qtype 0.1.1__tar.gz → 0.1.3__tar.gz

{qtype-0.1.1/qtype.egg-info → qtype-0.1.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: qtype
-Version: 0.1.1
+Version: 0.1.3
 Summary: DSL for Generative AI Prototyping
 Author-email: Lou Kratz <lou.kratz+qtype@bazaarvoice.com>
 License-Expression: Apache-2.0
@@ -23,16 +23,26 @@ Provides-Extra: interpreter
 Requires-Dist: aiostream>=0.7.1; extra == "interpreter"
 Requires-Dist: arize-phoenix-otel>=0.12.1; extra == "interpreter"
 Requires-Dist: boto3>=1.34.0; extra == "interpreter"
+Requires-Dist: datasets>=4.4.1; extra == "interpreter"
+Requires-Dist: diskcache>=5.6.3; extra == "interpreter"
 Requires-Dist: docling>=2.55.1; extra == "interpreter"
-Requires-Dist: diskcache[interpreter]>=5.6.3; extra == "interpreter"
+Requires-Dist: docx2txt>=0.9; extra == "interpreter"
 Requires-Dist: fastapi>=0.116.1; extra == "interpreter"
+Requires-Dist: jsonpath-ng>=1.7.0; extra == "interpreter"
+Requires-Dist: langfuse>=3.9.0; extra == "interpreter"
 Requires-Dist: llama-index-embeddings-bedrock>=0.5.2; extra == "interpreter"
 Requires-Dist: llama-index-embeddings-openai>=0.3.1; extra == "interpreter"
 Requires-Dist: llama-index-llms-bedrock-converse>=0.10.5; extra == "interpreter"
 Requires-Dist: llama-index-llms-bedrock>=0.3.8; extra == "interpreter"
+Requires-Dist: llama-index-llms-vertex>=0.6.1; extra == "interpreter"
+Requires-Dist: llama-index-postprocessor-bedrock-rerank>=0.5.1; extra == "interpreter"
+Requires-Dist: llama-index-readers-huggingface-fs>=0.4.1; extra == "interpreter"
+Requires-Dist: llama-index-vector-stores-qdrant>=0.8.6; extra == "interpreter"
 Requires-Dist: llama-index>=0.12.45; extra == "interpreter"
 Requires-Dist: openinference-instrumentation-llama-index>=4.3.4; extra == "interpreter"
 Requires-Dist: opensearch-py>=2.7.0; extra == "interpreter"
+Requires-Dist: opentelemetry-exporter-otlp>=1.35.0; extra == "interpreter"
+Requires-Dist: opentelemetry-sdk>=1.35.0; extra == "interpreter"
 Requires-Dist: pandas>=2.2.3; extra == "interpreter"
 Requires-Dist: psycopg2-binary>=2.9.10; extra == "interpreter"
 Requires-Dist: pyarrow>=21.0.0; extra == "interpreter"
@@ -41,15 +51,6 @@ Requires-Dist: python-magic>=0.4.27; extra == "interpreter"
 Requires-Dist: s3fs>=2025.7.0; extra == "interpreter"
 Requires-Dist: sqlalchemy>=2.0.42; extra == "interpreter"
 Requires-Dist: uvicorn[standard]>=0.35.0; extra == "interpreter"
-Requires-Dist: llama-index-llms-vertex>=0.6.1; extra == "interpreter"
-Requires-Dist: langfuse>=3.9.0; extra == "interpreter"
-Requires-Dist: opentelemetry-exporter-otlp>=1.35.0; extra == "interpreter"
-Requires-Dist: opentelemetry-sdk>=1.35.0; extra == "interpreter"
-Requires-Dist: docx2txt>=0.9; extra == "interpreter"
-Requires-Dist: llama-index-vector-stores-qdrant>=0.8.6; extra == "interpreter"
-Requires-Dist: jsonpath-ng>=1.7.0; extra == "interpreter"
-Requires-Dist: llama-index-readers-huggingface-fs>=0.4.1; extra == "interpreter"
-Requires-Dist: datasets>=4.4.1; extra == "interpreter"
 Dynamic: license-file
 
