Skip to content

Extractor

The Extractor class is the main interface for structured data extraction.

API Requirements

All methods require keyword arguments. The * in method signatures indicates that all parameters after it must be passed as keyword arguments:

# ✅ Correct usage
result = extractor.extract(data="file.pdf", query="extract information")
result = extractor.extract_queries(data="file.pdf", queries=["query1", "query2"])
model = extractor.get_schema(data="file.pdf", query="extract information")
refined = extractor.refine_data_model(model=ExistingModel, refinement_instructions="add field")

# ❌ Incorrect usage - will raise TypeError
result = extractor.extract("file.pdf", "extract information")
result = extractor.extract_queries("file.pdf", ["query1", "query2"])

Architecture Overview

View Architecture Diagram
graph TB
    subgraph "User Interface"
        A[Extractor Class]
        A1[extract] 
        A2[extract_queries]
        A3[extract_async]
        A4[refine_data_model]
    end

    subgraph "Core Processing Modules"
        B[LLM Core]
        C[Model Utils]
        D[Data Content Processor]
        E[Model Operations]
        F[Extraction Engine]
    end

    subgraph "File Processing Pipeline"
        G[File Reader]
        H[Format Detection]
        I[Document Conversion]
        J[PDF Generation]
        K[Multimodal Processing]
    end

    subgraph "LLM Integration"
        L[Instructor Client]
        M[LiteLLM Support]
        N[Provider Abstraction]
        O[Token Tracking]
    end

    subgraph "Output Management"
        P[Result Manager]
        Q[Type Safety]
        R[Error Handling]
        S[Usage Statistics]
    end

    A --> A1
    A --> A2
    A --> A3
    A --> A4

    A1 --> B
    A1 --> G
    B --> L
    G --> H
    H --> I
    I --> J
    J --> K

    B --> C
    B --> D
    C --> E
    E --> F
    F --> P

    L --> M
    M --> N
    N --> O
    P --> Q
    P --> R
    P --> S

Main class for structured data extraction - now acts as an orchestrator.

This class coordinates the various specialized components to perform structured data extraction from different types of sources.

Parameters:

Name Type Description Default
client Instructor

Instructor-patched client

required
model_name str

Name of the model to use

required
config Optional[Union[Dict, str, Path, ExtractionConfig]]

Configuration for extraction steps

None
max_threads int

Maximum number of concurrent threads

10
batch_size int

Size of batches for processing

100
max_retries int

Maximum number of retries for extraction

3
min_wait int

Minimum seconds to wait between retries

1
max_wait int

Maximum seconds to wait between retries

10
Source code in structx/extraction/extractor.py
 26
 27
 28
 29
 30
 31
 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
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
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
class Extractor:
    """
    Main class for structured data extraction - now acts as an orchestrator.

    This class coordinates the various specialized components to perform
    structured data extraction from different types of sources.

    Args:
        client: Instructor-patched client
        model_name: Name of the model to use
        config: Configuration for extraction steps
        max_threads: Maximum number of concurrent threads
        batch_size: Size of batches for processing
        max_retries: Maximum number of retries for extraction
        min_wait: Minimum seconds to wait between retries
        max_wait: Maximum seconds to wait between retries
    """

    def __init__(
        self,
        client: Instructor,
        model_name: str,
        config: Optional[Union[Dict, str, Path, ExtractionConfig]] = None,
        max_threads: int = 10,
        batch_size: int = 100,
        max_retries: int = 3,
        min_wait: int = 1,
        max_wait: int = 10,
    ):
        """Initialize extractor."""
        self.model_name = model_name

        # Setup configuration
        if not config:
            self.config = ExtractionConfig()
        elif isinstance(config, (dict, str, Path)):
            self.config = ExtractionConfig(
                config=config if isinstance(config, dict) else None,
                config_path=config if isinstance(config, (str, Path)) else None,
            )
        elif isinstance(config, ExtractionConfig):
            self.config = config
        else:
            raise ConfigurationError("Invalid configuration type")

        # Initialize core components
        self.llm_core = LLMCore(
            client=client,
            model_name=model_name,
            config=self.config,
            max_retries=max_retries,
            min_wait=min_wait,
            max_wait=max_wait,
        )

