Skip to content

Relation Extraction

The RelationExtraction task performs joint entity and relation extraction, identifying relationships between entities in text.

Usage

relations = {
    "works_for": "A person works for a company or organization.",
    "located_in": "A place or organization is located in a city, country, or region.",
    "founded": "A person founded a company or organization.",
}

fewshot_examples = [
    relation_extraction.FewshotExample(
        text="Henri Dunant founded the Red Cross in Geneva.",
        triplets=[
            RelationTriplet(
                head=RelationEntity(text="Henri Dunant", entity_type="PERSON"),
                relation="founded",
                tail=RelationEntity(text="Red Cross", entity_type="ORGANIZATION"),
                score=1.0,
            ),
            RelationTriplet(
                head=RelationEntity(text="Red Cross", entity_type="ORGANIZATION"),
                relation="located_in",
                tail=RelationEntity(text="Geneva", entity_type="LOCATION"),
                score=1.0,
            ),
        ],
    ),
    relation_extraction.FewshotExample(
        text="Eglantyne Jebb founded Save the Children in London.",
        triplets=[
            RelationTriplet(
                head=RelationEntity(text="Eglantyne Jebb", entity_type="PERSON"),
                relation="founded",
                tail=RelationEntity(text="Save the Children", entity_type="ORGANIZATION"),
                score=1.0,
            ),
            RelationTriplet(
                head=RelationEntity(text="Save the Children", entity_type="ORGANIZATION"),
                relation="located_in",
                tail=RelationEntity(text="London", entity_type="LOCATION"),
                score=1.0,
            ),
        ],
    ),
]

fewshot_args = {"fewshot_examples": fewshot_examples} if fewshot else {}

task = relation_extraction.RelationExtraction(
    relations=relations,
    model=batch_runtime.model,
    model_settings=batch_runtime.model_settings,
    batch_size=batch_runtime.batch_size,
    entity_types=["PERSON", "ORGANIZATION", "LOCATION"],
    **fewshot_args
)

pipe = Pipeline(task)
docs = list(pipe(relation_extraction_docs))

Results

The RelationExtraction task returns a unified Result object containing a list of RelationTriplet objects.

Each triplet includes a confidence score: - GLiNER2: Always present and derived from logits. - LLMs: Self-reported and may be None if not provided by the model.

class Result(pydantic.BaseModel):
    """Result of a relation extraction task.

    Attributes:
        triplets: List of extracted relation triplets.
    """

    triplets: list[RelationTriplet]

Each RelationTriplet consists of: - head: A RelationEntity representing the subject. - relation: The string identifier of the relationship. - tail: A RelationEntity representing the object.

A RelationEntity includes the surface text, entity_type, and character start/end offsets.

Evaluation

Performance of the relation extraction task can be measured using the .evaluate() method.

  • Metric: Corpus-wide Micro-F1 Score (F1). Triplets are matched based on the head entity text, the relation type, and the tail entity text.
  • Requirement: Each document must have ground-truth triplets stored in doc.gold[task_id].
report = task.evaluate(docs)
print(f"Relation F1-Score: {report.metrics['F1']}")

Ground Truth Formats

Ground truth has to be specified in doc.meta using Result instances.

Relation extraction predictive task.

RelationExtraction

Bases: PredictiveTask[TaskPromptSignature, TaskResult, _TaskBridge]

Extract relations between entities in text.

Source code in sieves/tasks/predictive/relation_extraction/core.py 
 32
 33
 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
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
class RelationExtraction(PredictiveTask[TaskPromptSignature, TaskResult, _TaskBridge]):
    """Extract relations between entities in text."""

    def __init__(
        self,
        relations: Sequence[str] | dict[str, str],
        model: TaskModel,
        entity_types: Sequence[str] | dict[str, str] | None = None,
        task_id: str | None = None,
        include_meta: bool = True,
        batch_size: int = -1,
        prompt_instructions: str | None = None,
        fewshot_examples: Sequence[FewshotExample] = (),
        model_settings: ModelSettings = ModelSettings(),
        condition: Callable[[Doc], bool] | None = None,
    ) -> None:
        """Initialize RelationExtraction task.

        :param relations: Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.
        :param model: Model to use.
        :param entity_types: Optional constraints on entity types involved in relations.
        :param task_id: Task ID.
        :param include_meta: Whether to include meta information generated by the task.
        :param batch_size: Batch size to use for inference. Use -1 to process all documents at once.
        :param prompt_instructions: Custom prompt instructions. If None, default instructions are used.
        :param fewshot_examples: Few-shot examples.
        :param model_settings: Settings for structured generation.
        :param condition: Optional callable that determines whether to process each document.
        """
        self._relations = relations
        self._entity_types = entity_types

        super().__init__(
            model=model,
            task_id=task_id,
            include_meta=include_meta,
            batch_size=batch_size,
            overwrite=False,
            prompt_instructions=prompt_instructions,
            fewshot_examples=fewshot_examples,
            model_settings=model_settings,
            condition=condition,
        )

    @property
    @override
    def fewshot_example_type(self) -> type[FewshotExample]:
        """Return few-shot example type.

        :return: Few-shot example type.
        """
        return FewshotExample

    @property
    @override
    def prompt_signature(self) -> type[pydantic.BaseModel]:
        """Return the unified Pydantic prompt signature for this task.

        :return: Unified Pydantic prompt signature.
        """
        # Define relation type literal if relations are provided.
        if isinstance(self._relations, dict):
            relation_names = list(self._relations.keys())
        else:
            relation_names = list(self._relations)

        RelationType = Literal[*(tuple(relation_names))] if relation_names else str  # type: ignore[invalid-type-form]

        # Define entity type literal if entity types are provided.
        entity_type_names: list[str] = []
        if self._entity_types:
            if isinstance(self._entity_types, dict):
                entity_type_names = list(self._entity_types.keys())
            else:
                entity_type_names = list(self._entity_types)

        EntityType = Literal[*(tuple(entity_type_names))] if entity_type_names else str  # type: ignore[invalid-type-form]

        # Create dynamic models.
        DynamicEntity = pydantic.create_model(
            "RelationEntity",
            text=(str, pydantic.Field(..., description="Surface text of the entity as it appears in the document.")),
            entity_type=(
                EntityType,
                pydantic.Field(..., description="The category or type of the entity."),
            ),
            __doc__="An entity involved in a relation.",
            __base__=pydantic.BaseModel,
        )

