Skip to content

DSPy

Bases: InternalEngine[PromptSignature, Result, Model, InferenceMode]

Engine for DSPy.

Source code in sieves/engines/dspy_.py 
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
class DSPy(InternalEngine[PromptSignature, Result, Model, InferenceMode]):
    """Engine for DSPy."""

    def __init__(
        self,
        model: Model,
        config_kwargs: dict[str, Any],
        init_kwargs: dict[str, Any],
        inference_kwargs: dict[str, Any],
        strict_mode: bool,
        batch_size: int,
    ):
        """
        :param model: Model to run. Note: DSPy only runs with APIs. If you want to run a model locally from v2.5
            onwards, serve it with OLlama - see here: # https://dspy.ai/learn/programming/language_models/?h=models#__tabbed_1_5.
            In a nutshell:
            > curl -fsSL https://ollama.ai/install.sh | sh
            > ollama run MODEL_ID
            > `model = dspy.LM(MODEL_ID, api_base='http://localhost:11434', api_key='')`
        :param config_kwargs: Optional kwargs supplied to dspy.configure().
        :param init_kwargs: Optional kwargs to supply to engine executable at init time.
        :param inference_kwargs: Optional kwargs to supply to engine executable at inference time.
        :param strict_mode: If True, exception is raised if prompt response can't be parsed correctly.
        :param batch_size: Batch size in processing prompts. -1 will batch all documents in one go. Not all engines
            support batching.
        """
        super().__init__(model, init_kwargs, inference_kwargs, strict_mode, batch_size)
        config_kwargs = {"max_tokens": DSPy._MAX_TOKENS} | (config_kwargs or {})
        dspy.configure(lm=model, **config_kwargs)

    @property
    def inference_modes(self) -> type[InferenceMode]:
        return InferenceMode

    @property
    def supports_few_shotting(self) -> bool:
        return True

    def build_executable(
        self,
        inference_mode: InferenceMode,
        prompt_template: str | None,  # noqa: UP007
        prompt_signature: type[PromptSignature] | PromptSignature,
        fewshot_examples: Iterable[pydantic.BaseModel] = tuple(),
    ) -> Executable[Result | None]:
        # Note: prompt_template is ignored here, as DSPy doesn't use it directly (only prompt_signature_description).
        assert isinstance(prompt_signature, type)

        # Handled differently than the other supported modules: dspy.Module serves as both the signature as well as
        # the inference generator.
        if inference_mode == InferenceMode.module:
            assert isinstance(prompt_signature, dspy.Module), ValueError(
                "In inference mode 'module' the provided prompt signature has to be of type dspy.Module."
            )
            generator = inference_mode.value(**self._init_kwargs)
        else:
            assert issubclass(prompt_signature, dspy.Signature)
            generator = inference_mode.value(signature=prompt_signature, **self._init_kwargs)

        def execute(values: Iterable[dict[str, Any]]) -> Iterable[Result | None]:
            # Compile predictor with few-shot examples.
            fewshot_examples_dicts = DSPy._convert_fewshot_examples(fewshot_examples)
            generator_fewshot: dspy.Module | None = None
            if len(fewshot_examples_dicts):
                examples = [dspy.Example(**fs_example) for fs_example in fewshot_examples_dicts]
                generator_fewshot = dspy.LabeledFewShot(k=5).compile(student=generator, trainset=examples)
            generator_async = dspy.asyncify(generator_fewshot or generator)

            batch_size = self._batch_size if self._batch_size != -1 else sys.maxsize
            # Ensure values are read as generator for standardized batch handling (otherwise we'd have to use
            # different batch handling depending on whether lists/tuples or generators are used).
            values = (v for v in values)

            while batch := [vals for vals in itertools.islice(values, batch_size)]:
                if len(batch) == 0:
                    break

                try:
                    calls = [generator_async(**doc_values, **self._inference_kwargs) for doc_values in batch]
                    yield from asyncio.run(self._execute_async_calls(calls))

                except ValueError as err:
                    if self._strict_mode:
                        raise ValueError(
                            "Encountered problem when executing prompt. Ensure your few-shot examples and document "
                            "chunks contain sensible information."
                        ) from err
                    else:
                        yield from [None] * len(batch)

        # return functools.partial(execute, generator=inference_generator)
        return execute

model property

Return model instance.