        # Initialize specialized processors
        self.model_operations = ModelOperations(self.llm_core)
        self.extraction_engine = ExtractionEngine(self.llm_core)
        self.data_processor = DataProcessor(max_threads, batch_size)
        self.result_manager = ResultManager()
        self.content_analyzer = ContentAnalyzer()

        logger.info(f"Initialized Extractor with configuration: {self.config.conf}")

    def _initialize_extraction(
        self, df: pd.DataFrame, query: str, generate_model: bool = True
    ) -> tuple[Any, Any, Optional[Type[BaseModel]]]:
        """Initialize the extraction process by refining query and generating models if needed."""
        # Refine query
        refined_query = self.llm_core.refine_query(query)
        logger.info(f"Refined Query: {refined_query.refined_query}")

        # Generate guide
        guide = self.llm_core.generate_extraction_guide(
            refined_query, df.columns.tolist()
        )
        logger.info(f"Target Columns: {guide.target_columns}")

        if not generate_model:
            return refined_query, guide, None

        # Get sample text for schema generation
        # Check if this is file-based extraction (contains pdf_path or source columns)
        is_file_based = "pdf_path" in df.columns or "source" in df.columns

        if is_file_based:
            # For file-based extractions, extract actual content samples
            sample_text = self.content_analyzer.extract_content_sample_for_schema(df)
            # Add context about the content type
            content_context = self.content_analyzer.detect_content_type_and_context(df)
            sample_text = f"Content type: {content_context}\n\n{sample_text}"
        else:
            # For traditional tabular data, use the existing approach
            sample_text = df[guide.target_columns].iloc[0]

        # Generate model
        schema_request = self.model_operations.generate_extraction_schema(
            sample_text, refined_query, guide
        )
        extraction_model = self.model_operations.create_model_from_schema(
            schema_request
        )

        return refined_query, guide, extraction_model

    def _create_extraction_worker(
        self,
        extraction_model: Type[BaseModel],
        refined_query: Any,
        guide: Any,
        result_df: pd.DataFrame,
        result_list: List[Any],
        failed_rows: List[Dict],
        return_df: bool,
        expand_nested: bool,
        is_custom_model: bool = False,
    ):
        """Create a worker function for threaded extraction."""

        def extract_worker(
            row_data: Union[str, Dict],
            row_idx: int,
            semaphore: threading.Semaphore,
            pbar,
        ):
            with semaphore:
                try:
                    items = self.extraction_engine.extract_from_row_data(
                        row_data=row_data,
                        extraction_model=extraction_model,
                        refined_query=refined_query,
                        guide=guide,
                        is_custom_model=is_custom_model,
                    )

                    if return_df:
                        self.result_manager.update_dataframe(
                            result_df, items, row_idx, expand_nested
                        )
                    else:
                        result_list.extend(items)

                except Exception as e:
                    row_text = row_data if isinstance(row_data, str) else str(row_data)
                    self.result_manager.handle_extraction_error(
                        result_df, failed_rows, row_idx, row_text, e
                    )
                finally:
                    pbar.update(1)

        return extract_worker

    @handle_errors(error_message="Data processing failed", error_type=ExtractionError)
    def _process_data(
        self,
        df: pd.DataFrame,
        query: str,
        return_df: bool,
        expand_nested: bool = False,
        extraction_model: Optional[Type[BaseModel]] = None,
    ) -> ExtractionResult:
        """Process DataFrame with extraction."""
        # Reset usage tracking
        self.llm_core.reset_usage()

        # Initialize extraction
        if extraction_model:
            # When a custom model is provided, generate refinement and guide from the model
            # instead of from the query to avoid conflicts
            refined_query, guide = self.model_operations.generate_from_custom_model(
                model=extraction_model, query=query, data_columns=df.columns.tolist()
            )
            ExtractionModel = extraction_model
        else:
            refined_query, guide, ExtractionModel = self._initialize_extraction(
                df, query, generate_model=True
            )

        # Initialize results
        result_df, result_list, failed_rows = self.result_manager.initialize_results(
            df, ExtractionModel
        )

        # Create worker function - pass is_custom_model flag when using a provided model
        worker_fn = self._create_extraction_worker(
            extraction_model=ExtractionModel,
            refined_query=refined_query,
            guide=guide,
            result_df=result_df,
            result_list=result_list,
            failed_rows=failed_rows,
            return_df=return_df,
            expand_nested=expand_nested,
            is_custom_model=extraction_model is not None,
        )

