Skip to content

Creating Custom BlendSQL Ingredients

All the built-in LLM ingredients inherit from the base classes QAIngredient, MapIngredient, JoinIngredient, and AliasIngredient.

These are intended to be helpful abstractions, so that the user can easily implement their own functions to run within a BlendSQL script.

The processing logic for a custom ingredient should go in a run() class function, and accept **kwargs in their signature.

AliasIngredient

Bases: Ingredient

This ingredient performs no other function than to act as a stand-in for complex chainings of other ingredients. This allows us (or our lms) to write less verbose BlendSQL queries, while maximizing the information we embed.

The run() function should return a tuple containing both the query text that should get subbed in, and any ingredient classes which are dependencies for executing the aliased query.

Examples:

from textwrap import dedent
from typing import Tuple, Collection

from blendsql.ingredients import AliasIngredient, LLMQA

class FetchDefinition(AliasIngredient):
    def run(self, term: str, *args, **kwargs) -> Tuple[str, Collection[Ingredient]]:
        new_query = dedent(
        f"""
        {{{{
            LLMQA(
                "What does {term} mean?"
            )
        }}}}
        """)
        ingredient_dependencies = {LLMQA}
        return (new_query, ingredient_dependencies)

# Now, we can use the ingredient like below
blendsql_query = """
SELECT {{FetchDefinition('delve')}} AS "Definition"
"""
Source code in blendsql/ingredients/ingredient.py
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
@attrs
class AliasIngredient(Ingredient):
    '''This ingredient performs no other function than to act as a stand-in for
    complex chainings of other ingredients. This allows us (or our lms) to write less verbose
    BlendSQL queries, while maximizing the information we embed.

    The `run()` function should return a tuple containing both the query text that should get subbed in,
    and any ingredient classes which are dependencies for executing the aliased query.

    Examples:
        ```python
        from textwrap import dedent
        from typing import Tuple, Collection

        from blendsql.ingredients import AliasIngredient, LLMQA

        class FetchDefinition(AliasIngredient):
            def run(self, term: str, *args, **kwargs) -> Tuple[str, Collection[Ingredient]]:
                new_query = dedent(
                f"""
                {{{{
                    LLMQA(
                        "What does {term} mean?"
                    )
                }}}}
                """)
                ingredient_dependencies = {LLMQA}
                return (new_query, ingredient_dependencies)

        # Now, we can use the ingredient like below
        blendsql_query = """
        SELECT {{FetchDefinition('delve')}} AS "Definition"
        """
        ```
    '''

    ingredient_type: str = IngredientType.ALIAS.value
    allowed_output_types: t.Tuple[t.Type] = (t.Tuple[str, Collection[Ingredient]],)

    def __call__(self, *args, **kwargs):
        return self._run(*args, **kwargs)

QAIngredient

Bases: Ingredient

Given a table subset in the form of a pd.DataFrame 'context', returns a scalar or array of scalars (in the form of a tuple).

Useful for end-to-end question answering tasks.

Examples:

import pandas as pd
import guidance

from blendsql.models import Model, LocalModel, RemoteModel
from blendsql.ingredients import QAIngredient
from blendsql._program import Program


class SummaryProgram(Program):
    """Program to call Model and return summary of the passed table.
    """

    def __call__(self, model: Model, serialized_db: str):
        prompt = f"Summarize the table. {serialized_db}"
        if isinstance(model, LocalModel):
            # Below we follow the guidance pattern for unconstrained text generation
            # https://github.com/guidance-ai/guidance
            response = (model.model_obj + guidance.gen(max_tokens=20, name="response"))._variables["response"]
        else:
            response = model.generate(
                messages_list=[[{"role": "user", "content": prompt}]],
                max_tokens=20
            )[0]
        # Finally, return (response, prompt) tuple
        # Returning the prompt here allows the underlying BlendSQL classes to track token usage
        return (response, prompt)


    class TableSummary(QAIngredient):
        def run(self, model: Model, context: pd.DataFrame, **kwargs) -> str:
            result = model.predict(program=SummaryProgram, serialized_db=context.to_string())
            return f"'{result}'"