 # QType

{qtype-0.1.1 → qtype-0.1.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "qtype"
-version = "0.1.1"
+version = "0.1.3"
 description = "DSL for Generative AI Prototyping"
 authors = [{ name="Lou Kratz", email="lou.kratz+qtype@bazaarvoice.com" }]
 readme = "README.md"
@@ -29,16 +29,26 @@ interpreter = [
     "aiostream>=0.7.1",
     "arize-phoenix-otel>=0.12.1",
     "boto3>=1.34.0",
+    "datasets>=4.4.1",
+    "diskcache>=5.6.3",
     "docling>=2.55.1",
-    "diskcache[interpreter]>=5.6.3",
+    "docx2txt>=0.9",
     "fastapi>=0.116.1",
+    "jsonpath-ng>=1.7.0",
+    "langfuse>=3.9.0",
     "llama-index-embeddings-bedrock>=0.5.2",
     "llama-index-embeddings-openai>=0.3.1",
     "llama-index-llms-bedrock-converse>=0.10.5",
     "llama-index-llms-bedrock>=0.3.8",
+    "llama-index-llms-vertex>=0.6.1",
+    "llama-index-postprocessor-bedrock-rerank>=0.5.1",
+    "llama-index-readers-huggingface-fs>=0.4.1",
+    "llama-index-vector-stores-qdrant>=0.8.6",
     "llama-index>=0.12.45",
     "openinference-instrumentation-llama-index>=4.3.4",
     "opensearch-py>=2.7.0",
+    "opentelemetry-exporter-otlp>=1.35.0",
+    "opentelemetry-sdk>=1.35.0",
     "pandas>=2.2.3",
     "psycopg2-binary>=2.9.10",
     "pyarrow>=21.0.0",
@@ -47,15 +57,6 @@ interpreter = [
     "s3fs>=2025.7.0",
     "sqlalchemy>=2.0.42",
     "uvicorn[standard]>=0.35.0",
-    "llama-index-llms-vertex>=0.6.1",
-    "langfuse>=3.9.0",
-    "opentelemetry-exporter-otlp>=1.35.0",
-    "opentelemetry-sdk>=1.35.0",
-    "docx2txt>=0.9",
-    "llama-index-vector-stores-qdrant>=0.8.6",
-    "jsonpath-ng>=1.7.0",
-    "llama-index-readers-huggingface-fs>=0.4.1",
-    "datasets>=4.4.1",
 ]
 
 [dependency-groups]

{qtype-0.1.1 → qtype-0.1.3}/qtype/application/facade.py

@@ -27,13 +27,13 @@ class QTypeFacade:
     """
 
     def telemetry(self, spec: SemanticDocumentType) -> None:
+        from qtype.interpreter.telemetry import register
+
         if isinstance(spec, SemanticApplication) and spec.telemetry:
             logger.info(
                 f"Telemetry enabled with endpoint: {spec.telemetry.endpoint}"
             )
             # Register telemetry if needed
-            from qtype.interpreter.telemetry import register
-
             register(spec.telemetry, self.secret_manager(spec), spec.id)
 
     def secret_manager(self, spec: SemanticDocumentType):
@@ -75,11 +75,17 @@ class QTypeFacade:
             DataFrame with results (one row per input)
         """
         import pandas as pd
+        from opentelemetry import trace
 
+        from qtype.interpreter.base.executor_context import ExecutorContext
+        from qtype.interpreter.converters import (
+            dataframe_to_flow_messages,
+            flow_messages_to_dataframe,
+        )
+        from qtype.interpreter.flow import run_flow
+        from qtype.interpreter.types import Session
         from qtype.semantic.loader import load
 
-        logger.info(f"Executing workflow from {path}")
-
         # Load the semantic application
         semantic_model, type_registry = load(Path(path))
         assert isinstance(semantic_model, SemanticApplication)
@@ -100,7 +106,10 @@ class QTypeFacade:
             else:
                 raise ValueError("No flows found in application")
 
+        logger.info(f"Executing flow {target_flow.id} from {path}")
+
         # Convert inputs to DataFrame (normalize single dict to 1-row DataFrame)
+
         if isinstance(inputs, dict):
             input_df = pd.DataFrame([inputs])
         elif isinstance(inputs, pd.DataFrame):
@@ -111,12 +120,6 @@ class QTypeFacade:
             )
 
         # Create session
-        from qtype.interpreter.converters import (
-            dataframe_to_flow_messages,
-            flow_messages_to_dataframe,
-        )
-        from qtype.interpreter.types import Session
-
         session = Session(
             session_id=kwargs.pop("session_id", "default"),
             conversation_history=kwargs.pop("conversation_history", []),
@@ -126,12 +129,8 @@ class QTypeFacade:
         initial_messages = dataframe_to_flow_messages(input_df, session)
 