        DynamicTriplet = pydantic.create_model(
            "RelationTriplet",
            head=(DynamicEntity, pydantic.Field(..., description="The subject entity (head) of the relation.")),
            relation=(RelationType, pydantic.Field(..., description="The type of relation between the head and tail.")),
            tail=(DynamicEntity, pydantic.Field(..., description="The object entity (tail) of the relation.")),
            score=(
                float | None,
                pydantic.Field(
                    default=None, description="Provide a confidence score for this relation triplet, between 0 and 1."
                ),
            ),
            __doc__="A relation triplet consisting of a head entity, a relation type, and a tail entity.",
            __base__=pydantic.BaseModel,
        )

        return pydantic.create_model(
            "RelationExtractionOutput",
            __doc__="Result of relation extraction. Contains a list of extracted relation triplets.",
            triplets=(
                list[DynamicTriplet],  # type: ignore[invalid-type-form]
                pydantic.Field(..., description="List of extracted relation triplets."),
            ),
        )

    @property
    @override
    def metric(self) -> str:
        return "F1"

    @override
    def _compute_metrics(self, truths: list[Any], preds: list[Any], judge: dspy.LM | None = None) -> dict[str, float]:
        """Compute corpus-level metrics.

        :param truths: List of ground truths.
        :param preds: List of predictions.
        :param judge: Optional DSPy LM instance to use as judge for generative tasks.
        :return: Dictionary of metrics.
        """
        tp = 0
        fp = 0
        fn = 0

        for gold, pred in zip(truths, preds):
            if gold is not None:
                assert isinstance(gold, TaskResult)
                true_triplets = {(t.head.text.lower(), t.relation.lower(), t.tail.text.lower()) for t in gold.triplets}
            else:
                true_triplets = set()

            if pred is not None:
                assert isinstance(pred, TaskResult)
                pred_triplets = {(t.head.text.lower(), t.relation.lower(), t.tail.text.lower()) for t in pred.triplets}
            else:
                pred_triplets = set()

            tp += len(true_triplets & pred_triplets)
            fp += len(pred_triplets - true_triplets)
            fn += len(true_triplets - pred_triplets)

        precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
        recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
        f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0

        return {self.metric: f1}

    @override
    def _init_bridge(self, model_type: ModelType) -> _TaskBridge:
        if model_type == ModelType.gliner:
            if self._entity_types is not None:
                warnings.warn(
                    "GliNER2 backend does not support entity type constraints for relation extraction. "
                    "The `entity_types` parameter will be ignored.",
                )

            return GliNERBridge(
                task_id=self._task_id,
                prompt_instructions=self._custom_prompt_instructions,
                prompt_signature=self.prompt_signature,
                model_settings=self._model_settings,
                inference_mode=gliner_.InferenceMode.relations,
            )

        bridge_types: dict[ModelType, type[DSPyRelationExtraction | PydanticRelationExtraction]] = {
            ModelType.dspy: DSPyRelationExtraction,
            ModelType.langchain: PydanticRelationExtraction,
            ModelType.outlines: PydanticRelationExtraction,
        }

        try:
            return bridge_types[model_type](
                task_id=self._task_id,
                relations=self._relations,
                entity_types=self._entity_types,
                prompt_instructions=self._custom_prompt_instructions,
                model_settings=self._model_settings,
                prompt_signature=self.prompt_signature,
                model_type=model_type,
                fewshot_examples=self._fewshot_examples,
            )
        except KeyError as err:
            raise KeyError(f"Model type {model_type} is not supported by {self.__class__.__name__}.") from err

    @staticmethod
    @override
    def supports() -> set[ModelType]:
        return {
            ModelType.dspy,
            ModelType.gliner,
            ModelType.langchain,
            ModelType.outlines,
        }

    @override
    @property
    def _state(self) -> dict[str, Any]:
        return {
            **super()._state,
            "relations": self._relations,
            "entity_types": self._entity_types,
        }

    @override
    def to_hf_dataset(self, docs: Iterable[Doc], threshold: float | None = None) -> datasets.Dataset:
        # Define metadata and features.
        entity_feature = datasets.Features(
            {
                "text": datasets.Value("string"),
                "entity_type": datasets.Value("string"),
            }
        )
        triplet_feature = datasets.Features(
            {
                "head": entity_feature,
                "relation": datasets.Value("string"),
                "tail": entity_feature,
                "score": datasets.Value("float32"),
            }
        )
        features = datasets.Features({"text": datasets.Value("string"), "triplets": datasets.Sequence(triplet_feature)})
        info = datasets.DatasetInfo(
            description=f"Relation extraction dataset. Generated with sieves v{Config.get_version()}.",
            features=features,
        )

        # Fetch data used for generating dataset.
        try:
            data: list[dict[str, Any]] = []
            for doc in docs:
                triplets: list[dict[str, Any]] = []
                for triplet in doc.results[self._task_id].triplets:
                    triplets.append(
                        {
                            "head": triplet.head.model_dump(),
                            "relation": triplet.relation,
                            "tail": triplet.tail.model_dump(),
                            "score": triplet.score,
                        }
                    )
                data.append({"text": doc.text, "triplets": triplets})
        except KeyError as err:
            raise KeyError(f"Not all documents have results for this task with ID {self._task_id}") from err

        # Create dataset.
        return datasets.Dataset.from_list(data, features=features, info=info)

    @override
    def distill(
        self,
        base_model_id: str,
        framework: DistillationFramework,
        data: datasets.Dataset | Sequence[Doc],
        output_path: Path | str,
        val_frac: float,
        init_kwargs: dict[str, Any] | None = None,
        train_kwargs: dict[str, Any] | None = None,
        seed: int | None = None,
    ) -> None:
        raise NotImplementedError

    @override
    def _evaluate_dspy_example(self, truth: dspy.Example, pred: dspy.Prediction, trace: Any, model: dspy.LM) -> float:
        # Compute triplet-level F1 score based on (head_text, relation, tail_text) triples.
        # Use lowercase for robust matching.
        true_triplets = {
            (t["head"]["text"].lower(), t["relation"].lower(), t["tail"]["text"].lower()) for t in truth["triplets"]
        }
        pred_triplets = {
            (t["head"]["text"].lower(), t["relation"].lower(), t["tail"]["text"].lower())
            for t in pred.get("triplets", [])
        }

        if not true_triplets:
            base_f1 = 1.0 if not pred_triplets else 0.0
        else:
            precision = len(true_triplets & pred_triplets) / len(pred_triplets) if pred_triplets else 0
            recall = len(true_triplets & pred_triplets) / len(true_triplets)
            base_f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0

        # If score is available, incorporate it.
        if "triplets" in truth and "triplets" in pred:
            true_scores = [t.get("score") for t in truth["triplets"] if t.get("score") is not None]
            pred_scores = [t.get("score") for t in pred["triplets"] if t.get("score") is not None]

            if true_scores and pred_scores:
                true_avg = sum(true_scores) / len(true_scores)
                pred_avg = sum(pred_scores) / len(pred_scores)
                score_accuracy = 1 - abs(true_avg - max(min(pred_avg, 1), 0))
                return (base_f1 + score_accuracy) / 2

        return base_f1

fewshot_example_type property

Return few-shot example type.