    if __name__ == "__main__":
        from blendsql import blend
        from blendsql.db import SQLite
        from blendsql.utils import fetch_from_hub
        from blendsql.models import LiteLLM

        blendsql = """
        SELECT {{
            TableSummary(
                context=(SELECT * FROM transactions LIMIT 10)
            )
        }} AS "Summary"
        """

        smoothie = blend(
            query=blendsql,
            default_model=LiteLLM("openai/gpt-4o-mini"),
            db=SQLite(fetch_from_hub("single_table.db")),
            ingredients={TableSummary}
        )
        # Now, we can get results
        print(smoothie.df)
        # 'The table summarizes a series of cash flow transactions made through Zelle'
        # ...and token usage
        print(smoothie.meta.prompt_tokens)
        print(smoothie.meta.completion_tokens)
Source code in blendsql/ingredients/ingredient.py
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
@attrs
class QAIngredient(Ingredient):
    '''
    Given a table subset in the form of a pd.DataFrame 'context',
    returns a scalar or array of scalars (in the form of a tuple).

    Useful for end-to-end question answering tasks.

    Examples:
        ```python
        import pandas as pd
        import guidance

        from blendsql.models import Model, LocalModel, RemoteModel
        from blendsql.ingredients import QAIngredient
        from blendsql._program import Program


        class SummaryProgram(Program):
            """Program to call Model and return summary of the passed table.
            """

            def __call__(self, model: Model, serialized_db: str):
                prompt = f"Summarize the table. {serialized_db}"
                if isinstance(model, LocalModel):
                    # Below we follow the guidance pattern for unconstrained text generation
                    # https://github.com/guidance-ai/guidance
                    response = (model.model_obj + guidance.gen(max_tokens=20, name="response"))._variables["response"]
                else:
                    response = model.generate(
                        messages_list=[[{"role": "user", "content": prompt}]],
                        max_tokens=20
                    )[0]
                # Finally, return (response, prompt) tuple
                # Returning the prompt here allows the underlying BlendSQL classes to track token usage
                return (response, prompt)


            class TableSummary(QAIngredient):
                def run(self, model: Model, context: pd.DataFrame, **kwargs) -> str:
                    result = model.predict(program=SummaryProgram, serialized_db=context.to_string())
                    return f"'{result}'"


            if __name__ == "__main__":
                from blendsql import blend
                from blendsql.db import SQLite
                from blendsql.utils import fetch_from_hub
                from blendsql.models import LiteLLM

                blendsql = """
                SELECT {{
                    TableSummary(
                        context=(SELECT * FROM transactions LIMIT 10)
                    )
                }} AS "Summary"
                """

                smoothie = blend(
                    query=blendsql,
                    default_model=LiteLLM("openai/gpt-4o-mini"),
                    db=SQLite(fetch_from_hub("single_table.db")),
                    ingredients={TableSummary}
                )
                # Now, we can get results
                print(smoothie.df)
                # 'The table summarizes a series of cash flow transactions made through Zelle'
                # ...and token usage
                print(smoothie.meta.prompt_tokens)
                print(smoothie.meta.completion_tokens)
        ```
    '''

    ingredient_type: str = IngredientType.QA.value
    allowed_output_types: t.Tuple[t.Type] = (t.Union[str, int, float, tuple],)

    def __call__(
        self,
        question: t.Optional[str] = None,
        context: t.Optional[t.Union[str, pd.DataFrame]] = None,
        options: t.Optional[t.Union[list, str]] = None,
        *args,
        **kwargs,
    ) -> t.Tuple[t.Union[str, int, float, tuple], t.Optional[exp.Expression]]:
        # Unpack kwargs
        aliases_to_tablenames: t.Dict[str, str] = kwargs["aliases_to_tablenames"]

        subtable: t.Union[pd.DataFrame, None] = None
        if context is not None:
            if isinstance(context, str):
                tablename, colname = utils.get_tablename_colname(context)
                tablename = aliases_to_tablenames.get(tablename, tablename)
                # Optionally materialize a CTE
                if tablename in self.db.lazy_tables:
                    subtable: pd.DataFrame = pd.DataFrame(
                        self.db.lazy_tables.pop(tablename).collect()[colname]
                    )
                else:
                    subtable: pd.DataFrame = self.db.execute_to_df(
                        f'SELECT "{colname}" FROM "{tablename}"'
                    )
            elif isinstance(context, pd.DataFrame):
                subtable: pd.DataFrame = context
            else:
                raise ValueError(
                    f"Unknown type for `identifier` arg in QAIngredient: {type(context)}"
                )
            if subtable.empty:
                raise IngredientException("Empty subtable passed to QAIngredient!")

