palimpzest 0.7.21__py3-none-any.whl → 0.8.0__py3-none-any.whl

palimpzest/query/operators/distinct.py

@@ -0,0 +1,62 @@
+from __future__ import annotations
+
+from palimpzest.core.elements.records import DataRecord, DataRecordSet
+from palimpzest.core.models import OperatorCostEstimates, RecordOpStats
+from palimpzest.query.operators.physical import PhysicalOperator
+
+
+class DistinctOp(PhysicalOperator):
+    def __init__(self, distinct_cols: list[str], distinct_seen: set | None = None, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.distinct_cols = distinct_cols
+        self._distinct_seen = set() if distinct_seen is None else distinct_seen
+
+    def __str__(self):
+        op = super().__str__()
+        op += f"    Distinct Cols: {self.distinct_cols}\n"
+        return op
+
+    def get_id_params(self):
+        id_params = super().get_id_params()
+        return {"distinct_cols": self.distinct_cols, **id_params}
+
+    def get_op_params(self):
+        op_params = super().get_op_params()
+        return {"distinct_cols": self.distinct_cols, "distinct_seen": self._distinct_seen, **op_params}
+
+    def naive_cost_estimates(self, source_op_cost_estimates: OperatorCostEstimates) -> OperatorCostEstimates:
+        # assume applying the distinct operator takes negligible additional time (and no cost in USD)
+        return OperatorCostEstimates(
+            cardinality=source_op_cost_estimates.cardinality,
+            time_per_record=0,
+            cost_per_record=0,
+            quality=1.0,
+        )
+
+    def __call__(self, candidate: DataRecord) -> DataRecordSet:
+        # create new DataRecord
+        dr = DataRecord.from_parent(schema=candidate.schema, parent_record=candidate)
+
+        # output record only if it has not been seen before
+        record_str = dr.to_json_str(project_cols=self.distinct_cols, bytes_to_str=True, sorted=True)
+        record_hash = f"{hash(record_str)}"
+        dr.passed_operator = record_hash not in self._distinct_seen
+        if dr.passed_operator:
+            self._distinct_seen.add(record_hash)
+
+        # create RecordOpStats object
+        record_op_stats = RecordOpStats(
+            record_id=dr.id,
+            record_parent_ids=dr.parent_ids,
+            record_source_indices=dr.source_indices,
+            record_state=dr.to_dict(include_bytes=False),
+            full_op_id=self.get_full_op_id(),
+            logical_op_id=self.logical_op_id,
+            op_name=self.op_name(),
+            time_per_record=0.0,
+            cost_per_record=0.0,
+            passed_operator=dr.passed_operator,
+            op_details={k: str(v) for k, v in self.get_id_params().items()},
+        )
+
+        return DataRecordSet([dr], [record_op_stats])

palimpzest/query/operators/filter.py

@@ -4,6 +4,8 @@ import time
 from abc import ABC, abstractmethod
 from typing import Any
 
+from pydantic.fields import FieldInfo
+
 from palimpzest.constants import (
     MODEL_CARDS,
     NAIVE_EST_FILTER_SELECTIVITY,
@@ -12,19 +14,17 @@ from palimpzest.constants import (
     Model,
     PromptStrategy,
 )
-from palimpzest.core.data.dataclasses import GenerationStats, OperatorCostEstimates, RecordOpStats
 from palimpzest.core.elements.filters import Filter
 from palimpzest.core.elements.records import DataRecord, DataRecordSet
-from palimpzest.core.lib.fields import BooleanField
-from palimpzest.query.generators.generators import generator_factory
+from palimpzest.core.models import GenerationStats, OperatorCostEstimates, RecordOpStats
+from palimpzest.query.generators.generators import Generator
 from palimpzest.query.operators.physical import PhysicalOperator
-from palimpzest.utils.model_helpers import get_vision_models
 
 
 class FilterOp(PhysicalOperator, ABC):
     def __init__(self, filter: Filter, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        assert self.input_schema.get_desc() == self.output_schema.get_desc(), "Input and output schemas must match for FilterOp"
+        assert self.input_schema == self.output_schema, "Input and output schemas must match for FilterOp"
         self.filter_obj = filter
 
