pico-dspy

A stripped-down version of dspy: https://github.com/stanfordnlp/dspy

Use case

dspy may be too bulky and slow to import in some cases, especially during highly iterative development with code auto reloaders. pico-dspy lets you use the usual signature+predict flow during development and switch your dspy-compatible signatures to the regular dspy when you need GEPA, fine tuning, etc.

Limitations

Several features or limitations compared to dspy:

• uses pydantic-ai instead of litellm, as litellm is infamously slow to import (https://github.com/BerriAI/litellm/issues/7605)

• uses lazy imports (with wrapt) to reduce import time

• just inference, no GEPA or anything

• only JSONAdapter, only one LM type

• only Predict supported

• text inference only

• no tool support

• no async api

• no loading and saving

• no caching

• no dspy.configure

• tested only with gemini

What works

import pico_dspy as dspy


class EmailParse(dspy.Signature):
    """
    Parse a raw email.
    """
    raw_email = dspy.InputField(desc="Raw email")
    subject = dspy.OutputField(desc="Email subject")
    body = dspy.OutputField(desc="Email text body")
    attachments: list[str] = dspy.OutputField(desc="Email text attachments")


RAW_EMAIL = """\
From : foo@example.net
To: bar@example.net
Subject: hello

Hello there
-- bar
"""

lm = dspy.LM('gemini/gemini-2.5-flash-lite')
pred = dspy.Predict(signature=EmailParse)
result = pred(lm=lm, raw_email=RAW_EMAIL)
print(dict(result))
print(result.get_lm_usage())

Install

pip install pico-dspy

Install your preferred variant of pydantic-ai, eg. pip install pydantic-ai-slim[google].