        self.num_values_passed += len(subtable) if subtable is not None else 0

        if options is not None:
            options = self.unpack_options(
                options=options,
                aliases_to_tablenames=aliases_to_tablenames,
            )

        if question is not None:
            question = self.unpack_question(
                question=question, aliases_to_tablenames=aliases_to_tablenames
            )

        response: t.Union[str, int, float, tuple] = self._run(
            question=question,
            context=subtable,
            options=options,
            *args,
            **self.__dict__ | kwargs,
        )
        if isinstance(response, tuple):
            response = format_tuple(
                response, kwargs.get("wrap_tuple_in_parentheses", True)
            )
        return response

    @abstractmethod
    def run(self, *args, **kwargs) -> t.Union[str, int, float, tuple]:
        ...

MapIngredient

Bases: Ingredient

For a given table/column pair, maps an external function to each of the given values, creating a new column.

Examples:

from typing import List
from blendsql.ingredients import MapIngredient
import requests


class GetQRCode(MapIngredient):
    """Calls API to generate QR code for a given URL.
    Saves bytes to file in qr_codes/ and returns list of paths.
    https://goqr.me/api/doc/create-qr-code/"""


    def run(self, values: List[str], **kwargs) -> List[str]:
        imgs_as_bytes = []
        for value in values:
            qr_code_bytes = requests.get(
                "https://api.qrserver.com/v1/create-qr-code/?data=https://{}/&size=100x100".format(value)
            ).content
            imgs_as_bytes.append(qr_code_bytes)
        return imgs_as_bytes


    if __name__ == "__main__":
        from blendsql import BlendSQL
        from blendsql.db import SQLite
        from blendsql.utils import fetch_from_hub

        bsql = BlendSQL(fetch_from_hub('urls.db'), ingredients={GetQRCode})

        smoothie = bsql.execute("SELECT genre, url, {{GetQRCode('QR Code as Bytes:', 'w::url')}} FROM w WHERE genre = 'social'")

        smoothie.df
        # | genre  | url           | QR Code as Bytes:      |
        # |--------|---------------|-----------------------|
        # | social | facebook.com  | b'...'                |
Source code in blendsql/ingredients/ingredient.py
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
@attrs
class MapIngredient(Ingredient):
    '''For a given table/column pair, maps an external function
    to each of the given values, creating a new column.

    Examples:
        ```python
        from typing import List
        from blendsql.ingredients import MapIngredient
        import requests


        class GetQRCode(MapIngredient):
            """Calls API to generate QR code for a given URL.
            Saves bytes to file in qr_codes/ and returns list of paths.
            https://goqr.me/api/doc/create-qr-code/"""


            def run(self, values: List[str], **kwargs) -> List[str]:
                imgs_as_bytes = []
                for value in values:
                    qr_code_bytes = requests.get(
                        "https://api.qrserver.com/v1/create-qr-code/?data=https://{}/&size=100x100".format(value)
                    ).content
                    imgs_as_bytes.append(qr_code_bytes)
                return imgs_as_bytes


            if __name__ == "__main__":
                from blendsql import BlendSQL
                from blendsql.db import SQLite
                from blendsql.utils import fetch_from_hub

                bsql = BlendSQL(fetch_from_hub('urls.db'), ingredients={GetQRCode})

                smoothie = bsql.execute("SELECT genre, url, {{GetQRCode('QR Code as Bytes:', 'w::url')}} FROM w WHERE genre = 'social'")

                smoothie.df
                # | genre  | url           | QR Code as Bytes:      |
                # |--------|---------------|-----------------------|
                # | social | facebook.com  | b'...'                |
        ```
    '''