     def __str__(self):
@@ -81,8 +81,8 @@ class FilterOp(PhysicalOperator, ABC):
         # create RecordOpStats object
         record_op_stats = RecordOpStats(
             record_id=dr.id,
-            record_parent_id=dr.parent_id,
-            record_source_idx=dr.source_idx,
+            record_parent_ids=dr.parent_ids,
+            record_source_indices=dr.source_indices,
             record_state=dr.to_dict(include_bytes=False),
             full_op_id=self.get_full_op_id(),
             logical_op_id=self.logical_op_id,
@@ -174,19 +174,22 @@ class LLMFilter(FilterOp):
         self,
         model: Model,
         prompt_strategy: PromptStrategy = PromptStrategy.COT_BOOL,
+        reasoning_effort: str | None = None,
         *args,
         **kwargs,
     ):
         super().__init__(*args, **kwargs)
         self.model = model
         self.prompt_strategy = prompt_strategy
-        self.generator = generator_factory(model, prompt_strategy, Cardinality.ONE_TO_ONE, self.verbose)
+        self.reasoning_effort = reasoning_effort
+        self.generator = Generator(model, prompt_strategy, reasoning_effort, self.api_base, Cardinality.ONE_TO_ONE, self.verbose)
 
     def get_id_params(self):
         id_params = super().get_id_params()
         id_params = {
             "model": self.model.value,
             "prompt_strategy": self.prompt_strategy.value,
+            "reasoning_effort": self.reasoning_effort,
             **id_params,
         }
 
@@ -197,6 +200,7 @@ class LLMFilter(FilterOp):
         op_params = {
             "model": self.model,
             "prompt_strategy": self.prompt_strategy,
+            "reasoning_effort": self.reasoning_effort,
             **op_params,
         }
 
@@ -206,7 +210,7 @@ class LLMFilter(FilterOp):
         return self.model.value
 
     def is_image_filter(self) -> bool:
-        return self.model in get_vision_models()
+        return self.prompt_strategy is PromptStrategy.COT_BOOL_IMAGE
 
     def naive_cost_estimates(self, source_op_cost_estimates: OperatorCostEstimates):
         # estimate number of input tokens from source
@@ -225,8 +229,13 @@ class LLMFilter(FilterOp):
         )
 
         # get est. of conversion cost (in USD) per record from model card
+        usd_per_input_token = (
+            MODEL_CARDS[self.model.value]["usd_per_audio_input_token"]
+            if self.prompt_strategy.is_audio_prompt()
+            else MODEL_CARDS[self.model.value]["usd_per_input_token"]
+        )
         model_conversion_usd_per_record = (
-            MODEL_CARDS[self.model.value]["usd_per_input_token"] * est_num_input_tokens
+            usd_per_input_token * est_num_input_tokens
             + MODEL_CARDS[self.model.value]["usd_per_output_token"] * est_num_output_tokens
         )
 
@@ -235,7 +244,7 @@ class LLMFilter(FilterOp):
         cardinality = selectivity * source_op_cost_estimates.cardinality
 
         # estimate quality of output based on the strength of the model being used
-        quality = (MODEL_CARDS[self.model.value]["overall"] / 100.0) * source_op_cost_estimates.quality
+        quality = (MODEL_CARDS[self.model.value]["overall"] / 100.0)
 
         return OperatorCostEstimates(
             cardinality=cardinality,
@@ -251,8 +260,8 @@ class LLMFilter(FilterOp):
         # construct kwargs for generation
         gen_kwargs = {"project_cols": input_fields, "filter_condition": self.filter_obj.filter_condition}
 
-        # generate output; NOTE: BooleanField is used to indicate the output type; thus, the desc is not needed
-        fields = {"passed_operator": BooleanField(desc="")}
+        # generate output; NOTE: FieldInfo is used to indicate the output type; thus, the desc is not needed
+        fields = {"passed_operator": FieldInfo(annotation=bool, description="Whether the record passed the filter operation")}
         field_answers, _, generation_stats, _ = self.generator(candidate, fields, **gen_kwargs)
 