Returns:

Type Description
type[FewshotExample]

Few-shot example type.

fewshot_examples property

Return few-shot examples.

Returns:

Type Description
Sequence[FewshotExample]

Few-shot examples.

id property

Return task ID.

Used by pipeline for results and dependency management.

Returns:

Type Description
str

Task ID.

prompt_signature property

Return the unified Pydantic prompt signature for this task.

Returns:

Type Description
type[BaseModel]

Unified Pydantic prompt signature.

prompt_signature_description property

Return prompt signature description.

Returns:

Type Description
str | None

Prompt signature description.

prompt_template property

Return prompt template.

Returns:

Type Description
str

Prompt template.

__add__(other)

Chain this task with another task or pipeline using the + operator.

This returns a new Pipeline that executes this task first, followed by the task(s) in other. The original task(s)/pipeline are not mutated.

Cache semantics: - If other is a Pipeline, the resulting pipeline adopts other's use_cache setting (because the left-hand side is a single task). - If other is a Task, the resulting pipeline defaults to use_cache=True.

Parameters:

Name Type Description Default
other Task | Pipeline

A Task or Pipeline to execute after this task.

 required

Returns:

Type Description
Pipeline

A new Pipeline representing the chained execution.

Raises:

Type Description
TypeError

If other is not a Task or Pipeline.
Source code in sieves/tasks/core.py 
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
def __add__(self, other: Task | Pipeline) -> Pipeline:
    """Chain this task with another task or pipeline using the ``+`` operator.

    This returns a new ``Pipeline`` that executes this task first, followed by the
    task(s) in ``other``. The original task(s)/pipeline are not mutated.

    Cache semantics:
    - If ``other`` is a ``Pipeline``, the resulting pipeline adopts ``other``'s
      ``use_cache`` setting (because the left-hand side is a single task).
    - If ``other`` is a ``Task``, the resulting pipeline defaults to ``use_cache=True``.

    :param other: A ``Task`` or ``Pipeline`` to execute after this task.
    :return: A new ``Pipeline`` representing the chained execution.
    :raises TypeError: If ``other`` is not a ``Task`` or ``Pipeline``.
    """
    # Lazy import to avoid circular dependency at module import time.
    from sieves.pipeline import Pipeline

    if isinstance(other, Pipeline):
        return Pipeline(tasks=[self, *other.tasks], use_cache=other.use_cache)

    if isinstance(other, Task):
        return Pipeline(tasks=[self, other])

    raise TypeError(f"Cannot chain Task with {type(other).__name__}")

__call__(docs)

Execute task with conditional logic.

Checks the condition for each document without materializing all docs upfront. Passes all documents that pass the condition to _call() for proper batching. Documents that fail the condition have results[task_id] set to None.

Parameters:

Name Type Description Default
docs Iterable[Doc]

Docs to process.

 required

Returns:

Type Description
Iterable[Doc]

Processed docs (in original order).
Source code in sieves/tasks/core.py 
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
def __call__(self, docs: Iterable[Doc]) -> Iterable[Doc]:
    """Execute task with conditional logic.

    Checks the condition for each document without materializing all docs upfront.
    Passes all documents that pass the condition to _call() for proper batching.
    Documents that fail the condition have results[task_id] set to None.

    :param docs: Docs to process.
    :return: Processed docs (in original order).
    """
    docs = iter(docs) if not isinstance(docs, Iterator) else docs

    # Materialize docs in batches. This doesn't incur additional memory overhead, as docs are materialized in
    # batches downstream anyway.
    batch_size = self._batch_size if self._batch_size > 0 else sys.maxsize
    while docs_batch := [doc for doc in itertools.islice(docs, batch_size)]:
        # First pass: determine which docs pass the condition by index.
        passing_indices: set[int] = {
            idx for idx, doc in enumerate(docs_batch) if self._condition is None or self._condition(doc)
        }

        # Process all passing docs in one batch.
        processed = self._call(d for i, d in enumerate(docs_batch) if i in passing_indices)
        processed_iter = iter(processed) if not isinstance(processed, Iterator) else processed

        # Iterate through original docs in order and yield results.
        for idx, doc in enumerate(docs_batch):
            if idx in passing_indices:
                # Doc passed condition - use processed result.
                yield next(processed_iter)
            else:
                # Doc failed condition - set `None` result and yield original.
                doc.results[self.id] = None
                yield doc

__init__(relations, model, entity_types=None, task_id=None, include_meta=True, batch_size=-1, prompt_instructions=None, fewshot_examples=(), model_settings=ModelSettings(), condition=None)

Initialize RelationExtraction task.

Parameters:

Name Type Description Default
relations Sequence[str] | dict[str, str]

Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.

 required
model TaskModel

Model to use.

 required
entity_types Sequence[str] | dict[str, str] | None

Optional constraints on entity types involved in relations.

 None
task_id str | None

Task ID.

 None
include_meta bool

Whether to include meta information generated by the task.

 True
batch_size int

Batch size to use for inference. Use -1 to process all documents at once.

 -1
prompt_instructions str | None

Custom prompt instructions. If None, default instructions are used.

 None
fewshot_examples Sequence[FewshotExample]

Few-shot examples.

 ()
model_settings ModelSettings

Settings for structured generation.

 ModelSettings()
condition Callable[[Doc], bool] | None

Optional callable that determines whether to process each document.

 None
Source code in sieves/tasks/predictive/relation_extraction/core.py 
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
def __init__(
    self,
    relations: Sequence[str] | dict[str, str],
    model: TaskModel,
    entity_types: Sequence[str] | dict[str, str] | None = None,
    task_id: str | None = None,
    include_meta: bool = True,
    batch_size: int = -1,
    prompt_instructions: str | None = None,
    fewshot_examples: Sequence[FewshotExample] = (),
    model_settings: ModelSettings = ModelSettings(),
    condition: Callable[[Doc], bool] | None = None,
) -> None:
    """Initialize RelationExtraction task.

