[FR]: optimizing prompts for custom models

[taborzbislaw]

Proposal summary

I want to optimize prompts for som ecustom model which returns structured response e.g.:

class ResponseSchema(BaseModel): query: List[str] = Field( ..., description="Some list" )

class CustomModel(OpikBaseModel): def init(self, MODEL:str, TEMP:float, CONTEXT:int, Response:BaseModel): super().init(model_name = None) self.MODEL_NAME = MODEL self.TEMPERATURE = TEMP self.CONTEXT_LENGTH = CONTEXT self.Response = Response

def generate_provider_response(self, **kwargs: Any) -> str:
    pass
    
def agenerate_provider_response(self, **kwargs: Any) -> str:
    pass
    
def agenerate_provider_response_stream(self, **kwargs: Any) -> str:
    pass
    
def generate_string(self, input: str, **kwargs: Any) -> str:
    """Simplified interface to generate a string output from the model."""
    #print(input)
    
    response = ollama.chat(
      messages=[
        {
          'role': 'user',
          'content': input,
        }
      ],
      model = self.MODEL_NAME,
      format = self.Response.model_json_schema(),
      options = {"context_length": self.CONTEXT_LENGTH,"temperature":self.TEMPERATURE}
    )
    
    return response["message"]["content"]

def agenerate_string(self, input: str, **kwargs: Any) -> str:
    pass
    
def agenerate_prompt(self, input: str, **kwargs: Any) -> str:
    pass

my_custom_model = CustomModel('qwen3:235b-a22b',0,4096,ResponseSchema)

Currently, only strings can be passed to optimizer constructors, like:

optimizer = FewShotBayesianOptimizer( project_name = SOME_PROJECT_NAME, model = "openai/gpt-4o-mini", temperature=0.1, max_tokens=5000, )

while desired option would be like that:

optimizer = FewShotBayesianOptimizer( project_name = SOME_PROJECT_NAME, model = my_custom_model )

Motivation

Structured output is currently a standard. When passing only a string with model name, it is not possible, using prompt only, to force a model to return structured output in all cases. For this reason application of opik prompt optimization in its current version is seriously limited.

[jverre]

Hi @taborzbislaw

Great suggestion, we support this in evaluations but not in the optimizer - I'll add it today

[taborzbislaw]

Hi,

thank you very much for your prompt answer. I am using custom models in evaluations and it works very well. Just one more question. When optimizing prompts I prepare TaskConfig like:

task_config = TaskConfig(
    instruction_prompt=initial_prompt,
    input_dataset_fields=["some_query"],
    output_dataset_field="expected_output",
    use_chat_prompt=True,
)

where initial_prompt must be string.

Optimization of prompt templates would be also a required functionality, extending applications of opik oprimizers. I am thinking about something like:

client = opik.Opik()
prompt_template = client.get_prompt(name="prompt_template")

prompt template could be like:

Here is a question: {{question}}.
Here is some context {{context}}
Answer the question given the context

task_config = TaskConfig(
    instruction_prompt = prompt_template,
    prompt_input_dataset_fields = {"question": "some_dataset_field1", "context":"some_dataset_field2"}
    output_dataset_field="expected_output",
)

So, again, things like the ones above are available for evaluations so extending them to optimizers would be really valuable.

[teenaxta]

hello any update not this?

[dsblank]

Hello @teenaxta and @taborzbislaw,

We are busy designing updates to the Opik optimizer system that will allow such variations. This is a big project; stay tuned!

[jverre]

Hi @teenaxta, @taborzbislaw

We've finished the refactoring and now support optimizing structured outputs, I've updated the docs here: https://www.comet.com/docs/opik/agent_optimization/opik_optimizer/models#customizing-the-llm-response

Let me know if you have any feedback

[taborzbislaw]

Hi,

can you, please, give some working example of optimization using structured output? There is no complete code in opik webpage so I was trying to compose thr code from snippets at different pages. I tried code like this (with a few variants):

###########################################
from pydantic import BaseModel
from openai import OpenAI
from opik_optimizer import ChatPrompt

from opik.evaluation.metrics import LevenshteinRatio
from opik_optimizer import MetaPromptOptimizer, ChatPrompt
from opik_optimizer.datasets import tiny_test
import opik_optimizer

client = OpenAI()

class AnswerModel(BaseModel):
    answer: str

def invoke_llm(model, messages, tools):
    completion = client.chat.completions.parse(
        model=model,
        messages=messages,
        response_format=AnswerModel,
    )

    return completion.choices[0].message.parse

prompt = ChatPrompt(
  name="demo_prompt",
  model="openai/gpt-4o",
  messages=[
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "{question}"}
  ],
  invoke=invoke_llm
)

# You can use a demo dataset for testing, or your own dataset
dataset = tiny_test()
print(f"Using dataset: {dataset.name}, with {len(dataset.get_items())} items.")