         return field_answers, generation_stats

palimpzest/query/operators/join.py

@@ -0,0 +1,402 @@
+from __future__ import annotations
+
+import time
+from abc import ABC, abstractmethod
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+from pydantic.fields import FieldInfo
+
+from palimpzest.constants import (
+    MODEL_CARDS,
+    NAIVE_EST_JOIN_SELECTIVITY,
+    NAIVE_EST_NUM_INPUT_TOKENS,
+    Cardinality,
+    Model,
+    PromptStrategy,
+)
+from palimpzest.core.elements.records import DataRecord, DataRecordSet
+from palimpzest.core.models import OperatorCostEstimates, RecordOpStats
+from palimpzest.query.generators.generators import Generator
+from palimpzest.query.operators.physical import PhysicalOperator
+
+
+class JoinOp(PhysicalOperator, ABC):
+    def __init__(self, condition: str, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.input_schema == self.output_schema, "Input and output schemas must match for JoinOp"
+        self.condition = condition
+
+    def __str__(self):
+        op = super().__str__()
+        op += f"    Condition: {self.condition}\n"
+        return op
+
+    def get_id_params(self):
+        id_params = super().get_id_params()
+        return {"condition": self.condition, **id_params}
+
+    def get_op_params(self):
+        op_params = super().get_op_params()
+        return {"condition": self.condition, **op_params}
+
+    @abstractmethod
+    def is_image_join(self) -> bool:
+        """Return True if the join operation processes image(s), False otherwise."""
+        pass
+
+    @abstractmethod
+    def naive_cost_estimates(self, left_source_op_cost_estimates: OperatorCostEstimates, right_source_op_cost_estimates: OperatorCostEstimates) -> OperatorCostEstimates:
+        pass
+
+
+class BlockingNestedLoopsJoin(JoinOp):
+    def __init__(
+        self,
+        model: Model,
+        prompt_strategy: PromptStrategy = PromptStrategy.COT_JOIN,
+        join_parallelism: int = 64,
+        reasoning_effort: str | None = None,
+        *args,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        self.model = model
+        self.prompt_strategy = prompt_strategy
+        self.join_parallelism = join_parallelism
+        self.reasoning_effort = reasoning_effort
+        self.generator = Generator(model, prompt_strategy, reasoning_effort, self.api_base, Cardinality.ONE_TO_ONE, self.verbose)
+        self.join_idx = 0
+
+    def get_id_params(self):
+        id_params = super().get_id_params()
+        id_params = {
+            "model": self.model.value,
+            "prompt_strategy": self.prompt_strategy.value,
+            "join_parallelism": self.join_parallelism,
+            "reasoning_effort": self.reasoning_effort,
+            **id_params,
+        }
+
+        return id_params
+
+    def get_op_params(self):
+        op_params = super().get_op_params()
+        op_params = {
+            "model": self.model,
+            "prompt_strategy": self.prompt_strategy,
+            "join_parallelism": self.join_parallelism,
+            "reasoning_effort": self.reasoning_effort,
+            **op_params,
+        }
+
+        return op_params
+
+    def get_model_name(self):
+        return self.model.value
+
+    def is_image_join(self) -> bool:
+        return self.prompt_strategy is PromptStrategy.COT_JOIN_IMAGE
+
+    def naive_cost_estimates(self, left_source_op_cost_estimates: OperatorCostEstimates, right_source_op_cost_estimates: OperatorCostEstimates):
+        # estimate number of input tokens from source
+        est_num_input_tokens = 2 * NAIVE_EST_NUM_INPUT_TOKENS
+        if self.is_image_join():
+            est_num_input_tokens = 2 * 765 / 10  # 1024x1024 image is 765 tokens
+
+        # NOTE: the output often generates an entire reasoning sentence, thus the true value may be higher
+        # the filter operation's LLM call should only output TRUE or FALSE, thus we expect its
+        # number of output tokens to be ~1.25
+        est_num_output_tokens = 1.25
+
+        # get est. of conversion time per record from model card;
+        model_conversion_time_per_record = (
+            MODEL_CARDS[self.model.value]["seconds_per_output_token"] * est_num_output_tokens
+        )
+
+        # get est. of conversion cost (in USD) per record from model card
+        usd_per_input_token = (
+            MODEL_CARDS[self.model.value]["usd_per_audio_input_token"]
+            if self.prompt_strategy.is_audio_prompt()
+            else MODEL_CARDS[self.model.value]["usd_per_input_token"]
+        )
+        model_conversion_usd_per_record = (
+            usd_per_input_token * est_num_input_tokens
+            + MODEL_CARDS[self.model.value]["usd_per_output_token"] * est_num_output_tokens
+        )
+