        # Process in batches
        self.data_processor.process_in_batches(df, worker_fn, guide.target_columns)

        # Log statistics
        self.result_manager.log_extraction_stats(len(df), failed_rows)

        # Create a deep copy of usage for the result
        result_usage = copy.deepcopy(self.llm_core.get_usage())

        # Reset the extractor's usage for the next operation
        self.llm_core.reset_usage()

        # Return results
        return ExtractionResult(
            data=result_df if return_df else result_list,
            failed=pd.DataFrame(failed_rows),
            model=ExtractionModel,
            usage=result_usage,
        )

    @handle_errors(error_message="Extraction failed", error_type=ExtractionError)
    def extract(
        self,
        *,
        data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
        query: str,
        model: Optional[Type[BaseModel]] = None,
        return_df: bool = False,
        expand_nested: bool = False,
        **kwargs: Any,
    ) -> ExtractionResult:
        """
        Extract structured data from text.

        Args:
            data: Input data (file path, DataFrame, list of dicts, or raw text)
            query: Natural language query
            model: Optional pre-generated Pydantic model class (if None, a model will be generated)
            return_df: Whether to return DataFrame
            expand_nested: Whether to flatten nested structures
            **kwargs: Additional options for file reading

        Returns:
            Extraction result with extracted data, failed rows, and model (if requested)
        """
        df = self.data_processor.prepare_data(data, **kwargs)
        return self._process_data(df, query, return_df, expand_nested, model)

    async def extract_async(
        self,
        *,
        data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
        query: str,
        return_df: bool = False,
        expand_nested: bool = False,
        **kwargs: Any,
    ) -> ExtractionResult:
        """
        Asynchronous version of `extract`.

        Args:
            data: Input data (file path, DataFrame, list of dicts, or raw text)
            query: Natural language query
            return_df: Whether to return DataFrame
            expand_nested: Whether to flatten nested structures
            **kwargs: Additional options for file reading

        Returns:
            ExtractionResult containing extracted data, failed rows, and the model
        """
        return await self.data_processor.run_async(
            self.extract, data, query, None, return_df, expand_nested, **kwargs
        )

    @handle_errors(error_message="Batch extraction failed", error_type=ExtractionError)
    def extract_queries(
        self,
        *,
        data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
        queries: List[str],
        return_df: bool = True,
        expand_nested: bool = False,
        **kwargs: Any,
    ) -> Dict[str, ExtractionResult]:
        """
        Process multiple queries on the same data.

        Args:
            data: Input data (file path, DataFrame, list of dicts, or raw text)
            queries: List of queries to process
            return_df: Whether to return DataFrame
            expand_nested: Whether to flatten nested structures
            **kwargs: Additional options for file reading

        Returns:
            Dictionary mapping queries to their results (extracted data and failed extractions)
        """
        results = {}

        for query in queries:
            logger.info(f"\nProcessing query: {query}")
            result = self.extract(
                data=data,
                query=query,
                return_df=return_df,
                expand_nested=expand_nested,
                **kwargs,
            )
            results[query] = result

        return results

    async def extract_queries_async(
        self,
        *,
        data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
        queries: List[str],
        return_df: bool = False,
        expand_nested: bool = False,
        **kwargs: Any,
    ) -> Dict[str, ExtractionResult]:
        """
        Asynchronous version of `extract_queries`.

        Args:
            data: Input data
            queries: List of queries
            return_df: Whether to return DataFrame
            expand_nested: Whether to flatten nested structures
            **kwargs: Additional options

        Returns:
            Dictionary mapping queries to ExtractionResult objects
        """
        return await self.data_processor.run_async(
            self.extract_queries, data, queries, return_df, expand_nested, **kwargs
        )

    @handle_errors(error_message="Schema generation failed", error_type=ExtractionError)
    def get_schema(
        self,
        *,
        data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
        query: str,
        **kwargs: Any,
    ) -> Type[BaseModel]:
        """
        Get extraction model without performing extraction.