         # Execute the flow
-        from opentelemetry import trace
-
-        from qtype.interpreter.base.executor_context import ExecutorContext
-        from qtype.interpreter.flow import run_flow
-
         secret_manager = self.secret_manager(semantic_model)
+
         context = ExecutorContext(
             secret_manager=secret_manager,
             tracer=trace.get_tracer(__name__),

{qtype-0.1.1 → qtype-0.1.3}/qtype/cli.py

@@ -135,7 +135,7 @@ def main() -> None:
     # Set logging level based on user input
     logging.basicConfig(
         level=getattr(logging, args.log_level),
-        format="%(levelname)s: %(message)s",
+        format="%(asctime)s - %(levelname)s: %(message)s",
     )
 
     # Dispatch to the selected subcommand

{qtype-0.1.1 → qtype-0.1.3}/qtype/commands/generate.py

@@ -188,7 +188,7 @@ def parser(subparsers: argparse._SubParsersAction) -> None:
 
         has_semantic_deps = True
     except ImportError:
-        logger.warning(
+        logger.debug(
             "NetworkX or Ruff is not installed. Skipping semantic model generation."
         )
         has_semantic_deps = False

{qtype-0.1.1 → qtype-0.1.3}/qtype/commands/run.py

@@ -25,7 +25,7 @@ warnings.filterwarnings("ignore", category=UnsupportedFieldAttributeWarning)
 
 
 # supress qdrant logging
-for name in ["httpx", "urllib3", "qdrant_client"]:
+for name in ["httpx", "urllib3", "qdrant_client", "opensearch"]:
     logging.getLogger(name).setLevel(logging.WARNING)
 
 
@@ -40,12 +40,16 @@ def read_data_from_file(file_path: str) -> pd.DataFrame:
     mime_type = magic.Magic(mime=True).from_file(file_path)
 
     if mime_type == "text/csv":
-        return pd.read_csv(file_path)
+        # TODO: Restore na values and convert to optional once we support them https://github.com/bazaarvoice/qtype/issues/101
+        df = pd.read_csv(file_path)
+        return df.fillna("")
     elif mime_type == "text/plain":
         # For text/plain, use file extension to determine format
         file_ext = Path(file_path).suffix.lower()
         if file_ext == ".csv":
-            return pd.read_csv(file_path)
+            # TODO: Restore na values and convert to optional once we support them https://github.com/bazaarvoice/qtype/issues/101
+            df = pd.read_csv(file_path)
+            return df.fillna("")
         elif file_ext == ".json":
             return pd.read_json(file_path)
         else:

{qtype-0.1.1 → qtype-0.1.3}/qtype/dsl/domain_types.py

@@ -93,12 +93,33 @@ class RAGChunk(Embedding):
     )
 
 
-class RAGSearchResult(StrictBaseModel):
-    """A standard, built-in representation of a search result from a RAG vector search."""
+class SearchResult(StrictBaseModel):
+    """A standard, built-in representation of a search result."""
 
-    chunk: RAGChunk = Field(
+    content: Any = Field(..., description="The content of the search result.")
+    doc_id: str = Field(
+        ...,
+        description="The identifier of the document from which the result was retrieved.",
+    )
+    score: float = Field(
+        ...,
+        description="The relevance score of the search result with respect to the query.",
+    )
+
+
+class RAGSearchResult(SearchResult):
+    """A standard, built-in representation of a search result from a RAG vector search.
+
+    Note: doc_id is duplicated from content.document_id for convenience.
+    """
+
+    content: RAGChunk = Field(
         ..., description="The RAG chunk returned as a search result."
     )
+    doc_id: str = Field(
+        ...,
+        description="The document ID (duplicated from content.document_id).",
+    )
     score: float = Field(
         ...,
         description="The similarity score of the chunk with respect to the query.",

