out.response})
    return out.response

What pack_dspy does

• Reads dspy.History.messages (a list of OpenAI-shape dicts).
• Runs the configured strategy and tokenizer.
• Returns either a plain list[dict] (default) or a fresh dspy.History when as_history=True.

Compose with everything else

pack_dspy is just a thin layer over the OpenAI provider adapter; cache markers, tool-pair atomicity, async summarisation, and pinning all work the same way as in any other recipe.