# This example uses Levenshtein distance to measure output quality
def levenshtein_ratio(dataset_item, llm_output):
    metric = LevenshteinRatio()
    return metric.score(reference=dataset_item['label'], output=llm_output)

print("Starting optimization...")
result = optimizer.optimize_prompt(
    prompt=prompt,
    dataset=dataset,
    metric=levenshtein_ratio,
)

##################################################

When running, I got errors like:

Starting optimization...
╭────────────────────────────────────────────────────────────────────╮
│ ● Running Opik Evaluation - MetaPromptOptimizer                    │
│                                                                    │
│ -> View optimization details ]8;id=555929;http://localhost:5173/api/v1/session/redirect/optimizations/?optimization_id=0197ffef-67ba-7444-8056-dc209e1a9af7&dataset_id=01970e14-9fab-7307-87ad-d4faa00fc7c5&path=aHR0cDovL2xvY2FsaG9zdDo1MTczL2FwaS8=\in your Opik dashboard]8;;\                │
╰────────────────────────────────────────────────────────────────────╯


> Let's optimize the prompt:

╭─ system ───────────────────────────────────────────────────────────╮
│                                                                    │
│  You are a helpful assistant.                                      │
│                                                                    │
╰────────────────────────────────────────────────────────────────────╯
╭─ user ─────────────────────────────────────────────────────────────╮
│                                                                    │
│  {question}                                                        │
│                                                                    │
╰────────────────────────────────────────────────────────────────────╯

Using MetaPromptOptimizer with the parameters: 
  - n_samples: None
  - auto_continue: False


> First we will establish the baseline performance:

  Evaluation ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━   0% -:--:--

[2025-07-12 20:39:30] ERROR    Error calling model with prompt: 'Completions' object   [meta_prompt_optimizer.py](file:///home/zt/azai/lib/python3.11/site-packages/opik_optimizer/meta_prompt_optimizer/meta_prompt_optimizer.py):[312](file:///home/zt/azai/lib/python3.11/site-packages/opik_optimizer/meta_prompt_optimizer/meta_prompt_optimizer.py#312)
                               has no attribute 'parse'                                                            

                      ERROR    Failed prompt: [{'role': 'system', 'content': 'You are  [meta_prompt_optimizer.py](file:///home/zt/azai/lib/python3.11/site-packages/opik_optimizer/meta_prompt_optimizer/meta_prompt_optimizer.py):[313](file:///home/zt/azai/lib/python3.11/site-packages/opik_optimizer/meta_prompt_optimizer/meta_prompt_optimizer.py#313)
                               a helpful assistant.'}, {'role': 'user', 'content':                                 
                               '{question}'}]                                                                      

                      ERROR    Prompt length: 38                                       [meta_prompt_optimizer.py](file:///home/zt/azai/lib/python3.11/site-packages/opik_optimizer/meta_prompt_optimizer/meta_prompt_optimizer.py):[314](file:///home/zt/azai/lib/python3.11/site-packages/opik_optimizer/meta_prompt_optimizer/meta_prompt_optimizer.py#314)

                      ERROR    Error calling model with prompt: 'Completions' object   [meta_prompt_optimizer.py](file:///home/zt/azai/lib/python3.11/site-packages/opik_optimizer/meta_prompt_optimizer/meta_prompt_optimizer.py):[312](file:///home/zt/azai/lib/python3.11/site-packages/opik_optimizer/meta_prompt_optimizer/meta_prompt_optimizer.py#312)
                               has no attribute 'parse'

Thank you in advance

[dsblank]

Your invoke_llm() function should return a string. Does completion.choices[0].message have a property parse? It doesn't look like it from the error message.

[taborzbislaw]

Thank you for the hint "invoke_llm() function should return a string". Now optimization works

[dsblank]

Glad that worked!