    ingredient_type: str = IngredientType.MAP.value
    allowed_output_types: t.Tuple[t.Type] = (t.Iterable[t.Any],)

    def unpack_default_kwargs(self, **kwargs):
        return unpack_default_kwargs(**kwargs)

    def __call__(
        self,
        question: t.Optional[str] = None,
        values: t.Optional[ValueArray] = None,
        options: t.Optional[ValueArray] = None,
        *args,
        **kwargs,
    ) -> tuple:
        """Returns tuple with format (arg, tablename, colname, new_table)"""
        # Unpack kwargs
        aliases_to_tablenames: t.Dict[str, str] = kwargs["aliases_to_tablenames"]
        get_temp_subquery_table: t.Callable = kwargs["get_temp_subquery_table"]
        get_temp_session_table: t.Callable = kwargs["get_temp_session_table"]
        prev_subquery_map_columns: t.Set[str] = kwargs["prev_subquery_map_columns"]

        # TODO: make sure we support all types of ValueArray references here
        tablename, colname = utils.get_tablename_colname(values)
        tablename = aliases_to_tablenames.get(tablename, tablename)

        # Check for previously created temporary tables
        value_source_tablename, _ = self.maybe_get_temp_table(
            temp_table_func=get_temp_subquery_table, tablename=tablename
        )
        temp_session_tablename, temp_session_table_exists = self.maybe_get_temp_table(
            temp_table_func=get_temp_session_table, tablename=tablename
        )

        # Optionally materialize a CTE
        if tablename in self.db.lazy_tables:
            original_table = self.db.lazy_tables.pop(tablename).collect()
        else:
            original_table = self.db.execute_to_df(
                select_all_from_table_query(tablename)
            )

        # Need to be sure the new column doesn't already exist here
        new_arg_column = question or str(uuid.uuid4())[:4]
        while (
            new_arg_column in set(self.db.iter_columns(tablename))
            or new_arg_column in prev_subquery_map_columns
        ):
            new_arg_column = "_" + new_arg_column

        # Get a list of values to map
        # First, check if we've already dumped some `MapIngredient` output to the main session table
        if temp_session_table_exists:
            temp_session_table = self.db.execute_to_df(
                select_all_from_table_query(temp_session_tablename)
            )
            # We don't need to run this function on everything,
            #   if a previous subquery already got to certain values
            if new_arg_column in temp_session_table.columns:
                unpacked_values = self.db.execute_to_list(
                    f'SELECT DISTINCT "{colname}" FROM "{temp_session_tablename}" WHERE "{new_arg_column}" IS NULL',
                )
            # Base case: this is the first time we've used this particular ingredient
            # BUT, temp_session_tablename still exists
            else:
                unpacked_values = self.db.execute_to_list(
                    f'SELECT DISTINCT "{colname}" FROM "{temp_session_tablename}"',
                )
        else:
            unpacked_values = self.db.execute_to_list(
                f'SELECT DISTINCT "{colname}" FROM "{value_source_tablename}"',
            )

        # No need to run ingredient if we have no values to map onto
        if len(unpacked_values) == 0:
            original_table[new_arg_column] = None
            return (new_arg_column, tablename, colname, original_table)

        if options is not None:
            # Override any pattern with our new unpacked options
            options = self.unpack_options(
                options=options,
                aliases_to_tablenames=aliases_to_tablenames,
            )

        if question is not None:
            question = self.unpack_question(
                question=question, aliases_to_tablenames=aliases_to_tablenames
            )

        mapped_values: Collection[t.Any] = self._run(
            question=question,
            values=unpacked_values,
            options=options,
            tablename=tablename,
            colname=colname,
            *args,
            **self.__dict__ | kwargs,
        )
        self.num_values_passed += len(mapped_values)
        df_as_dict: t.Dict[str, list] = {colname: [], new_arg_column: []}
        for value, mapped_value in zip(unpacked_values, mapped_values):
            df_as_dict[colname].append(value)
            df_as_dict[new_arg_column].append(mapped_value)
        subtable = pd.DataFrame(df_as_dict)
        if all(
            isinstance(x, (int, type(None))) and not isinstance(x, bool)
            for x in mapped_values
        ):
            subtable[new_arg_column] = subtable[new_arg_column].astype("Int64")
        # Add new_table to original table
        new_table = original_table.merge(subtable, how="left", on=colname)
        if new_table.shape[0] != original_table.shape[0]:
            raise IngredientException(
                f"subtable from MapIngredient.run() needs same length as # rows from original\nOriginal has {original_table.shape[0]}, new_table has {new_table.shape[0]}"
            )
        # Now, new table has original columns + column with the name of the question we answered
        return (new_arg_column, tablename, colname, new_table)