    :param relations: Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.
    :param model: Model to use.
    :param entity_types: Optional constraints on entity types involved in relations.
    :param task_id: Task ID.
    :param include_meta: Whether to include meta information generated by the task.
    :param batch_size: Batch size to use for inference. Use -1 to process all documents at once.
    :param prompt_instructions: Custom prompt instructions. If None, default instructions are used.
    :param fewshot_examples: Few-shot examples.
    :param model_settings: Settings for structured generation.
    :param condition: Optional callable that determines whether to process each document.
    """
    self._relations = relations
    self._entity_types = entity_types

    super().__init__(
        model=model,
        task_id=task_id,
        include_meta=include_meta,
        batch_size=batch_size,
        overwrite=False,
        prompt_instructions=prompt_instructions,
        fewshot_examples=fewshot_examples,
        model_settings=model_settings,
        condition=condition,
    )

deserialize(config, **kwargs) classmethod

Generate PredictiveTask 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
PredictiveTask[TaskPromptSignature, TaskResult, TaskBridge]

Deserialized PredictiveTask instance.
Source code in sieves/tasks/predictive/core.py 
256
257
258
259
260
261
262
263
264
265
266
267
268
269
@classmethod
def deserialize(
    cls, config: Config, **kwargs: dict[str, Any]
) -> PredictiveTask[TaskPromptSignature, TaskResult, TaskBridge]:
    """Generate PredictiveTask instance from config.

    :param config: Config to generate instance from.
    :param kwargs: Values to inject into loaded config.
    :return PredictiveTask[TaskPromptSignature, TaskResult, _TaskBridge]: Deserialized PredictiveTask instance.
    """
    init_dict = config.to_init_dict(cls, **kwargs)
    init_dict["model_settings"] = ModelSettings.model_validate(init_dict["model_settings"])

    return cls(**init_dict)

evaluate(docs, judge=None, failure_threshold=0.5)

Evaluate task performance using DSPy-based evaluation.

Parameters:

Name Type Description Default
docs Iterable[Doc]

Documents to evaluate.

 required
judge LM | None

Optional DSPy LM instance to use as judge for generative tasks.

 None
failure_threshold float

Decision threshold for whether to mark predicitions as failures.

 0.5

Returns:

Type Description
TaskEvaluationReport

Evaluation report.
Source code in sieves/tasks/predictive/core.py 
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
def evaluate(
    self, docs: Iterable[Doc], judge: dspy.LM | None = None, failure_threshold: float = 0.5
) -> TaskEvaluationReport:
    """Evaluate task performance using DSPy-based evaluation.

    :param docs: Documents to evaluate.
    :param judge: Optional DSPy LM instance to use as judge for generative tasks.
    :param failure_threshold: Decision threshold for whether to mark predicitions as failures.
    :return: Evaluation report.
    """
    truths: list[Any] = []
    preds: list[Any] = []
    failures: list[Doc] = []

    # Evaluate each doc individually to identify failed predictions.
    for doc in docs:
        if self.id not in doc.results:
            continue

        pred = doc.results[self.id]
        gold = doc.gold.get(self.id, None)

        # Accumulate for corpus-level metrics.
        truths.append(gold)
        preds.append(pred)

        # If gold or prediction is None: we cannot do proper evalution, so we just check whether they're both None
        # to compute score for failure analysis.
        if gold is None or pred is None:
            if gold is not None or pred is not None:
                failures.append(doc)
        else:
            # Convert result and gold to DSPy representation.
            truth = dspy.Example(**self._task_result_to_dspy_dict(gold))
            pred_dspy = dspy.Prediction(**self._task_result_to_dspy_dict(pred))

            # Call internal evaluation logic for per-doc failure analysis.
            score = self._evaluate_dspy_example(truth, pred_dspy, trace=None, model=judge)

            if score < failure_threshold:
                failures.append(doc)

    # Evaluate on corpus level to obtain representative metrics.
    metrics = self._compute_metrics(truths, preds, judge=judge)

    return TaskEvaluationReport(
        metrics=metrics,
        task_id=self.id,
        failures=failures,
    )

optimize(optimizer, verbose=True)

Optimize task prompt and few-shot examples with the available optimization config.

Updates task to use best prompt and few-shot examples found by the optimizer.

Parameters:

Name Type Description Default
optimizer Optimizer

Optimizer to run.

 required
verbose bool

Whether to suppress output. DSPy produces a good amount of logs, so this can be useful to not pollute your terminal. Only warnings and errors will be printed.

 True

Returns:

Type Description
tuple[str, Sequence[FewshotExample]]

Best found prompt and few-shot examples.
Source code in sieves/tasks/predictive/core.py 
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
def optimize(self, optimizer: optimization.Optimizer, verbose: bool = True) -> tuple[str, Sequence[FewshotExample]]:
    """Optimize task prompt and few-shot examples with the available optimization config.

    Updates task to use best prompt and few-shot examples found by the optimizer.

    :param optimizer: Optimizer to run.
    :param verbose: Whether to suppress output. DSPy produces a good amount of logs, so this can be useful to
        not pollute your terminal. Only warnings and errors will be printed.

    :return tuple[str, Sequence[FewshotExample]]: Best found prompt and few-shot examples.
    """
    assert len(self._fewshot_examples) > 1, "At least two few-shot examples need to be provided to optimize."

    # Run optimizer to get best prompt and few-shot examples.
    signature = self._get_task_signature()
    dspy_examples = [ex.to_dspy() for ex in self._fewshot_examples]

    def _pred_eval(truth: dspy.Example, pred: dspy.Prediction, trace: Any | None = None) -> float:
        """Wrap optimization evaluation, inject model.