        Args:
            query: Natural language query
            data: Input data (file path, DataFrame, list of dicts, or raw text)
            **kwargs: Additional options for file reading

        Returns:
            Pydantic model for extraction with `.usage` attribute for token tracking
        """
        if isinstance(data, str) and not Path(data).exists():
            sample_text = data
            columns = ["text"]
        else:
            df = self.data_processor.prepare_data(data, **kwargs)
            is_file_based = "pdf_path" in df.columns or "source" in df.columns
            columns = df.columns.tolist()

            if is_file_based:
                sample_text = self.content_analyzer.extract_content_sample_for_schema(
                    df
                )
                content_context = self.content_analyzer.detect_content_type_and_context(
                    df
                )
                sample_text = f"Content type: {content_context}\n\n{sample_text}"
            else:
                # For traditional tabular data, create a representative sample
                sample_text = "\n".join(df.head().to_string(index=False).splitlines())

        # Refine query
        refined_query = self.llm_core.refine_query(query)

        # Generate guide
        guide = self.llm_core.generate_extraction_guide(refined_query, columns)

        # Generate schema
        schema_request = self.model_operations.generate_extraction_schema(
            sample_text, refined_query, guide
        )

        # Create model
        extraction_model = self.model_operations.create_model_from_schema(
            schema_request
        )

        # Create a deep copy of usage for the model
        model_usage = copy.deepcopy(self.llm_core.get_usage())

        # Reset the extractor's usage for the next operation
        self.llm_core.reset_usage()

        # Add usage to model
        extraction_model.usage = model_usage

        return extraction_model

    async def get_schema_async(
        self,
        *,
        data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
        query: str,
        **kwargs: Any,
    ) -> Type[BaseModel]:
        """
        Asynchronous version of `get_schema`.

        Args:
            query: Natural language query
            data: Input data (file path, DataFrame, list of dicts, or raw text)
            **kwargs: Additional options for file reading

        Returns:
            Dynamically generated Pydantic model class
        """
        return await self.data_processor.run_async(
            self.get_schema, query, data, **kwargs
        )

    def refine_data_model(
        self,
        *,
        model: Type[BaseModel],
        refinement_instructions: str,
        model_name: Optional[str] = None,
    ) -> Type[BaseModel]:
        """
        Refine an existing data model based on natural language instructions.

        Args:
            model: Existing Pydantic model to refine
            refinement_instructions: Natural language instructions for refinement
            model_name: Optional name for the refined model (defaults to original name with 'Refined' prefix)

        Returns:
            A new refined Pydantic model with `.usage` attribute for token tracking
        """
        # Default model name if not provided
        if model_name is None:
            model_name = f"Refined{model.__name__}"

        refined_model = self.model_operations.refine_existing_model(
            model, refinement_instructions, model_name
        )

        # Create a deep copy of usage for the model
        model_usage = copy.deepcopy(self.llm_core.get_usage())

        # Reset the extractor's usage for the next operation
        self.llm_core.reset_usage()

        # Add usage to model
        refined_model.usage = model_usage

        return refined_model

    @classmethod
    def from_litellm(
        cls,
        *,
        model: str,
        api_key: Optional[str] = None,
        config: Optional[Union[Dict, str]] = None,
        max_threads: int = 10,
        batch_size: int = 100,
        max_retries: int = 3,
        min_wait: int = 1,
        max_wait: int = 10,
        **litellm_kwargs: Any,
    ) -> "Extractor":
        """
        Create Extractor instance using litellm.

        Args:
            model: Model identifier (e.g., "gpt-4", "claude-2", "azure/gpt-4")
            api_key: API key for the model provider
            config: Extraction configuration
            max_threads: Maximum number of concurrent threads
            batch_size: Size of processing batches
            max_retries: Maximum number of retries for extraction
            min_wait: Minimum seconds to wait between retries
            max_wait: Maximum seconds to wait between retries
            **litellm_kwargs: Additional kwargs for litellm (e.g., api_base, organization)
        """
        import instructor
        import litellm
        from litellm import completion

        # Set up litellm
        if api_key:
            litellm.api_key = api_key

        # drop unnecessary parameters
        litellm.drop_params = True

        # Set additional litellm configs
        for key, value in litellm_kwargs.items():
            setattr(litellm, key, value)

        # Create patched client
        client = instructor.from_litellm(completion)

        return cls(
            client=client,
            model_name=model,
            config=config,
            max_threads=max_threads,
            batch_size=batch_size,
            max_retries=max_retries,
            min_wait=min_wait,
            max_wait=max_wait,
        )

extract(*, data, query, model=None, return_df=False, expand_nested=False, **kwargs)

Extract structured data from text.