    @abstractmethod
    def run(self, *args, **kwargs) -> Iterable[t.Any]:
        ...

JoinIngredient

Bases: Ingredient

Executes an INNER JOIN using dict mapping. 'Join on color of food'

Examples:

from blendsql.ingredients import JoinIngredient

class do_join(JoinIngredient):
    """A very silly, overcomplicated way to do a traditional SQL join.
    But useful for testing.
    """

    def run(self, left_values: List[str], right_values: List[str], **kwargs) -> dict:
        return {left_value: left_value for left_value in left_values}

blendsql_query = """
SELECT Account, Quantity FROM returns
JOIN {{
    do_join(
        left_on='account_history::Symbol',
        right_on='returns::Symbol'
    )
}}
"""
Source code in blendsql/ingredients/ingredient.py
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
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
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
@attrs
class JoinIngredient(Ingredient):
    '''Executes an `INNER JOIN` using dict mapping.
    'Join on color of food'
    {"tomato": "red", "broccoli": "green", "lemon": "yellow"}

    Examples:
        ```python
        from blendsql.ingredients import JoinIngredient

        class do_join(JoinIngredient):
            """A very silly, overcomplicated way to do a traditional SQL join.
            But useful for testing.
            """

            def run(self, left_values: List[str], right_values: List[str], **kwargs) -> dict:
                return {left_value: left_value for left_value in left_values}

        blendsql_query = """
        SELECT Account, Quantity FROM returns
        JOIN {{
            do_join(
                left_on='account_history::Symbol',
                right_on='returns::Symbol'
            )
        }}
        """
        ```
    '''

    use_skrub_joiner: bool = attrib(default=True)

    ingredient_type: str = IngredientType.JOIN.value
    allowed_output_types: t.Tuple[t.Type] = (dict,)

    def __call__(
        self,
        left_on: t.Optional[str] = None,
        right_on: t.Optional[str] = None,
        join_criteria: t.Optional[str] = None,
        *args,
        **kwargs,
    ) -> tuple:
        # Unpack kwargs
        aliases_to_tablenames: t.Dict[str, str] = kwargs["aliases_to_tablenames"]
        get_temp_subquery_table: t.Callable = kwargs["get_temp_subquery_table"]
        get_temp_session_table: t.Callable = kwargs["get_temp_session_table"]
        # Depending on the size of the underlying data, it may be optimal to swap
        #   the order of 'left_on' and 'right_on' columns during processing
        swapped = False
        values = []
        original_lr_identifiers = []
        modified_lr_identifiers = []
        mapping: t.Dict[str, str] = {}
        for on_arg in [left_on, right_on]:
            # Since LLMJoin is unique, in that we need to inject the referenced tablenames back to the query,
            #   make sure we keep the `referenced_tablename` variable.
            # So the below works:
            #     SELECT f.name, colors.name FROM fruits f
            #     JOIN {{LLMJoin('f::name', 'colors::name', join_criteria='Align the fruit to its color')}}
            referenced_tablename, colname = utils.get_tablename_colname(on_arg)
            tablename = aliases_to_tablenames.get(
                referenced_tablename, referenced_tablename
            )
            original_lr_identifiers.append((referenced_tablename, colname))
            tablename, _ = self.maybe_get_temp_table(
                temp_table_func=get_temp_subquery_table,
                tablename=tablename,
            )
            values.append(
                self.db.execute_to_list(
                    f'SELECT DISTINCT "{colname}" FROM "{tablename}"', to_type=str
                )
            )
            modified_lr_identifiers.append((tablename, colname))
        sorted_values = sorted(values, key=len)
        # check swapping only once, at the beginning
        if sorted_values != values:
            swapped = True
        if join_criteria is None:
            # First, check which values we actually need to call Model on
            # We don't want to join when there's already an intuitive alignment
            # First, make sure outer loop is shorter of the two lists
            outer, inner = sorted_values
            _outer = []
            inner = set(inner)
            mapping = {}
            for l in outer:
                if l in inner:
                    # Define this mapping, and remove from Model inference call
                    mapping[l] = l
                    inner.remove(l)
                else:
                    _outer.append(l)
                if len(inner) == 0:
                    break
            # Remained _outer and inner lists preserved the sorting order in length:
            # len(_outer) = len(outer) - #matched <= len(inner original) - matched = len(inner)
            if self.use_skrub_joiner and all(len(x) > 1 for x in [inner, _outer]):
                from skrub import Joiner