        :param truth: Ground truth.
        :param pred: Predicted value.
        :param trace: Optional trace information.
        :return: Metric value between 0.0 and 1.0.
        :raises KeyError: If target fields are missing from truth or prediction.
        :raises ValueError: If similarity score cannot be parsed from LLM response.
        """
        return self._evaluate_dspy_example(truth, pred, trace, model=optimizer.model)

    if verbose:
        best_prompt, best_examples = optimizer(signature, dspy_examples, _pred_eval, verbose=verbose)
    else:
        # Temporarily suppress DSPy logs.
        dspy_logger = logging.getLogger("dspy")
        optuna_logger = logging.getLogger("optuna")
        original_dspy_level = dspy_logger.level
        original_optuna_level = optuna_logger.level

        try:
            dspy_logger.setLevel(logging.ERROR)
            optuna_logger.setLevel(logging.ERROR)
            with warnings.catch_warnings():
                warnings.simplefilter("ignore")
                best_prompt, best_examples = optimizer(signature, dspy_examples, _pred_eval, verbose=verbose)
        finally:
            dspy_logger.setLevel(original_dspy_level)
            optuna_logger.setLevel(original_optuna_level)

    # Update few-shot examples and prompt instructions.
    fewshot_example_cls = self._fewshot_examples[0].__class__
    self._fewshot_examples = [fewshot_example_cls.from_dspy(ex) for ex in best_examples]
    self._validate_fewshot_examples()
    self._custom_prompt_instructions = best_prompt

    # Reinitialize bridge to use new prompt and few-shot examples.
    self._bridge = self._init_bridge(ModelType.get_model_type(self._model_wrapper))

    return best_prompt, self._fewshot_examples

serialize()

Serialize task.

Returns:

Type Description
Config

Config instance.
Source code in sieves/tasks/core.py 
147
148
149
150
151
152
def serialize(self) -> Config:
    """Serialize task.

    :return: Config instance.
    """
    return Config.create(self.__class__, {k: Attribute(value=v) for k, v in self._state.items()})

Bridges for relation extraction task.

DSPyRelationExtraction

Bases: RelationExtractionBridge[PromptSignature, Result, InferenceMode]

DSPy bridge for relation extraction.

Source code in sieves/tasks/predictive/relation_extraction/bridges.py 
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
class DSPyRelationExtraction(RelationExtractionBridge[dspy_.PromptSignature, dspy_.Result, dspy_.InferenceMode]):
    """DSPy bridge for relation extraction."""

    @override
    def _validate(self) -> None:
        assert self._model_type == ModelType.dspy

    @property
    @override
    def _chunk_extractor(self) -> Callable[[Any], Iterable[pydantic.BaseModel]]:
        return lambda res: res.triplets

    @override
    @property
    def _default_prompt_instructions(self) -> str:
        return ""

    @override
    @property
    def inference_mode(self) -> dspy_.InferenceMode:
        return self._model_settings.inference_mode or dspy_.InferenceMode.predict

model_settings property

Return model settings.

Returns:

Type Description
ModelSettings

Model settings.

model_type property

Return model type.

Returns:

Type Description
ModelType

Model type.

prompt_template property

Return prompt template.

Chains _prompt_instructions, _prompt_example_xml and _prompt_conclusion.

Note: different model have different expectations as to how a prompt should look like. E.g. outlines supports the Jinja 2 templating format for insertion of values and few-shot examples, whereas DSPy integrates these things in a different value in the workflow and hence expects the prompt not to include these things. Mind model-specific expectations when creating a prompt template.

Returns:

Type Description
str

Prompt template as string. None if not used by model wrapper.

__init__(task_id, relations, entity_types, prompt_instructions, model_settings, prompt_signature, model_type, fewshot_examples=())

Initialize relation extraction bridge.

Parameters:

Name Type Description Default
task_id str

Task ID.

 required
relations Sequence[str] | dict[str, str]

Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.

 required
entity_types Sequence[str] | dict[str, str] | None

Entity types constraints.

 required
prompt_instructions str | None

Custom prompt instructions. If None, default instructions are used.

 required
model_settings ModelSettings

Settings for structured generation.

 required
prompt_signature type[BaseModel]

Unified Pydantic prompt signature.

 required
model_type ModelType

Model type.

 required
fewshot_examples Sequence[BaseModel]

Few-shot examples.

 ()
Source code in sieves/tasks/predictive/relation_extraction/bridges.py 
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
def __init__(
    self,
    task_id: str,
    relations: Sequence[str] | dict[str, str],
    entity_types: Sequence[str] | dict[str, str] | None,
    prompt_instructions: str | None,
    model_settings: ModelSettings,
    prompt_signature: type[pydantic.BaseModel],
    model_type: ModelType,
    fewshot_examples: Sequence[pydantic.BaseModel] = (),
):
    """Initialize relation extraction bridge.

    :param task_id: Task ID.
    :param relations: Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.
    :param entity_types: Entity types constraints.
    :param prompt_instructions: Custom prompt instructions. If None, default instructions are used.
    :param model_settings: Settings for structured generation.
    :param prompt_signature: Unified Pydantic prompt signature.
    :param model_type: Model type.
    :param fewshot_examples: Few-shot examples.
    """
    super().__init__(
        task_id=task_id,
        prompt_instructions=prompt_instructions,
        overwrite=False,
        model_settings=model_settings,
        prompt_signature=prompt_signature,
        model_type=model_type,
        fewshot_examples=fewshot_examples,
    )
    if isinstance(relations, dict):
        self._relations = list(relations.keys())
        self._relation_descriptions = relations
    else:
        self._relations = list(relations)
        self._relation_descriptions = {}

    if isinstance(entity_types, dict):
        self._entity_types = list(entity_types.keys())
        self._entity_type_descriptions = entity_types
    elif entity_types is not None:
        self._entity_types = list(entity_types)
        self._entity_type_descriptions = {}
    else:
        self._entity_types = None
        self._entity_type_descriptions = {}

    self._consolidation_strategy = MultiEntityConsolidation(extractor=self._chunk_extractor)

extract(docs)

Extract all values from doc instances that are to be injected into the prompts.

Parameters:

Name Type Description Default
docs Sequence[Doc]

Docs to extract values from.

 required

Returns:

Type Description
Sequence[dict[str, Any]]

All values from doc instances that are to be injected into the prompts as a sequence.
Source code in sieves/tasks/predictive/bridges.py 
178
179
180
181
182
183
184
def extract(self, docs: Sequence[Doc]) -> Sequence[dict[str, Any]]:
    """Extract all values from doc instances that are to be injected into the prompts.

    :param docs: Docs to extract values from.
    :return: All values from doc instances that are to be injected into the prompts as a sequence.
    """
    return [{"text": doc.text if doc.text else None} for doc in docs]

PydanticRelationExtraction

Bases: RelationExtractionBridge[BaseModel, BaseModel | list[Any], ModelWrapperInferenceMode], ABC

Base class for Pydantic-based relation extraction bridges.