Parameters:

Name Type Description Default
data Union[str, Path, DataFrame, List[Dict[str, str]]]

Input data (file path, DataFrame, list of dicts, or raw text)

required
query str

Natural language query

required
model Optional[Type[BaseModel]]

Optional pre-generated Pydantic model class (if None, a model will be generated)

None
return_df bool

Whether to return DataFrame

False
expand_nested bool

Whether to flatten nested structures

False
**kwargs Any

Additional options for file reading

{}

Returns:

Type Description
ExtractionResult

Extraction result with extracted data, failed rows, and model (if requested)

Source code in structx/extraction/extractor.py
@handle_errors(error_message="Extraction failed", error_type=ExtractionError)
def extract(
    self,
    *,
    data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
    query: str,
    model: Optional[Type[BaseModel]] = None,
    return_df: bool = False,
    expand_nested: bool = False,
    **kwargs: Any,
) -> ExtractionResult:
    """
    Extract structured data from text.

    Args:
        data: Input data (file path, DataFrame, list of dicts, or raw text)
        query: Natural language query
        model: Optional pre-generated Pydantic model class (if None, a model will be generated)
        return_df: Whether to return DataFrame
        expand_nested: Whether to flatten nested structures
        **kwargs: Additional options for file reading

    Returns:
        Extraction result with extracted data, failed rows, and model (if requested)
    """
    df = self.data_processor.prepare_data(data, **kwargs)
    return self._process_data(df, query, return_df, expand_nested, model)

extract_async(*, data, query, return_df=False, expand_nested=False, **kwargs) async

Asynchronous version of extract.

Parameters:

Name Type Description Default
data Union[str, Path, DataFrame, List[Dict[str, str]]]

Input data (file path, DataFrame, list of dicts, or raw text)

required
query str

Natural language query

required
return_df bool

Whether to return DataFrame

False
expand_nested bool

Whether to flatten nested structures

False
**kwargs Any

Additional options for file reading

{}

Returns:

Type Description
ExtractionResult

ExtractionResult containing extracted data, failed rows, and the model

Source code in structx/extraction/extractor.py
async def extract_async(
    self,
    *,
    data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
    query: str,
    return_df: bool = False,
    expand_nested: bool = False,
    **kwargs: Any,
) -> ExtractionResult:
    """
    Asynchronous version of `extract`.

    Args:
        data: Input data (file path, DataFrame, list of dicts, or raw text)
        query: Natural language query
        return_df: Whether to return DataFrame
        expand_nested: Whether to flatten nested structures
        **kwargs: Additional options for file reading

    Returns:
        ExtractionResult containing extracted data, failed rows, and the model
    """
    return await self.data_processor.run_async(
        self.extract, data, query, None, return_df, expand_nested, **kwargs
    )

extract_queries(*, data, queries, return_df=True, expand_nested=False, **kwargs)

Process multiple queries on the same data.

Parameters:

Name Type Description Default
data Union[str, Path, DataFrame, List[Dict[str, str]]]

Input data (file path, DataFrame, list of dicts, or raw text)

required
queries List[str]

List of queries to process

required
return_df bool

Whether to return DataFrame

True
expand_nested bool

Whether to flatten nested structures

False
**kwargs Any

Additional options for file reading

{}

Returns:

Type Description
Dict[str, ExtractionResult]

Dictionary mapping queries to their results (extracted data and failed extractions)

Source code in structx/extraction/extractor.py
@handle_errors(error_message="Batch extraction failed", error_type=ExtractionError)
def extract_queries(
    self,
    *,
    data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
    queries: List[str],
    return_df: bool = True,
    expand_nested: bool = False,
    **kwargs: Any,
) -> Dict[str, ExtractionResult]:
    """
    Process multiple queries on the same data.

    Args:
        data: Input data (file path, DataFrame, list of dicts, or raw text)
        queries: List of queries to process
        return_df: Whether to return DataFrame
        expand_nested: Whether to flatten nested structures
        **kwargs: Additional options for file reading

    Returns:
        Dictionary mapping queries to their results (extracted data and failed extractions)
    """
    results = {}

    for query in queries:
        logger.info(f"\nProcessing query: {query}")
        result = self.extract(
            data=data,
            query=query,
            return_df=return_df,
            expand_nested=expand_nested,
            **kwargs,
        )
        results[query] = result

    return results

extract_queries_async(*, data, queries, return_df=False, expand_nested=False, **kwargs) async

Asynchronous version of extract_queries.

