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
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412 | class GliNERBridge(Bridge[gliner2.inference.engine.Schema, gliner_.Result, gliner_.InferenceMode]):
"""Bridge for GLiNER2 models."""
def __init__(
self,
task_id: str,
prompt_instructions: str | None,
prompt_signature: type[pydantic.BaseModel],
model_settings: ModelSettings,
inference_mode: gliner_.InferenceMode,
mode: Literal["multi", "single"] = "multi",
):
"""Initialize GLiNER2 bridge.
Important: currently only GLiNER2 schemas/structures with one key each are supported. We do NOT support
composite requests like `create_schema().entities().classification(). ...`.
:param task_id: Task ID.
:param prompt_instructions: Custom prompt instructions. If None, default instructions are used.
:param prompt_signature: Unified Pydantic prompt signature.
:param model_settings: Model settings including inference_mode.
:param mode: Extraction mode. If "multi", all occurrences of the entity are extracted. If "single", exactly one
(or no) entity is extracted. Only used if `inference_mode == InferenceMode.structure`, ignored otherwise.
"""
super().__init__(
task_id=task_id,
prompt_instructions=prompt_instructions,
overwrite=False,
model_settings=model_settings,
prompt_signature=prompt_signature,
model_type=ModelType.gliner,
)
self._inference_mode = inference_mode
self._mode = mode
@override
@property
def model_type(self) -> ModelType:
return ModelType.gliner
@override
@property
def prompt_signature(self) -> gliner2.inference.engine.Schema | gliner2.inference.engine.StructureBuilder:
# Map internal inference mode to GliNER utility mode.
mode_map = {
gliner_.InferenceMode.classification: "classification",
gliner_.InferenceMode.entities: "entities",
gliner_.InferenceMode.structure: "structure",
gliner_.InferenceMode.relations: "relations",
}
mode = mode_map[self.inference_mode]
model_cls = self._pydantic_signature
# Unwrap unified signature container if necessary.
if mode == "structure" or mode == "entities":
if "entities" in model_cls.model_fields:
model_cls = model_cls.model_fields["entities"].annotation
if typing.get_origin(model_cls) is list:
model_cls = typing.get_args(model_cls)[0]
elif "entity" in model_cls.model_fields:
model_cls = model_cls.model_fields["entity"].annotation
# Handle Optional/Union.
if typing.get_origin(model_cls) in (typing.Union, types.UnionType):
args = typing.get_args(model_cls)
model_cls = [t for t in args if t is not type(None)][0]
elif mode == "relations":
if "triplets" in model_cls.model_fields:
model_cls = model_cls.model_fields["triplets"].annotation
if typing.get_origin(model_cls) is list:
model_cls = typing.get_args(model_cls)[0]
kwargs = {"mode": mode}
if mode == "classification":
kwargs["task"] = self._task_id
prompt_signature = convert_to_signature(
model_cls=model_cls,
model_type=ModelType.gliner,
**kwargs,
)
assert isinstance(prompt_signature, gliner2.inference.engine.Schema | gliner2.inference.engine.StructureBuilder)
return prompt_signature
@property
def prompt_signature_pydantic(self) -> type[pydantic.BaseModel] | None:
"""Return Pydantic model representation of GLiNER2 schema.
:return: Pydantic model representation of GLiNER2 schema.
"""
return self._pydantic_signature
@override
@property
def _default_prompt_instructions(self) -> str:
# GLiNER2 doesn't support custom instructions.
return ""
@override
@property
def inference_mode(self) -> gliner_.InferenceMode:
return self._model_settings.inference_mode or self._inference_mode
@override
def integrate(self, results: Sequence[gliner_.Result], docs: list[Doc]) -> list[Doc]:
for doc, result in zip(docs, results):
match self._inference_mode:
# Used by: Classification
case gliner_.InferenceMode.classification:
if self._mode == "multi":
label_scores: list[tuple[str, float]] = []
for res in sorted(result, key=lambda x: x["score"], reverse=True):
assert isinstance(res, dict)
label_scores.append((res["label"], res["score"]))
doc.results[self._task_id] = ResultMultiLabel(label_scores=label_scores)
else:
doc.results[self._task_id] = ResultSingleLabel(
label=result[0]["label"], score=result[0]["score"]
)
# Used by: NER
case gliner_.InferenceMode.entities:
if isinstance(result, list):