Source code in sieves/tasks/predictive/relation_extraction/bridges.py 
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
class PydanticRelationExtraction(
    RelationExtractionBridge[pydantic.BaseModel, pydantic.BaseModel | list[Any], ModelWrapperInferenceMode], abc.ABC
):
    """Base class for Pydantic-based relation extraction bridges."""

    @override
    def _validate(self) -> None:
        assert self._model_type in {ModelType.langchain, ModelType.outlines}

    @property
    @override
    def _chunk_extractor(self) -> Callable[[Any], Iterable[pydantic.BaseModel]]:
        return lambda res: res.triplets

    @override
    @property
    def _default_prompt_instructions(self) -> str:
        return (
            "Extract relations between entities in the text.\n"
            f"Relations: {self._relations}\n"
            f"Entity Types: {self._entity_types or 'Any'}\n"
            "Return a list of triplets with head, relation, tail, and a confidence score between 0.0 and 1.0."
        )

    @override
    @property
    def _prompt_conclusion(self) -> str | None:
        return "========\n<text>{{ text }}</text>"

    @override
    @property
    def model_type(self) -> ModelType:
        return self._model_type

    @override
    @property
    def inference_mode(self) -> outlines_.InferenceMode | langchain_.InferenceMode:
        if self._model_type == ModelType.outlines:
            return self._model_settings.inference_mode or outlines_.InferenceMode.json
        elif self._model_type == ModelType.langchain:
            return self._model_settings.inference_mode or langchain_.InferenceMode.structured

        raise ValueError(f"Unsupported model type: {self._model_type}")

model_settings property

Return model settings.

Returns:

Type Description
ModelSettings

Model settings.

prompt_template property

Return prompt template.

Chains _prompt_instructions, _prompt_example_xml and _prompt_conclusion.

Note: different model have different expectations as to how a prompt should look like. E.g. outlines supports the Jinja 2 templating format for insertion of values and few-shot examples, whereas DSPy integrates these things in a different value in the workflow and hence expects the prompt not to include these things. Mind model-specific expectations when creating a prompt template.

Returns:

Type Description
str

Prompt template as string. None if not used by model wrapper.

__init__(task_id, relations, entity_types, prompt_instructions, model_settings, prompt_signature, model_type, fewshot_examples=())

Initialize relation extraction bridge.

Parameters:

Name Type Description Default
task_id str

Task ID.

 required
relations Sequence[str] | dict[str, str]

Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.

 required
entity_types Sequence[str] | dict[str, str] | None

Entity types constraints.

 required
prompt_instructions str | None

Custom prompt instructions. If None, default instructions are used.

 required
model_settings ModelSettings

Settings for structured generation.

 required
prompt_signature type[BaseModel]

Unified Pydantic prompt signature.

 required
model_type ModelType

Model type.

 required
fewshot_examples Sequence[BaseModel]

Few-shot examples.

 ()
Source code in sieves/tasks/predictive/relation_extraction/bridges.py 
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
def __init__(
    self,
    task_id: str,
    relations: Sequence[str] | dict[str, str],
    entity_types: Sequence[str] | dict[str, str] | None,
    prompt_instructions: str | None,
    model_settings: ModelSettings,
    prompt_signature: type[pydantic.BaseModel],
    model_type: ModelType,
    fewshot_examples: Sequence[pydantic.BaseModel] = (),
):
    """Initialize relation extraction bridge.

    :param task_id: Task ID.
    :param relations: Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.
    :param entity_types: Entity types constraints.
    :param prompt_instructions: Custom prompt instructions. If None, default instructions are used.
    :param model_settings: Settings for structured generation.
    :param prompt_signature: Unified Pydantic prompt signature.
    :param model_type: Model type.
    :param fewshot_examples: Few-shot examples.
    """
    super().__init__(
        task_id=task_id,
        prompt_instructions=prompt_instructions,
        overwrite=False,
        model_settings=model_settings,
        prompt_signature=prompt_signature,
        model_type=model_type,
        fewshot_examples=fewshot_examples,
    )
    if isinstance(relations, dict):
        self._relations = list(relations.keys())
        self._relation_descriptions = relations
    else:
        self._relations = list(relations)
        self._relation_descriptions = {}

    if isinstance(entity_types, dict):
        self._entity_types = list(entity_types.keys())
        self._entity_type_descriptions = entity_types
    elif entity_types is not None:
        self._entity_types = list(entity_types)
        self._entity_type_descriptions = {}
    else:
        self._entity_types = None
        self._entity_type_descriptions = {}

    self._consolidation_strategy = MultiEntityConsolidation(extractor=self._chunk_extractor)

extract(docs)

Extract all values from doc instances that are to be injected into the prompts.

Parameters:

Name Type Description Default
docs Sequence[Doc]

Docs to extract values from.

 required

Returns:

Type Description
Sequence[dict[str, Any]]

All values from doc instances that are to be injected into the prompts as a sequence.
Source code in sieves/tasks/predictive/bridges.py 
178
179
180
181
182
183
184
def extract(self, docs: Sequence[Doc]) -> Sequence[dict[str, Any]]:
    """Extract all values from doc instances that are to be injected into the prompts.

    :param docs: Docs to extract values from.
    :return: All values from doc instances that are to be injected into the prompts as a sequence.
    """
    return [{"text": doc.text if doc.text else None} for doc in docs]

RelationExtractionBridge

Bases: Bridge[_BridgePromptSignature, _BridgeResult, ModelWrapperInferenceMode], ABC

Abstract base class for relation extraction bridges.

Source code in sieves/tasks/predictive/relation_extraction/bridges.py 
 33
 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
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
class RelationExtractionBridge(Bridge[_BridgePromptSignature, _BridgeResult, ModelWrapperInferenceMode], abc.ABC):
    """Abstract base class for relation extraction bridges."""

    def __init__(
        self,
        task_id: str,
        relations: Sequence[str] | dict[str, str],
        entity_types: Sequence[str] | dict[str, str] | None,
        prompt_instructions: str | None,
        model_settings: ModelSettings,
        prompt_signature: type[pydantic.BaseModel],
        model_type: ModelType,
        fewshot_examples: Sequence[pydantic.BaseModel] = (),
    ):
        """Initialize relation extraction bridge.

