DSPy

​

Install

pip install trulayer dspy

​

Instrument

import os
import dspy
import trulayer
from trulayer.instruments.dspy import instrument_dspy

trulayer.init(api_key=os.environ["TRULAYER_API_KEY"], project_name="my-app")

with trulayer.trace("dspy-run") as trace:
    instrument_dspy(trace)

    dspy.configure(lm=dspy.LM("openai/gpt-4o-mini"))
    qa = dspy.Predict("question -> answer")
    result = qa(question="What is the capital of France?")
    print(result.answer)

​

What gets captured

• One llm span per dspy.Predict.forward() call with the rendered prompt as input and the parsed output fields (e.g. answer) as output.
• Token counts, model name, and latency pulled from the underlying dspy.LM call.
• Errors from the LM or from DSPy’s field parsing surface as span.status = error.

​

Disabling

from trulayer.instruments.dspy import uninstrument_dspy

uninstrument_dspy()

​

Known gotchas

• Global patch. DSPy doesn’t expose a per-instance hook, so instrumentation patches dspy.Predict.forward globally. Safe to call multiple times — subsequent calls are no-ops.
• Compile-time calls are traced too. If you run dspy.teleprompt optimizers while instrumented, every candidate prompt evaluation shows up as a span. Consider calling uninstrument_dspy() during compilation and re-instrumenting at serve time.
• Async. DSPy’s async support is experimental; the current instrumentation is sync-only.