# Result is a list of entity dictionaries.
doc.results[self._task_id] = NERResult(
entities=[
NEREntity.model_validate(
{k if k != "confidence" else "score": v for k, v in res.items()}
)
for res in result
],
text=doc.text or "",
)
else:
doc.results[self._task_id] = result
# Used by: InformationExtraction
case gliner_.InferenceMode.structure:
if not result:
if self._mode == "multi":
doc.results[self._task_id] = IEResultMulti(entities=[])
else:
doc.results[self._task_id] = IEResultSingle(entity=None)
continue
entity_type_name = list(result.keys())[0]
# Unified signature container unwrapping
validation_model = self.prompt_signature_pydantic
if validation_model and "entities" in validation_model.model_fields:
# Multi-mode container
validation_model = validation_model.model_fields["entities"].annotation
if hasattr(validation_model, "__args__"):
validation_model = validation_model.__args__[0]
elif validation_model and "entity" in validation_model.model_fields:
# Single-mode container
validation_model = validation_model.model_fields["entity"].annotation
if hasattr(validation_model, "__args__"):
# Filter out None
validation_model = [t for t in validation_model.__args__ if t is not type(None)][0]
assert validation_model is not None and issubclass(validation_model, pydantic.BaseModel)
extracted_entities: list[pydantic.BaseModel] = []
for entity in result[entity_type_name]:
# Map fields and include score.
entity_data = {key: value["text"] for key, value in entity.items()}
# Calculate average confidence across all fields.
confidences = [v["confidence"] for v in entity.values() if "confidence" in v]
entity_data["score"] = sum(confidences) / len(confidences) if confidences else None
extracted_entities.append(validation_model.model_validate(entity_data))
# This covers single/multi mode for information extraction.
if self._mode == "multi":
doc.results[self._task_id] = IEResultMulti(entities=extracted_entities)
else:
doc.results[self._task_id] = IEResultSingle(
entity=extracted_entities[0] if extracted_entities else None
)
# Used by: RelationExtraction
case gliner_.InferenceMode.relations:
triplets: list[RERelationTriplet] = []
# GliNER2 relations output is a dict mapping task ids to relation types to lists of triplets:
# {'task_id': {'relation_type': [{'head': {...}, 'tail': {...}}, ...]}} # noqa: ERA001
relation_data = result.get(self._task_id) or result.get("relation_extraction", {})
assert isinstance(relation_data, dict)
for rel_type, triplets_list in relation_data.items():
for triplet_data in triplets_list:
# Calculate score as average of head and tail confidence if they exist.
head_conf = triplet_data.get("head", {}).get("confidence")
tail_conf = triplet_data.get("tail", {}).get("confidence")
confidences = [c for c in (head_conf, tail_conf) if c is not None]
score = sum(confidences) / len(confidences) if confidences else None
triplets.append(
RERelationTriplet(
head=RERelationEntity(
text=triplet_data["head"]["text"],
entity_type="UNKNOWN",
),
relation=rel_type,
tail=RERelationEntity(
text=triplet_data["tail"]["text"],
entity_type="UNKNOWN",
),
score=score,
)
)
doc.results[self._task_id] = RERelationResult(triplets=triplets)
return docs
@override
def consolidate(
self, results: Sequence[gliner_.Result], docs_offsets: list[tuple[int, int]]
) -> Sequence[gliner_.Result]:
consolidated_results: list[gliner_.Result] = []
# Determine label scores for chunks per document.
for doc_offset in docs_offsets:
scores: dict[Any, float] = defaultdict(lambda: 0.0)
entities: dict[str, list[str] | dict[str, str | list[str]]] = {}
all_entities_list: list[dict[str, Any]] = []
# In single mode for structure, we want to keep track of the highest confidence entity.
highest_confidence_entity: dict[str, Any] | None = None
max_confidence: float = -1.0
for res in results[doc_offset[0] : doc_offset[1]]:
match self._inference_mode:
case gliner_.InferenceMode.classification:
keys = list(res.keys())
assert len(keys) == 1, "Composite GliNER2 schemas are not supported."