{qtype-0.1.1 → qtype-0.1.3}/qtype/dsl/model.py

@@ -573,6 +573,10 @@ class FieldExtractor(Step):
         ...,
         description="JSONPath expression to extract data from the input. Uses jsonpath-ng syntax.",
     )
+    fail_on_missing: bool = Field(
+        default=True,
+        description="Whether to raise an error if the JSONPath matches no data. If False, returns None.",
+    )
 
 
 class InvokeTool(Step, ConcurrentStepMixin):
@@ -1077,6 +1081,14 @@ class DocumentIndex(Index):
         ...,
         description="URL endpoint for the search cluster (e.g., https://my-cluster.es.amazonaws.com).",
     )
+    id_field: str | None = Field(
+        default=None,
+        description=(
+            "Field name to use as document ID. "
+            "If not specified, auto-detects from: _id, id, doc_id, document_id, or uuid. "
+            "If all are missing, a UUID is generated."
+        ),
+    )
 
 
 class Search(Step, ABC):
@@ -1089,15 +1101,18 @@ class Search(Step, ABC):
     index: Reference[IndexType] | str = Field(
         ..., description="Index to search against (object or ID reference)."
     )
+    default_top_k: int | None = Field(
+        default=10,
+        description="Number of top results to retrieve if not provided in the inputs.",
+    )
 
 
 class VectorSearch(Search, BatchableStepMixin):
     """Performs vector similarity search against a vector index."""
 
     type: Literal["VectorSearch"] = "VectorSearch"
-    default_top_k: int | None = Field(
-        default=50,
-        description="Number of top results to retrieve if not provided in the inputs.",
+    index: Reference[VectorIndex] | str = Field(
+        ..., description="Index to search against (object or ID reference)."
     )
 
 
@@ -1105,6 +1120,43 @@ class DocumentSearch(Search, ConcurrentStepMixin):
     """Performs document search against a document index."""
 
     type: Literal["DocumentSearch"] = "DocumentSearch"
+    index: Reference[DocumentIndex] | str = Field(
+        ..., description="Index to search against (object or ID reference)."
+    )
+    query_args: dict[str, Any] = Field(
+        default={
+            "type": "best_fields",
+            "fields": ["*"],
+        },
+        description="The arguments (other than 'query') to specify to the query shape (see https://docs.opensearch.org/latest/query-dsl/full-text/multi-match/).",
+    )
+
+
+class Reranker(Step):
+    """Reranks a list of documents based on relevance to a query using an LLM."""
+
+    type: Literal["Reranker"] = "Reranker"
+
+
+# TODO: create a reranker that supports llamaindex rerankers...
+
+
+class BedrockReranker(Reranker, ConcurrentStepMixin):
+    """Reranks documents using an AWS Bedrock model."""
+
+    type: Literal["BedrockReranker"] = "BedrockReranker"
+    auth: Reference[AWSAuthProvider] | str | None = Field(
+        default=None,
+        description="AWS authorization provider for Bedrock access.",
+    )
+    model_id: str = Field(
+        ...,
+        description="Bedrock model ID to use for reranking. See https://docs.aws.amazon.com/bedrock/latest/userguide/rerank-supported.html",
+    )
+    num_results: int | None = Field(
+        default=None,
+        description="Return this many results.",
+    )
 
 
 # Create a union type for all tool types
@@ -1146,6 +1198,7 @@ StepType = Annotated[
     Union[
         Agent,
         Aggregate,
+        BedrockReranker,
         Decoder,
         DocToTextConverter,
         DocumentEmbedder,

{qtype-0.1.1 → qtype-0.1.3}/qtype/interpreter/base/base_step_executor.py

@@ -241,7 +241,7 @@ class StepExecutor(ABC):
 
                 # Track message counts for telemetry
                 message_count = 0
-                error_count = len(failed_messages)
+                error_count = 0
 
                 # Stream results and track progress
                 async with all_results.stream() as streamer:

{qtype-0.1.1 → qtype-0.1.3}/qtype/interpreter/base/executor_context.py

@@ -7,7 +7,8 @@ concerns threaded through the execution pipeline.
 
 from __future__ import annotations
 
-from dataclasses import dataclass
+from concurrent.futures import ThreadPoolExecutor
+from dataclasses import dataclass, field
 
 from opentelemetry.trace import Tracer
 
@@ -51,6 +52,9 @@ class ExecutorContext:
         on_progress: Optional callback for progress updates during execution.
         tracer: OpenTelemetry tracer for distributed tracing and observability.
             Defaults to a no-op tracer if telemetry is not configured.
+        thread_pool: Shared thread pool for running synchronous operations
+            in async contexts. Defaults to a pool with 100 threads to support
+            high concurrency workloads without thread exhaustion.
 