+        # estimate output cardinality using a constant assumption of the filter selectivity
+        selectivity = NAIVE_EST_JOIN_SELECTIVITY
+        cardinality = selectivity * (left_source_op_cost_estimates.cardinality * right_source_op_cost_estimates.cardinality)
+
+        # estimate quality of output based on the strength of the model being used
+        quality = (MODEL_CARDS[self.model.value]["overall"] / 100.0)
+
+        return OperatorCostEstimates(
+            cardinality=cardinality,
+            time_per_record=model_conversion_time_per_record,
+            cost_per_record=model_conversion_usd_per_record,
+            quality=quality,
+        )
+
+    def _process_join_candidate_pair(
+        self,
+        left_candidate: DataRecord,
+        right_candidate: DataRecord,
+        gen_kwargs: dict,
+    ) -> tuple[list[DataRecord], list[RecordOpStats]]:
+        start_time = time.time()
+
+        # generate output; NOTE: FieldInfo is used to indicate the output type; thus, the desc is not needed
+        fields = {"passed_operator": FieldInfo(annotation=bool, description="Whether the records satisfy the join condition")}
+        field_answers, _, generation_stats, _ = self.generator(left_candidate, fields, right_candidate=right_candidate, **gen_kwargs)
+
+        # determine whether or not the join was satisfied
+        passed_operator = field_answers["passed_operator"]
+
+        # compute output record and add to output_records
+        join_dr = DataRecord.from_join_parents(self.output_schema, left_candidate, right_candidate)
+        join_dr.passed_operator = passed_operator
+
+        # compute record stats and add to output_record_op_stats
+        record_op_stats = RecordOpStats(
+            record_id=join_dr.id,
+            record_parent_ids=join_dr.parent_ids,
+            record_source_indices=join_dr.source_indices,
+            record_state=join_dr.to_dict(include_bytes=False),
+            full_op_id=self.get_full_op_id(),
+            logical_op_id=self.logical_op_id,
+            op_name=self.op_name(),
+            time_per_record=time.time() - start_time,
+            cost_per_record=generation_stats.cost_per_record,
+            model_name=self.get_model_name(),
+            join_condition=self.condition,
+            total_input_tokens=generation_stats.total_input_tokens,
+            total_output_tokens=generation_stats.total_output_tokens,
+            total_input_cost=generation_stats.total_input_cost,
+            total_output_cost=generation_stats.total_output_cost,
+            llm_call_duration_secs=generation_stats.llm_call_duration_secs,
+            fn_call_duration_secs=generation_stats.fn_call_duration_secs,
+            total_llm_calls=generation_stats.total_llm_calls,
+            total_embedding_llm_calls=generation_stats.total_embedding_llm_calls,
+            answer=field_answers,
+            passed_operator=passed_operator,
+            image_operation=self.is_image_join(),
+            op_details={k: str(v) for k, v in self.get_id_params().items()},
+        )
+
+        return [join_dr], [record_op_stats]
+
+    def __call__(self, left_candidates: list[DataRecord], right_candidates: list[DataRecord]) -> tuple[DataRecordSet, int]:
+        # get the set of input fields from both records in the join
+        input_fields = self.get_input_fields()
+
+        # construct kwargs for generation
+        gen_kwargs = {"project_cols": input_fields, "join_condition": self.condition}
+
+        # apply the generator to each pair of candidates
+        output_records, output_record_op_stats, num_inputs_processed = [], [], 0
+        total_join_candidates = len(left_candidates) * len(right_candidates)
+        with ThreadPoolExecutor(max_workers=self.join_parallelism) as executor:
+            futures = []
+            for candidate in left_candidates:
+                for right_candidate in right_candidates:
+                    futures.append(executor.submit(self._process_join_candidate_pair, candidate, right_candidate, gen_kwargs))
+                    num_inputs_processed += 1
+
+            for future in as_completed(futures):
+                self.join_idx += 1
+                join_output_records, join_output_record_op_stats = future.result()
+                output_records.extend(join_output_records)
+                output_record_op_stats.extend(join_output_record_op_stats)
+                print(f"{self.join_idx}/{total_join_candidates} JOINED")
+
+        return DataRecordSet(output_records, output_record_op_stats), num_inputs_processed
+
+
+class NestedLoopsJoin(JoinOp):
+    def __init__(
+        self,
+        model: Model,
+        prompt_strategy: PromptStrategy = PromptStrategy.COT_JOIN,
+        join_parallelism: int = 64,