extracted_res = res[keys[0]]
# In case of single-label: pad to list so that we can process in a unified way.
if isinstance(extracted_res, dict):
extracted_res = [extracted_res]
for entry in extracted_res:
# GliNER might use two different structures here, depending on the version.
if "label" in entry:
scores[entry["label"]] += entry["confidence"]
else:
assert isinstance(entry, dict)
keys = list(entry.keys())
assert len(keys) == 1, "Composite GliNER2 schemas are not supported."
for label, confidence in entry[keys[0]]:
scores[label] += confidence
case gliner_.InferenceMode.entities:
for entity_type in res["entities"]:
items = res["entities"][entity_type]
if items:
if isinstance(items[0], dict):
# Flatten into a list of entities (dicts).
for item in items:
all_entities_list.append({"entity_type": entity_type, **item})
else:
if entity_type not in entities:
entities[entity_type] = []
relevant_entities: list[str] = entities[entity_type] # type: ignore[assignment]
relevant_entities.extend(items)
case gliner_.InferenceMode.structure:
for entity_type in res:
if entity_type not in entities:
entities[entity_type] = []
relevant_entities_struct: list[Any] = entities[entity_type] # type: ignore[assignment]
# Add entities from this chunk.
for entity in res[entity_type]:
# If single mode: we already have logic to track max_confidence below.
# But we also need to store all entities for general structure if not single?
# Actually, GLiNER structure mode output is a bit different.
relevant_entities_struct.append(entity)
if self._mode == "single":
for entity in res[entity_type]:
# Calculate average confidence for the entity structure.
assert isinstance(entity, dict)
confidences = [
field_val["confidence"]
for field_val in entity.values()
if "confidence" in field_val
]
if confidences:
avg_conf = sum(confidences) / len(confidences)
if avg_conf > max_confidence:
max_confidence = avg_conf
# We need to preserve the confidence in the structure
# so that integrate can pick it up.
highest_confidence_entity = {entity_type: [entity]}
case gliner_.InferenceMode.relations:
# Collect all triplets from all chunks for this document.
# Simple unification for now.
relation_data = res.get(self._task_id) or res.get("relation_extraction", {})
assert isinstance(relation_data, dict)
for rel_type, triplets_list in relation_data.items():
for triplet in triplets_list:
key = (
triplet["head"]["text"],
rel_type,
triplet["tail"]["text"],
triplet["head"]["start"],
triplet["tail"]["start"],
)
if key not in scores: # Use scores dict as a seen set for convenience.
scores[key] = 1.0
all_entities_list.append({"relation": rel_type, **triplet})
match self._inference_mode:
case gliner_.InferenceMode.classification:
# Ensure that all labels have been assigned - GLiNER2 is sometimes negligent about this.
# We extract labels from our Pydantic signature fields (excluding 'score' if present).
labels = [name for name in self._pydantic_signature.model_fields if name != "score"]
for label in labels:
if label not in scores:
scores[label] = 0.0
# Average score, sort in descending order.
sorted_scores: list[dict[str, str | float]] = sorted(
(
{"label": attr, "score": score / (doc_offset[1] - doc_offset[0])}
for attr, score in scores.items()
),
key=lambda x: x["score"],
reverse=True,
)
consolidated_results.append(sorted_scores)
case gliner_.InferenceMode.entities | gliner_.InferenceMode.structure | gliner_.InferenceMode.relations:
if self._inference_mode == gliner_.InferenceMode.structure and self._mode == "single":
if highest_confidence_entity:
consolidated_results.append(highest_confidence_entity)
else:
# Use the entity name from the pydantic signature (unwrapped).
model_cls = self._pydantic_signature
if "entity" in model_cls.model_fields:
model_cls = model_cls.model_fields["entity"].annotation
if typing.get_origin(model_cls) in (
typing.Union,
types.UnionType,
):
model_cls = [t for t in typing.get_args(model_cls) if t is not type(None)][0]
consolidated_results.append({model_cls.__name__: []})
elif all_entities_list:
if self._inference_mode == gliner_.InferenceMode.relations:
# Reconstruct the dict structure for integrate.
reconstructed: dict[str, list[Any]] = defaultdict(list)
for item in all_entities_list:
rel_type = item.pop("relation")
reconstructed[rel_type].append(item)
# Ensure we use the current task_id as the key so that integrate() can find it.
consolidated_results.append({self._task_id: dict(reconstructed)})
else:
consolidated_results.append(all_entities_list)
else:
consolidated_results.append(entities)
return consolidated_results
|