Parameters:

Name Type Description Default
data Union[str, Path, DataFrame, List[Dict[str, str]]]

Input data

required
queries List[str]

List of queries

required
return_df bool

Whether to return DataFrame

False
expand_nested bool

Whether to flatten nested structures

False
**kwargs Any

Additional options

{}

Returns:

Type Description
Dict[str, ExtractionResult]

Dictionary mapping queries to ExtractionResult objects

Source code in structx/extraction/extractor.py
async def extract_queries_async(
    self,
    *,
    data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
    queries: List[str],
    return_df: bool = False,
    expand_nested: bool = False,
    **kwargs: Any,
) -> Dict[str, ExtractionResult]:
    """
    Asynchronous version of `extract_queries`.

    Args:
        data: Input data
        queries: List of queries
        return_df: Whether to return DataFrame
        expand_nested: Whether to flatten nested structures
        **kwargs: Additional options

    Returns:
        Dictionary mapping queries to ExtractionResult objects
    """
    return await self.data_processor.run_async(
        self.extract_queries, data, queries, return_df, expand_nested, **kwargs
    )

from_litellm(*, model, api_key=None, config=None, max_threads=10, batch_size=100, max_retries=3, min_wait=1, max_wait=10, **litellm_kwargs) classmethod

Create Extractor instance using litellm.

Parameters:

Name Type Description Default
model str

Model identifier (e.g., "gpt-4", "claude-2", "azure/gpt-4")

required
api_key Optional[str]

API key for the model provider

None
config Optional[Union[Dict, str]]

Extraction configuration

None
max_threads int

Maximum number of concurrent threads

10
batch_size int

Size of processing batches

100
max_retries int

Maximum number of retries for extraction

3
min_wait int

Minimum seconds to wait between retries

1
max_wait int

Maximum seconds to wait between retries

10
**litellm_kwargs Any

Additional kwargs for litellm (e.g., api_base, organization)

{}
Source code in structx/extraction/extractor.py
@classmethod
def from_litellm(
    cls,
    *,
    model: str,
    api_key: Optional[str] = None,
    config: Optional[Union[Dict, str]] = None,
    max_threads: int = 10,
    batch_size: int = 100,
    max_retries: int = 3,
    min_wait: int = 1,
    max_wait: int = 10,
    **litellm_kwargs: Any,
) -> "Extractor":
    """
    Create Extractor instance using litellm.

    Args:
        model: Model identifier (e.g., "gpt-4", "claude-2", "azure/gpt-4")
        api_key: API key for the model provider
        config: Extraction configuration
        max_threads: Maximum number of concurrent threads
        batch_size: Size of processing batches
        max_retries: Maximum number of retries for extraction
        min_wait: Minimum seconds to wait between retries
        max_wait: Maximum seconds to wait between retries
        **litellm_kwargs: Additional kwargs for litellm (e.g., api_base, organization)
    """
    import instructor
    import litellm
    from litellm import completion

    # Set up litellm
    if api_key:
        litellm.api_key = api_key

    # drop unnecessary parameters
    litellm.drop_params = True

    # Set additional litellm configs
    for key, value in litellm_kwargs.items():
        setattr(litellm, key, value)

    # Create patched client
    client = instructor.from_litellm(completion)

    return cls(
        client=client,
        model_name=model,
        config=config,
        max_threads=max_threads,
        batch_size=batch_size,
        max_retries=max_retries,
        min_wait=min_wait,
        max_wait=max_wait,
    )

get_schema(*, data, query, **kwargs)

Get extraction model without performing extraction.

Parameters:

Name Type Description Default
query str

Natural language query

required
data Union[str, Path, DataFrame, List[Dict[str, str]]]

Input data (file path, DataFrame, list of dicts, or raw text)

required
**kwargs Any

Additional options for file reading

{}

Returns:

Type Description
Type[BaseModel]

Pydantic model for extraction with .usage attribute for token tracking

Source code in structx/extraction/extractor.py
@handle_errors(error_message="Schema generation failed", error_type=ExtractionError)
def get_schema(
    self,
    *,
    data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
    query: str,
    **kwargs: Any,
) -> Type[BaseModel]:
    """
    Get extraction model without performing extraction.