+        reasoning_effort: str | None = None,
+        *args,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        self.model = model
+        self.prompt_strategy = prompt_strategy
+        self.join_parallelism = join_parallelism
+        self.reasoning_effort = reasoning_effort
+        self.generator = Generator(model, prompt_strategy, reasoning_effort, self.api_base, Cardinality.ONE_TO_ONE, self.verbose)
+        self.join_idx = 0
+
+        # maintain list(s) of input records for the join
+        self._left_input_records: list[DataRecord] = []
+        self._right_input_records: list[DataRecord] = []
+
+    def get_id_params(self):
+        id_params = super().get_id_params()
+        id_params = {
+            "model": self.model.value,
+            "prompt_strategy": self.prompt_strategy.value,
+            "join_parallelism": self.join_parallelism,
+            "reasoning_effort": self.reasoning_effort,
+            **id_params,
+        }
+
+        return id_params
+
+    def get_op_params(self):
+        op_params = super().get_op_params()
+        op_params = {
+            "model": self.model,
+            "prompt_strategy": self.prompt_strategy,
+            "join_parallelism": self.join_parallelism,
+            "reasoning_effort": self.reasoning_effort,
+            **op_params,
+        }
+
+        return op_params
+
+    def get_model_name(self):
+        return self.model.value
+
+    def is_image_join(self) -> bool:
+        return self.prompt_strategy is PromptStrategy.COT_JOIN_IMAGE
+
+    def naive_cost_estimates(self, left_source_op_cost_estimates: OperatorCostEstimates, right_source_op_cost_estimates: OperatorCostEstimates):
+        # estimate number of input tokens from source
+        est_num_input_tokens = 2 * NAIVE_EST_NUM_INPUT_TOKENS
+        if self.is_image_join():
+            est_num_input_tokens = 2 * 765 / 10  # 1024x1024 image is 765 tokens
+
+        # NOTE: the output often generates an entire reasoning sentence, thus the true value may be higher
+        # the filter operation's LLM call should only output TRUE or FALSE, thus we expect its
+        # number of output tokens to be ~1.25
+        est_num_output_tokens = 1.25
+
+        # get est. of conversion time per record from model card;
+        model_conversion_time_per_record = (
+            MODEL_CARDS[self.model.value]["seconds_per_output_token"] * est_num_output_tokens
+        )
+
+        # get est. of conversion cost (in USD) per record from model card
+        usd_per_input_token = (
+            MODEL_CARDS[self.model.value]["usd_per_audio_input_token"]
+            if self.prompt_strategy.is_audio_prompt()
+            else MODEL_CARDS[self.model.value]["usd_per_input_token"]
+        )
+        model_conversion_usd_per_record = (
+            usd_per_input_token * est_num_input_tokens
+            + MODEL_CARDS[self.model.value]["usd_per_output_token"] * est_num_output_tokens
+        )
+
+        # estimate output cardinality using a constant assumption of the filter selectivity
+        selectivity = NAIVE_EST_JOIN_SELECTIVITY
+        cardinality = selectivity * (left_source_op_cost_estimates.cardinality * right_source_op_cost_estimates.cardinality)
+
+        # estimate quality of output based on the strength of the model being used
+        quality = (MODEL_CARDS[self.model.value]["overall"] / 100.0)
+
+        return OperatorCostEstimates(
+            cardinality=cardinality,
+            time_per_record=model_conversion_time_per_record,
+            cost_per_record=model_conversion_usd_per_record,
+            quality=quality,
+        )
+
+    def _process_join_candidate_pair(
+        self,
+        left_candidate: DataRecord,
+        right_candidate: DataRecord,
+        gen_kwargs: dict,
+    ) -> tuple[list[DataRecord], list[RecordOpStats]]:
+        start_time = time.time()
+
+        # generate output; NOTE: FieldInfo is used to indicate the output type; thus, the desc is not needed
+        fields = {"passed_operator": FieldInfo(annotation=bool, description="Whether the records satisfy the join condition")}
+        field_answers, _, generation_stats, _ = self.generator(left_candidate, fields, right_candidate=right_candidate, **gen_kwargs)
+
+        # determine whether or not the join was satisfied
+        passed_operator = field_answers["passed_operator"]
+
+        # compute output record and add to output_records
+        join_dr = DataRecord.from_join_parents(self.output_schema, left_candidate, right_candidate)
+        join_dr.passed_operator = passed_operator
+
+        # compute record stats and add to output_record_op_stats
+        record_op_stats = RecordOpStats(
+            record_id=join_dr.id,