     Example:
         ```python
@@ -72,3 +76,16 @@ class ExecutorContext:
     on_stream_event: StreamingCallback | None = None
     on_progress: ProgressCallback | None = None
     tracer: Tracer | None = None
+    thread_pool: ThreadPoolExecutor = field(
+        default_factory=lambda: ThreadPoolExecutor(max_workers=100)
+    )
+
+    def cleanup(self) -> None:
+        """
+        Clean up resources held by the context.
+
+        This should be called when the context is no longer needed to ensure
+        proper cleanup of the thread pool and any other resources.
+        """
+        if self.thread_pool:
+            self.thread_pool.shutdown(wait=True)

qtype-0.1.3/qtype/interpreter/base/factory.py

@@ -0,0 +1,84 @@
+from qtype.semantic.model import (
+    Agent,
+    Aggregate,
+    BedrockReranker,
+    Decoder,
+    DocToTextConverter,
+    DocumentEmbedder,
+    DocumentSearch,
+    DocumentSource,
+    DocumentSplitter,
+    Echo,
+    FieldExtractor,
+    FileSource,
+    FileWriter,
+    IndexUpsert,
+    InvokeEmbedding,
+    InvokeFlow,
+    InvokeTool,
+    LLMInference,
+    PromptTemplate,
+    SQLSource,
+    Step,
+    VectorSearch,
+)
+
+from .batch_step_executor import StepExecutor
+from .executor_context import ExecutorContext
+
+# Lazy-load executor classes only when needed
+# This avoids importing heavy dependencies until actually required
+EXECUTOR_REGISTRY = {
+    Agent: "qtype.interpreter.executors.agent_executor.AgentExecutor",
+    Aggregate: "qtype.interpreter.executors.aggregate_executor.AggregateExecutor",
+    BedrockReranker: "qtype.interpreter.executors.bedrock_reranker_executor.BedrockRerankerExecutor",
+    Decoder: "qtype.interpreter.executors.decoder_executor.DecoderExecutor",
+    DocToTextConverter: "qtype.interpreter.executors.doc_to_text_executor.DocToTextConverterExecutor",
+    DocumentEmbedder: "qtype.interpreter.executors.document_embedder_executor.DocumentEmbedderExecutor",
+    DocumentSearch: "qtype.interpreter.executors.document_search_executor.DocumentSearchExecutor",
+    DocumentSource: "qtype.interpreter.executors.document_source_executor.DocumentSourceExecutor",
+    DocumentSplitter: "qtype.interpreter.executors.document_splitter_executor.DocumentSplitterExecutor",
+    Echo: "qtype.interpreter.executors.echo_executor.EchoExecutor",
+    FieldExtractor: "qtype.interpreter.executors.field_extractor_executor.FieldExtractorExecutor",
+    FileSource: "qtype.interpreter.executors.file_source_executor.FileSourceExecutor",
+    FileWriter: "qtype.interpreter.executors.file_writer_executor.FileWriterExecutor",
+    IndexUpsert: "qtype.interpreter.executors.index_upsert_executor.IndexUpsertExecutor",
+    InvokeEmbedding: "qtype.interpreter.executors.invoke_embedding_executor.InvokeEmbeddingExecutor",
+    InvokeFlow: "qtype.interpreter.executors.invoke_flow_executor.InvokeFlowExecutor",
+    InvokeTool: "qtype.interpreter.executors.invoke_tool_executor.InvokeToolExecutor",
+    LLMInference: "qtype.interpreter.executors.llm_inference_executor.LLMInferenceExecutor",
+    PromptTemplate: "qtype.interpreter.executors.prompt_template_executor.PromptTemplateExecutor",
+    SQLSource: "qtype.interpreter.executors.sql_source_executor.SQLSourceExecutor",
+    VectorSearch: "qtype.interpreter.executors.vector_search_executor.VectorSearchExecutor",
+}
+
+
+def create_executor(
+    step: Step, context: ExecutorContext, **dependencies
+) -> StepExecutor:
+    """
+    Factory to create the appropriate executor for a given step.
+
+    Args:
+        step: The step to create an executor for
+        context: ExecutorContext containing cross-cutting concerns
+        **dependencies: Executor-specific dependencies
+
+    Returns:
+        StepExecutor: Configured executor instance
+    """
+    executor_path = EXECUTOR_REGISTRY.get(type(step))
+    if not executor_path:
+        raise ValueError(
+            f"No executor found for step type: {type(step).__name__}"
+        )
+
+    # Lazy-load the executor class
+    module_path, class_name = executor_path.rsplit(".", 1)
+    import importlib
+
+    module = importlib.import_module(module_path)
+    executor_class = getattr(module, class_name)
+
+    # This assumes the constructor takes the step, context, then dependencies
+    return executor_class(step, context, **dependencies)