    Args:
        query: Natural language query
        data: Input data (file path, DataFrame, list of dicts, or raw text)
        **kwargs: Additional options for file reading

    Returns:
        Pydantic model for extraction with `.usage` attribute for token tracking
    """
    if isinstance(data, str) and not Path(data).exists():
        sample_text = data
        columns = ["text"]
    else:
        df = self.data_processor.prepare_data(data, **kwargs)
        is_file_based = "pdf_path" in df.columns or "source" in df.columns
        columns = df.columns.tolist()

        if is_file_based:
            sample_text = self.content_analyzer.extract_content_sample_for_schema(
                df
            )
            content_context = self.content_analyzer.detect_content_type_and_context(
                df
            )
            sample_text = f"Content type: {content_context}\n\n{sample_text}"
        else:
            # For traditional tabular data, create a representative sample
            sample_text = "\n".join(df.head().to_string(index=False).splitlines())

    # Refine query
    refined_query = self.llm_core.refine_query(query)

    # Generate guide
    guide = self.llm_core.generate_extraction_guide(refined_query, columns)

    # Generate schema
    schema_request = self.model_operations.generate_extraction_schema(
        sample_text, refined_query, guide
    )

    # Create model
    extraction_model = self.model_operations.create_model_from_schema(
        schema_request
    )

    # Create a deep copy of usage for the model
    model_usage = copy.deepcopy(self.llm_core.get_usage())

    # Reset the extractor's usage for the next operation
    self.llm_core.reset_usage()

    # Add usage to model
    extraction_model.usage = model_usage

    return extraction_model

get_schema_async(*, data, query, **kwargs) async

Asynchronous version of get_schema.

Parameters:

Name Type Description Default
query str

Natural language query

required
data Union[str, Path, DataFrame, List[Dict[str, str]]]

Input data (file path, DataFrame, list of dicts, or raw text)

required
**kwargs Any

Additional options for file reading

{}

Returns:

Type Description
Type[BaseModel]

Dynamically generated Pydantic model class

Source code in structx/extraction/extractor.py
async def get_schema_async(
    self,
    *,
    data: Union[str, Path, pd.DataFrame, List[Dict[str, str]]],
    query: str,
    **kwargs: Any,
) -> Type[BaseModel]:
    """
    Asynchronous version of `get_schema`.

    Args:
        query: Natural language query
        data: Input data (file path, DataFrame, list of dicts, or raw text)
        **kwargs: Additional options for file reading

    Returns:
        Dynamically generated Pydantic model class
    """
    return await self.data_processor.run_async(
        self.get_schema, query, data, **kwargs
    )

refine_data_model(*, model, refinement_instructions, model_name=None)

Refine an existing data model based on natural language instructions.

Parameters:

Name Type Description Default
model Type[BaseModel]

Existing Pydantic model to refine

required
refinement_instructions str

Natural language instructions for refinement

required
model_name Optional[str]

Optional name for the refined model (defaults to original name with 'Refined' prefix)

None

Returns:

Type Description
Type[BaseModel]

A new refined Pydantic model with .usage attribute for token tracking

Source code in structx/extraction/extractor.py
def refine_data_model(
    self,
    *,
    model: Type[BaseModel],
    refinement_instructions: str,
    model_name: Optional[str] = None,
) -> Type[BaseModel]:
    """
    Refine an existing data model based on natural language instructions.

    Args:
        model: Existing Pydantic model to refine
        refinement_instructions: Natural language instructions for refinement
        model_name: Optional name for the refined model (defaults to original name with 'Refined' prefix)

    Returns:
        A new refined Pydantic model with `.usage` attribute for token tracking
    """
    # Default model name if not provided
    if model_name is None:
        model_name = f"Refined{model.__name__}"

    refined_model = self.model_operations.refine_existing_model(
        model, refinement_instructions, model_name
    )

    # Create a deep copy of usage for the model
    model_usage = copy.deepcopy(self.llm_core.get_usage())

    # Reset the extractor's usage for the next operation
    self.llm_core.reset_usage()

    # Add usage to model
    refined_model.usage = model_usage

    return refined_model