        :param task_id: Task ID.
        :param relations: Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.
        :param entity_types: Entity types constraints.
        :param prompt_instructions: Custom prompt instructions. If None, default instructions are used.
        :param model_settings: Settings for structured generation.
        :param prompt_signature: Unified Pydantic prompt signature.
        :param model_type: Model type.
        :param fewshot_examples: Few-shot examples.
        """
        super().__init__(
            task_id=task_id,
            prompt_instructions=prompt_instructions,
            overwrite=False,
            model_settings=model_settings,
            prompt_signature=prompt_signature,
            model_type=model_type,
            fewshot_examples=fewshot_examples,
        )
        if isinstance(relations, dict):
            self._relations = list(relations.keys())
            self._relation_descriptions = relations
        else:
            self._relations = list(relations)
            self._relation_descriptions = {}

        if isinstance(entity_types, dict):
            self._entity_types = list(entity_types.keys())
            self._entity_type_descriptions = entity_types
        elif entity_types is not None:
            self._entity_types = list(entity_types)
            self._entity_type_descriptions = {}
        else:
            self._entity_types = None
            self._entity_type_descriptions = {}

        self._consolidation_strategy = MultiEntityConsolidation(extractor=self._chunk_extractor)

    @override
    @property
    def prompt_signature(self) -> _BridgePromptSignature:
        return convert_to_signature(
            model_cls=self._pydantic_signature,
            model_type=self.model_type,
            mode="relations",
        )  # type: ignore[return-value]

    @property
    @abc.abstractmethod
    def _chunk_extractor(self) -> Callable[[Any], Iterable[pydantic.BaseModel]]:
        """Return a callable that extracts a list of entities from a raw chunk result.

        :return: Extractor callable.
        """

    def _get_relation_descriptions(self) -> str:
        """Return relation descriptions as a string.

        :return: Relation descriptions.
        """
        descs: list[str] = []
        for rel in self._relations:
            if rel in self._relation_descriptions:
                descs.append(
                    f"  <relation_description>\n    <relation>{rel}</relation>\n    <description>"
                    f"{self._relation_descriptions[rel]}</description>\n  </relation_description>"
                )
            else:
                descs.append(f"  <relation>{rel}</relation>")
        return "\n".join(descs)

    def _get_entity_type_descriptions(self) -> str:
        """Return entity type descriptions as a string.

        :return: Entity type descriptions.
        """
        if self._entity_types is None:
            return "Unbounded"

        descs: list[str] = []
        for et in self._entity_types:
            if et in self._entity_type_descriptions:
                descs.append(
                    f"  <entity_type_description>\n    <type>{et}</type>\n    <description>"
                    f"{self._entity_type_descriptions[et]}</description>\n  </entity_type_description>"
                )
            else:
                descs.append(f"  <type>{et}</type>")

        return "\n".join(descs)

    def _process_triplets(self, raw_triplets: list[Any]) -> list[RelationTriplet]:
        """Convert raw triplets from model to RelationTriplet objects.

        :param raw_triplets: Raw triplets from the model.
        :return: Processed RelationTriplet objects.
        """
        processed: list[RelationTriplet] = []
        for raw in raw_triplets:
            head_text = getattr(raw.head, "text", "")
            head_type = getattr(raw.head, "entity_type", "")

            tail_text = getattr(raw.tail, "text", "")
            tail_type = getattr(raw.tail, "entity_type", "")

            processed.append(
                RelationTriplet(
                    head=RelationEntity(text=head_text, entity_type=head_type),
                    relation=getattr(raw, "relation", ""),
                    tail=RelationEntity(text=tail_text, entity_type=tail_type),
                    score=getattr(raw, "score", None),
                )
            )

        return processed

    @override
    def integrate(self, results: Sequence[_BridgeResult | list[Any]], docs: list[Doc]) -> list[Doc]:
        for doc, result in zip(docs, results):
            # Handle both model result objects and raw lists from consolidation.
            raw_triplets = result if isinstance(result, list) else getattr(result, "triplets", [])
            doc.results[self._task_id] = Result(triplets=self._process_triplets(raw_triplets))

        return docs

    @override
    def consolidate(
        self,
        results: Sequence[_BridgeResult],
        docs_offsets: list[tuple[int, int]],
    ) -> Sequence[list[pydantic.BaseModel]]:
        return self._consolidation_strategy.consolidate(results, docs_offsets)

inference_mode abstractmethod property

Return inference mode.

Returns:

Type Description
ModelWrapperInferenceMode

Inference mode.

model_settings property

Return model settings.

Returns:

Type Description
ModelSettings

Model settings.

model_type property

Return model type.

Returns:

Type Description
ModelType

Model type.

prompt_template property

Return prompt template.

Chains _prompt_instructions, _prompt_example_xml and _prompt_conclusion.

Note: different model have different expectations as to how a prompt should look like. E.g. outlines supports the Jinja 2 templating format for insertion of values and few-shot examples, whereas DSPy integrates these things in a different value in the workflow and hence expects the prompt not to include these things. Mind model-specific expectations when creating a prompt template.

Returns:

Type Description
str

Prompt template as string. None if not used by model wrapper.

__init__(task_id, relations, entity_types, prompt_instructions, model_settings, prompt_signature, model_type, fewshot_examples=())

Initialize relation extraction bridge.

Parameters:

Name Type Description Default
task_id str

Task ID.

 required
relations Sequence[str] | dict[str, str]

Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.

 required
entity_types Sequence[str] | dict[str, str] | None

Entity types constraints.

 required
prompt_instructions str | None

Custom prompt instructions. If None, default instructions are used.

 required
model_settings ModelSettings

Settings for structured generation.

 required
prompt_signature type[BaseModel]

Unified Pydantic prompt signature.

 required
model_type ModelType

Model type.

 required
fewshot_examples Sequence[BaseModel]

Few-shot examples.

 ()
Source code in sieves/tasks/predictive/relation_extraction/bridges.py 
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
def __init__(
    self,
    task_id: str,
    relations: Sequence[str] | dict[str, str],
    entity_types: Sequence[str] | dict[str, str] | None,
    prompt_instructions: str | None,
    model_settings: ModelSettings,
    prompt_signature: type[pydantic.BaseModel],
    model_type: ModelType,
    fewshot_examples: Sequence[pydantic.BaseModel] = (),
):
    """Initialize relation extraction bridge.

