import dspy
lm = dspy.LM("gemini/gemini-2.5-flash") # api_key="my_api_key"
lm("Count all the R's in strawberry")['There are **3** R\'s in "strawberry".']DSPy sounds neat, but is it reliable?
2025-09-26
Someone just asked/told me:
DSPy sounds like a neat idea but is it practical?
TL;DR: yes.
My take on reliability and practicality is that, yes, it’s reliable—but that’s the wrong question to ask. The reason I reach for DSPy is first and foremost because it’s the simplest and most ergonomic way I know to call an LLM from Python. It’s the most lightweight way to send a prompt to an LLM, and it also happens to provide the richest tooling and possibilities: optimization of instructions and few-shots, agents, fine-tuning, structured output with validation and retries, and signatures.
Say you just want to send a string to an LLM and be productive. Without much setup or framework, you can just do this:
Nothing more, this is extremely reliable and much more ergonomic and productive than any other AI SDK I have tried (special mention to Claudette, which is also very nice).
But then you want to go from prompt to workflow or AI system. You just move to:
dspy.configure(lm=lm)
prg = dspy.Predict("text_input, letter -> letter_occurence: int")
prg(text_input="snowboarding is cool", letter="o")Prediction(
letter_occurence=3
)That, in my experience, is as good as writing the prompt myself, but already much more general, and much more productive and ergonomic! But then, say you want to optimize it. You add this:
examples = [dspy.Example(
text_input="snowboarding is cool", letter="o",
letter_occurence=4
),
dspy.Example(
text_input="writing the prompt", letter="i",
letter_occurence=2
),
dspy.Example(
text_input="that is extremely reliable", letter="e",
letter_occurence=5
)]
# mark input fields
trainset = [i.with_inputs("text_input", "letter") for i in examples]
def is_equal(gold, pred, _=None):
return gold.letter_occurence == pred.letter_occurence
optimizer = dspy.MIPROv2(metric=is_equal)
prg_opt = optimizer.compile(prg, trainset=trainset)… 2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: ===== Trial 10 / 10 =====
Average Metric: 1.00 / 2 (50.0%): 100%|██████████| 2/2 [00:00<00:00, 293.70it/s] 2025/09/26 21:02:37 INFO dspy.evaluate.evaluate: Average Metric: 1 / 2 (50.0%) 2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: Score: 50.0 with parameters [‘Predictor 0: Instruction 2’, ‘Predictor 0: Few-Shot Set 5’]. 2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: Scores so far: [50.0, 100.0, 100.0, 100.0, 50.0, 100.0, 100.0, 50.0, 100.0, 50.0] 2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: Best score so far: 100.0 2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: =========================
2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: ===== Trial 11 / 10 =====
Average Metric: 2.00 / 2 (100.0%): 100%|██████████| 2/2 [00:00<00:00, 178.00it/s] 2025/09/26 21:02:37 INFO dspy.evaluate.evaluate: Average Metric: 2 / 2 (100.0%) 2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: Score: 100.0 with parameters [‘Predictor 0: Instruction 1’, ‘Predictor 0: Few-Shot Set 3’]. 2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: Scores so far: [50.0, 100.0, 100.0, 100.0, 50.0, 100.0, 100.0, 50.0, 100.0, 50.0, 100.0] 2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: Best score so far: 100.0 2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: =========================
2025/09/26 21:02:37 INFO dspy.teleprompt.mipro_optimizer_v2: Returning best identified program with score 100.0!
Now we can use the optimized program:
prg_opt(text_input="Returning best identified program", letter="e")Prediction(
letter_occurence=4
)That is what DSPy sent to the LLM. For us:
{
"messages": [
{
"role": "system",
"content": [
"Your input fields are:",
"1. `text_input` (str):",
"2. `letter` (str):",
"Your output fields are:",
"1. `letter_occurence` (int):",
"All interactions will be structured in the following way, with the appropriate values filled in.",
"",
"[[ ## text_input ## ]]",
"{text_input}",
"",
"[[ ## letter ## ]]",
"{letter}",
"",
"[[ ## letter_occurence ## ]]",
"{letter_occurence} # note: the value you produce must be a single int value",
"",
"[[ ## completed ## ]]",
"In adhering to this structure, your objective is:",
"Count the occurrences of the specified `letter` within the `text_input` string, and provide this total as `letter_occurence`."
]
},
{
"role": "user",
"content": [
"[[ ## text_input ## ]]",
"snowboarding is cool",
"",
"[[ ## letter ## ]]",
"o"
]
},
{
"role": "assistant",
"content": [
"[[ ## letter_occurence ## ]]",
"4",
"",
"[[ ## completed ## ]]"
]
},
{
"role": "user",
"content": [
"[[ ## text_input ## ]]",
"Returning best identified program",
"",
"[[ ## letter ## ]]",
"e",
"",
"Respond with the corresponding output fields, starting with the field `[[ ## letter_occurence ## ]]` (must be formatted as a valid Python int), and then ending with the marker for `[[ ## completed ## ]]`."
]
},
{
"role": "assistant",
"content": [
"[[ ## letter_occurence ## ]]",
"4",
"",
"[[ ## completed ## ]]"
]
}
]
}The final instruction was:
In adhering to this structure, your objective is: Count the occurrences of the specified
letterwithin thetext_inputstring, and provide this total asletter_occurence.
And the selected few-shot example was this one:
...{"role": "user",
"content": [
"[[ ## text_input ## ]] snowboarding is cool
[[ ## letter ## ]] o"
]},{
"role": "assistant",
"content": [
"[[ ## letter_occurence ## ]] 4
[[ ## completed ## ]]"
]}...As you can see, dspy can be used gradually and is a delight at each step of the way :)