+            record_parent_ids=join_dr.parent_ids,
+            record_source_indices=join_dr.source_indices,
+            record_state=join_dr.to_dict(include_bytes=False),
+            full_op_id=self.get_full_op_id(),
+            logical_op_id=self.logical_op_id,
+            op_name=self.op_name(),
+            time_per_record=time.time() - start_time,
+            cost_per_record=generation_stats.cost_per_record,
+            model_name=self.get_model_name(),
+            join_condition=self.condition,
+            total_input_tokens=generation_stats.total_input_tokens,
+            total_output_tokens=generation_stats.total_output_tokens,
+            total_input_cost=generation_stats.total_input_cost,
+            total_output_cost=generation_stats.total_output_cost,
+            llm_call_duration_secs=generation_stats.llm_call_duration_secs,
+            fn_call_duration_secs=generation_stats.fn_call_duration_secs,
+            total_llm_calls=generation_stats.total_llm_calls,
+            total_embedding_llm_calls=generation_stats.total_embedding_llm_calls,
+            answer=field_answers,
+            passed_operator=passed_operator,
+            image_operation=self.is_image_join(),
+            op_details={k: str(v) for k, v in self.get_id_params().items()},
+        )
+
+        return [join_dr], [record_op_stats]
+
+    def __call__(self, left_candidates: list[DataRecord], right_candidates: list[DataRecord]) -> tuple[DataRecordSet | None, int]:
+        # get the set of input fields from both records in the join
+        input_fields = self.get_input_fields()
+
+        # construct kwargs for generation
+        gen_kwargs = {"project_cols": input_fields, "join_condition": self.condition}
+
+        # apply the generator to each pair of candidates
+        output_records, output_record_op_stats, num_inputs_processed = [], [], 0
+        with ThreadPoolExecutor(max_workers=self.join_parallelism) as executor:
+            futures = []
+            # join new left candidates with new right candidates
+            for candidate in left_candidates:
+                for right_candidate in right_candidates:
+                    futures.append(executor.submit(self._process_join_candidate_pair, candidate, right_candidate, gen_kwargs))
+                    num_inputs_processed += 1
+
+            # join new left candidates with stored right input records
+            for candidate in left_candidates:
+                for right_candidate in self._right_input_records:
+                    futures.append(executor.submit(self._process_join_candidate_pair, candidate, right_candidate, gen_kwargs))
+                    num_inputs_processed += 1
+
+            # join new right candidates with stored left input records
+            for candidate in self._left_input_records:
+                for right_candidate in right_candidates:
+                    futures.append(executor.submit(self._process_join_candidate_pair, candidate, right_candidate, gen_kwargs))
+                    num_inputs_processed += 1
+
+            # collect results as they complete
+            for future in as_completed(futures):
+                self.join_idx += 1
+                join_output_records, join_output_record_op_stats = future.result()
+                output_records.extend(join_output_records)
+                output_record_op_stats.extend(join_output_record_op_stats)
+                print(f"{self.join_idx} JOINED")
+
+        # store input records to join with new records added later
+        self._left_input_records.extend(left_candidates)
+        self._right_input_records.extend(right_candidates)
+
+        # return None if no output records were produced
+        if len(output_records) == 0:
+            return None, num_inputs_processed
+
+        return DataRecordSet(output_records, output_record_op_stats), num_inputs_processed

palimpzest/query/operators/limit.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
-from palimpzest.core.data.dataclasses import OperatorCostEstimates, RecordOpStats
 from palimpzest.core.elements.records import DataRecord, DataRecordSet
+from palimpzest.core.models import OperatorCostEstimates, RecordOpStats
 from palimpzest.query.operators.physical import PhysicalOperator
 
 
@@ -41,8 +41,8 @@ class LimitScanOp(PhysicalOperator):
         # create RecordOpStats object
         record_op_stats = RecordOpStats(
             record_id=dr.id,
-            record_parent_id=dr.parent_id,
-            record_source_idx=dr.source_idx,
+            record_parent_ids=dr.parent_ids,
+            record_source_indices=dr.source_indices,
             record_state=dr.to_dict(include_bytes=False),
             full_op_id=self.get_full_op_id(),
             logical_op_id=self.logical_op_id,