{qtype-0.1.1 → qtype-0.1.3}/qtype/interpreter/conversions.py

@@ -18,10 +18,15 @@ from llama_index.core.base.llms.types import (
 from llama_index.core.memory import Memory as LlamaMemory
 from llama_index.core.schema import Document as LlamaDocument
 from llama_index.core.vector_stores.types import BasePydanticVectorStore
-from opensearchpy import AWSV4SignerAuth, OpenSearch
+from opensearchpy import AsyncOpenSearch, AWSV4SignerAuth
 
 from qtype.base.types import PrimitiveTypeEnum
-from qtype.dsl.domain_types import ChatContent, ChatMessage, RAGDocument
+from qtype.dsl.domain_types import (
+    ChatContent,
+    ChatMessage,
+    RAGDocument,
+    RAGSearchResult,
+)
 from qtype.dsl.model import Memory
 from qtype.interpreter.auth.aws import aws
 from qtype.interpreter.auth.generic import auth
@@ -328,7 +333,7 @@ def to_embedding_model(model: Model) -> BaseEmbedding:
 @cached_resource
 def to_opensearch_client(
     index: DocumentIndex, secret_manager: SecretManagerBase
-) -> OpenSearch:
+) -> AsyncOpenSearch:
     """
     Convert a DocumentIndex to an OpenSearch/Elasticsearch client.
 
@@ -377,7 +382,7 @@ def to_opensearch_client(
                 f"Unsupported authentication type for DocumentIndex: {type(index.auth)}"
             )
 
-    return OpenSearch(**client_kwargs)
+    return AsyncOpenSearch(**client_kwargs)
 
 
 def to_content_block(content: ChatContent) -> ContentBlock:
@@ -575,7 +580,7 @@ def to_llama_vector_store_and_retriever(
     return vector_store, retriever
 
 
-def from_node_with_score(node_with_score) -> Any:
+def from_node_with_score(node_with_score) -> RAGSearchResult:
     """Convert a LlamaIndex NodeWithScore to a RAGSearchResult.
 
     Args:
@@ -603,4 +608,8 @@ def from_node_with_score(node_with_score) -> Any:
     )
 
     # Wrap in RAGSearchResult with score
-    return RAGSearchResult(chunk=chunk, score=node_with_score.score or 0.0)
+    return RAGSearchResult(
+        content=chunk,
+        doc_id=chunk.document_id,
+        score=node_with_score.score or 0.0,
+    )

{qtype-0.1.1 → qtype-0.1.3}/qtype/interpreter/converters.py

@@ -2,6 +2,9 @@
 
 from __future__ import annotations
 
+from collections.abc import AsyncIterator
+from typing import Any, cast
+
 import pandas as pd
 from pydantic import BaseModel
 
@@ -9,11 +12,11 @@ from qtype.interpreter.types import FlowMessage, Session
 from qtype.semantic.model import Flow
 
 
-def dataframe_to_flow_messages(
+async def dataframe_to_flow_messages(
     df: pd.DataFrame, session: Session
-) -> list[FlowMessage]:
+) -> AsyncIterator[FlowMessage]:
     """
-    Convert a DataFrame to a list of FlowMessages.
+    Convert a DataFrame to an async generator of FlowMessages.
 
     Each row in the DataFrame becomes a FlowMessage with the same session.
 
@@ -21,14 +24,15 @@ def dataframe_to_flow_messages(
         df: DataFrame where each row represents one set of inputs
         session: Session object to use for all messages
 
-    Returns:
-        List of FlowMessages, one per DataFrame row
+    Yields:
+        FlowMessages, one per DataFrame row
     """
-    messages = []
-    for _, row in df.iterrows():
-        variables = row.to_dict()
-        messages.append(FlowMessage(session=session, variables=variables))
-    return messages
+    # Use to_dict with orient='records' - much faster than iterrows
+    # This returns a list of dicts directly without Series overhead
+    records = cast(list[dict[str, Any]], df.to_dict(orient="records"))
+
+    for record in records:
+        yield FlowMessage(session=session, variables=record)
 
 
 def flow_messages_to_dataframe(
@@ -46,8 +50,6 @@ def flow_messages_to_dataframe(
     Returns:
         DataFrame with one row per message, columns for each output variable
     """
-    from typing import Any
-
     results = []
     for idx, message in enumerate(messages):
         row_data: dict[str, Any] = {"row": idx}