                # Create the main_table DataFrame
                main_table = pd.DataFrame(_outer, columns=["out"])
                # Create the aux_table DataFrame
                aux_table = pd.DataFrame(inner, columns=["in"])
                joiner = Joiner(
                    aux_table,
                    main_key="out",
                    aux_key="in",
                    max_dist=0.9,
                    add_match_info=False,
                )
                res = joiner.fit_transform(main_table)
                # Below is essentially set.difference on aux_table and those paired in res
                inner = aux_table.loc[~aux_table["in"].isin(res["in"]), "in"].tolist()
                # length(new inner) = length(inner) - #matched by fuzzy join
                _outer = res["out"][res["in"].isnull()].to_list()
                # length(new _outer) = length(_outer) - #matched by fuzzy join
                _skrub_mapping = (
                    res.dropna(subset=["in"]).set_index("out")["in"].to_dict()
                )
                logger.debug(
                    Fore.YELLOW
                    + "Made the following alignment with `skrub.Joiner`:"
                    + Fore.RESET
                )
                logger.debug(
                    Fore.YELLOW + json.dumps(_skrub_mapping, indent=4) + Fore.RESET
                )
                mapping = mapping | _skrub_mapping
            # order by length is still preserved regardless of using fuzzy join, so after initial matching and possible fuzzy join matching
            # This is because the lengths of each list will decrease at the same rate, so whichever list was larger at the beginning,
            # will be larger here at the end.
            # len(_outer) <= len(inner)
            sorted_values = [_outer, inner]

        # Now, we have our final values to process.
        left_values, right_values = sorted_values

        (left_tablename, left_colname), (
            right_tablename,
            right_colname,
        ) = original_lr_identifiers
        (_left_tablename, _left_colname), (
            _right_tablename,
            _right_colname,
        ) = modified_lr_identifiers

        if all(len(x) > 0 for x in [left_values, right_values]):
            # Some alignment still left to do
            self.num_values_passed += len(left_values) + len(right_values)

            _predicted_mapping: t.Dict[str, str] = self._run(
                left_values=left_values,
                right_values=right_values,
                join_criteria=join_criteria,
                *args,
                **self.__dict__ | kwargs,
            )
            mapping = mapping | _predicted_mapping
        # Using mapped left/right values, create intermediary mapping table
        temp_join_tablename = get_temp_session_table(str(uuid.uuid4())[:4])
        # Below, we check to see if 'swapped' is True
        # If so, we need to inverse what is 'left', and what is 'right'
        joined_values_df = pd.DataFrame(
            data={
                "left" if not swapped else "right": mapping.keys(),
                "right" if not swapped else "left": mapping.values(),
            }
        )
        self.db.to_temp_table(df=joined_values_df, tablename=temp_join_tablename)
        return (
            left_tablename,
            right_tablename,
            f"""JOIN "{temp_join_tablename}" ON "{left_tablename}"."{left_colname}" = "{temp_join_tablename}".left
              JOIN "{right_tablename}" ON "{right_tablename}"."{right_colname}" = "{temp_join_tablename}".right
              """,
            temp_join_tablename,
        )

    @abstractmethod
    def run(self, *args, **kwargs) -> dict:
        ...