Returns:

Type Description
EngineModel

Model instance.

__init__(model, config_kwargs, init_kwargs, inference_kwargs, strict_mode, batch_size)

Parameters:

Name Type Description Default
model Model

Model to run. Note: DSPy only runs with APIs. If you want to run a model locally from v2.5 onwards, serve it with OLlama - see here: # https://dspy.ai/learn/programming/language_models/?h=models#__tabbed_1_5. In a nutshell: > curl -fsSL https://ollama.ai/install.sh | sh > ollama run MODEL_ID > model = dspy.LM(MODEL_ID, api_base='http://localhost:11434', api_key='')

 required
config_kwargs dict[str, Any]

Optional kwargs supplied to dspy.configure().

 required
init_kwargs dict[str, Any]

Optional kwargs to supply to engine executable at init time.

 required
inference_kwargs dict[str, Any]

Optional kwargs to supply to engine executable at inference time.

 required
strict_mode bool

If True, exception is raised if prompt response can't be parsed correctly.

 required
batch_size int

Batch size in processing prompts. -1 will batch all documents in one go. Not all engines support batching.

 required
Source code in sieves/engines/dspy_.py 
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
def __init__(
    self,
    model: Model,
    config_kwargs: dict[str, Any],
    init_kwargs: dict[str, Any],
    inference_kwargs: dict[str, Any],
    strict_mode: bool,
    batch_size: int,
):
    """
    :param model: Model to run. Note: DSPy only runs with APIs. If you want to run a model locally from v2.5
        onwards, serve it with OLlama - see here: # https://dspy.ai/learn/programming/language_models/?h=models#__tabbed_1_5.
        In a nutshell:
        > curl -fsSL https://ollama.ai/install.sh | sh
        > ollama run MODEL_ID
        > `model = dspy.LM(MODEL_ID, api_base='http://localhost:11434', api_key='')`
    :param config_kwargs: Optional kwargs supplied to dspy.configure().
    :param init_kwargs: Optional kwargs to supply to engine executable at init time.
    :param inference_kwargs: Optional kwargs to supply to engine executable at inference time.
    :param strict_mode: If True, exception is raised if prompt response can't be parsed correctly.
    :param batch_size: Batch size in processing prompts. -1 will batch all documents in one go. Not all engines
        support batching.
    """
    super().__init__(model, init_kwargs, inference_kwargs, strict_mode, batch_size)
    config_kwargs = {"max_tokens": DSPy._MAX_TOKENS} | (config_kwargs or {})
    dspy.configure(lm=model, **config_kwargs)

deserialize(config, **kwargs) classmethod

Generate Engine instance from config.

Parameters:

Name Type Description Default
config Config

Config to generate instance from.

 required
kwargs dict[str, Any]

Values to inject into loaded config.

 {}

Returns:

Type Description
InternalEngine[EnginePromptSignature, EngineResult, EngineModel, EngineInferenceMode]

Deserialized Engine instance.
Source code in sieves/engines/core.py 
122
123
124
125
126
127
128
129
130
131
@classmethod
def deserialize(
    cls, config: Config, **kwargs: dict[str, Any]
) -> InternalEngine[EnginePromptSignature, EngineResult, EngineModel, EngineInferenceMode]:
    """Generate Engine instance from config.
    :param config: Config to generate instance from.
    :param kwargs: Values to inject into loaded config.
    :return: Deserialized Engine instance.
    """
    return cls(**config.to_init_dict(cls, **kwargs))

serialize()

Serializes engine.

Returns:

Type Description
Config

Config instance.
Source code in sieves/engines/core.py 
116
117
118
119
120
def serialize(self) -> Config:
    """Serializes engine.
    :return: Config instance.
    """
    return Config.create(self.__class__, self._attributes)