    :param task_id: Task ID.
    :param relations: Relations to extract. Can be a list of relation types or a dict mapping types to descriptions.
    :param entity_types: Entity types constraints.
    :param prompt_instructions: Custom prompt instructions. If None, default instructions are used.
    :param model_settings: Settings for structured generation.
    :param prompt_signature: Unified Pydantic prompt signature.
    :param model_type: Model type.
    :param fewshot_examples: Few-shot examples.
    """
    super().__init__(
        task_id=task_id,
        prompt_instructions=prompt_instructions,
        overwrite=False,
        model_settings=model_settings,
        prompt_signature=prompt_signature,
        model_type=model_type,
        fewshot_examples=fewshot_examples,
    )
    if isinstance(relations, dict):
        self._relations = list(relations.keys())
        self._relation_descriptions = relations
    else:
        self._relations = list(relations)
        self._relation_descriptions = {}

    if isinstance(entity_types, dict):
        self._entity_types = list(entity_types.keys())
        self._entity_type_descriptions = entity_types
    elif entity_types is not None:
        self._entity_types = list(entity_types)
        self._entity_type_descriptions = {}
    else:
        self._entity_types = None
        self._entity_type_descriptions = {}

    self._consolidation_strategy = MultiEntityConsolidation(extractor=self._chunk_extractor)

extract(docs)

Extract all values from doc instances that are to be injected into the prompts.

Parameters:

Name Type Description Default
docs Sequence[Doc]

Docs to extract values from.

 required

Returns:

Type Description
Sequence[dict[str, Any]]

All values from doc instances that are to be injected into the prompts as a sequence.
Source code in sieves/tasks/predictive/bridges.py 
178
179
180
181
182
183
184
def extract(self, docs: Sequence[Doc]) -> Sequence[dict[str, Any]]:
    """Extract all values from doc instances that are to be injected into the prompts.

    :param docs: Docs to extract values from.
    :return: All values from doc instances that are to be injected into the prompts as a sequence.
    """
    return [{"text": doc.text if doc.text else None} for doc in docs]

Schemas for relation extraction task.

FewshotExample

Bases: FewshotExample

Few-shot example for relation extraction.

Attributes: text: Input text. triplets: Expected relation triplets.

Source code in sieves/tasks/predictive/schemas/relation_extraction.py 
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
class FewshotExample(BaseFewshotExample):
    """Few-shot example for relation extraction.

    Attributes:
        text: Input text.
        triplets: Expected relation triplets.
    """

    triplets: list[RelationTriplet]

    @property
    def target_fields(self) -> tuple[str, ...]:
        """Return target fields.

        :return: Target fields.
        """
        return ("triplets",)

input_fields property

Defines which fields are inputs.

Returns:

Type Description
Sequence[str]

Sequence of field names.

target_fields property

Return target fields.

Returns:

Type Description
tuple[str, ...]

Target fields.

from_dspy(example) classmethod

Convert from dspy.Example.

Parameters:

Name Type Description Default
example Example

Example as dspy.Example.

 required

Returns:

Type Description
Self

Example as FewshotExample.
Source code in sieves/tasks/predictive/schemas/core.py 
63
64
65
66
67
68
69
70
@classmethod
def from_dspy(cls, example: dspy.Example) -> Self:
    """Convert from `dspy.Example`.

    :param example: Example as `dspy.Example`.
    :returns: Example as `FewshotExample`.
    """
    return cls(**example)

to_dspy()

Convert to dspy.Example.

Returns:

Type Description
Example

Example as dspy.Example.
Source code in sieves/tasks/predictive/schemas/core.py 
56
57
58
59
60
61
def to_dspy(self) -> dspy.Example:
    """Convert to `dspy.Example`.

    :returns: Example as `dspy.Example`.
    """
    return dspy.Example(**ModelWrapper.convert_fewshot_examples([self])[0]).with_inputs(*self.input_fields)

RelationEntity

Bases: BaseModel

Entity involved in a relation.

Attributes: text: Surface text of the entity. entity_type: Type of the entity.

Source code in sieves/tasks/predictive/schemas/relation_extraction.py 
13
14
15
16
17
18
19
20
21
22
class RelationEntity(pydantic.BaseModel, frozen=True):
    """Entity involved in a relation.

    Attributes:
        text: Surface text of the entity.
        entity_type: Type of the entity.
    """

    text: str
    entity_type: str

RelationEntityWithContext

Bases: BaseModel

Entity mention with text span, type, and context for span discovery.

Attributes: text: Surface text of the entity. context: Short context around the entity. entity_type: Type of the entity.

Source code in sieves/tasks/predictive/schemas/relation_extraction.py 
55
56
57
58
59
60
61
62
63
64
65
66
class RelationEntityWithContext(pydantic.BaseModel):
    """Entity mention with text span, type, and context for span discovery.

    Attributes:
        text: Surface text of the entity.
        context: Short context around the entity.
        entity_type: Type of the entity.
    """

    text: str
    context: str
    entity_type: str

RelationTriplet

Bases: BaseModel

Triplet representing a relation between two entities.

Attributes: head: The subject entity. relation: The type of relation. tail: The object entity. score: Confidence score.

Source code in sieves/tasks/predictive/schemas/relation_extraction.py 
25
26
27
28
29
30
31
32
33
34
35
36
37
38
class RelationTriplet(pydantic.BaseModel, frozen=True):
    """Triplet representing a relation between two entities.

    Attributes:
        head: The subject entity.
        relation: The type of relation.
        tail: The object entity.
        score: Confidence score.
    """

    head: RelationEntity
    relation: str
    tail: RelationEntity
    score: float | None = None

RelationTripletWithContext

Bases: BaseModel

Triplet with context for span discovery.

Attributes: head: The head entity with context. relation: The relation type. tail: The tail entity with context. score: Confidence score.

Source code in sieves/tasks/predictive/schemas/relation_extraction.py 
69
70
71
72
73
74
75
76
77
78
79
80
81
82
class RelationTripletWithContext(pydantic.BaseModel):
    """Triplet with context for span discovery.

    Attributes:
        head: The head entity with context.
        relation: The relation type.
        tail: The tail entity with context.
        score: Confidence score.
    """

    head: RelationEntityWithContext
    relation: str
    tail: RelationEntityWithContext
    score: float | None = None

Result

Bases: BaseModel

Result of a relation extraction task.

Attributes: triplets: List of extracted relation triplets.

Source code in sieves/tasks/predictive/schemas/relation_extraction.py 
42
43
44
45
46
47
48
49
class Result(pydantic.BaseModel):
    """Result of a relation extraction task.

    Attributes:
        triplets: List of extracted relation triplets.
    """

    triplets: list[RelationTriplet]