cocoindex 0.2.3__cp311-abi3-manylinux_2_28_aarch64.whl

cocoindex/flow.py

@@ -0,0 +1,1205 @@
+"""
+Flow is the main interface for building and running flows.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import datetime
+import functools
+import inspect
+import re
+from dataclasses import dataclass
+from enum import Enum
+from threading import Lock
+from typing import (
+    Any,
+    Callable,
+    Generic,
+    Iterable,
+    NamedTuple,
+    Sequence,
+    TypeVar,
+    cast,
+    get_args,
+    get_origin,
+)
+
+from rich.text import Text
+from rich.tree import Tree
+
+from . import _engine  # type: ignore
+from . import index
+from . import op
+from . import setting
+from .convert import (
+    dump_engine_object,
+    make_engine_value_decoder,
+    make_engine_value_encoder,
+)
+from .op import FunctionSpec
+from .runtime import execution_context
+from .setup import SetupChangeBundle
+from .typing import analyze_type_info, encode_enriched_type
+from .validation import (
+    validate_flow_name,
+    validate_full_flow_name,
+    validate_target_name,
+)
+
+
+class _NameBuilder:
+    _existing_names: set[str]
+    _next_name_index: dict[str, int]
+
+    def __init__(self) -> None:
+        self._existing_names = set()
+        self._next_name_index = {}
+
+    def build_name(self, name: str | None, /, prefix: str) -> str:
+        """
+        Build a name. If the name is None, generate a name with the given prefix.
+        """
+        if name is not None:
+            self._existing_names.add(name)
+            return name
+
+        next_idx = self._next_name_index.get(prefix, 0)
+        while True:
+            name = f"{prefix}{next_idx}"
+            next_idx += 1
+            self._next_name_index[prefix] = next_idx
+            if name not in self._existing_names:
+                self._existing_names.add(name)
+                return name
+
+
+_WORD_BOUNDARY_RE = re.compile("(?<!^)(?=[A-Z])")
+
+
+def _to_snake_case(name: str) -> str:
+    return _WORD_BOUNDARY_RE.sub("_", name).lower()
+
+
+def _create_data_slice(
+    flow_builder_state: _FlowBuilderState,
+    creator: Callable[[_engine.DataScopeRef | None, str | None], _engine.DataSlice],
+    name: str | None = None,
+) -> DataSlice[T]:
+    if name is None:
+        return DataSlice(
+            _DataSliceState(
+                flow_builder_state,
+                lambda target: creator(target[0], target[1])
+                if target is not None
+                else creator(None, None),
+            )
+        )
+    else:
+        return DataSlice(_DataSliceState(flow_builder_state, creator(None, name)))
+
+
+def _spec_kind(spec: Any) -> str:
+    return cast(str, spec.__class__.__name__)
+
+
+def _transform_helper(
+    flow_builder_state: _FlowBuilderState,
+    fn_spec: FunctionSpec | Callable[..., Any],
+    transform_args: list[tuple[Any, str | None]],
+    name: str | None = None,
+) -> DataSlice[Any]:
+    if isinstance(fn_spec, FunctionSpec):
+        kind = _spec_kind(fn_spec)
+        spec = fn_spec
+    elif callable(fn_spec) and (
+        op_kind := getattr(fn_spec, "__cocoindex_op_kind__", None)
+    ):
+        kind = op_kind
+        spec = op.EmptyFunctionSpec()
+    else:
+        raise ValueError("transform() can only be called on a CocoIndex function")
+
+    def _create_data_slice_inner(
+        target_scope: _engine.DataScopeRef | None, name: str | None
+    ) -> _engine.DataSlice:
+        result = flow_builder_state.engine_flow_builder.transform(
+            kind,
+            dump_engine_object(spec),
+            transform_args,
+            target_scope,
+            flow_builder_state.field_name_builder.build_name(
+                name, prefix=_to_snake_case(_spec_kind(fn_spec)) + "_"
+            ),
+        )
+        return result
+
+    return _create_data_slice(
+        flow_builder_state,
+        _create_data_slice_inner,
+        name,
+    )
+
+
+T = TypeVar("T")
+S = TypeVar("S")
+
+
+class _DataSliceState:
+    flow_builder_state: _FlowBuilderState
+
+    _lazy_lock: Lock | None = None  # None means it's not lazy.
+    _data_slice: _engine.DataSlice | None = None
+    _data_slice_creator: (
+        Callable[[tuple[_engine.DataScopeRef, str] | None], _engine.DataSlice] | None
+    ) = None
+
+    def __init__(
+        self,
+        flow_builder_state: _FlowBuilderState,
+        data_slice: _engine.DataSlice
+        | Callable[[tuple[_engine.DataScopeRef, str] | None], _engine.DataSlice],
+    ):
+        self.flow_builder_state = flow_builder_state
+
+        if isinstance(data_slice, _engine.DataSlice):
+            self._data_slice = data_slice
+        else:
+            self._lazy_lock = Lock()
+            self._data_slice_creator = data_slice
+
+    @property
+    def engine_data_slice(self) -> _engine.DataSlice:
+        """
+        Get the internal DataSlice.
+        This can be blocking.
+        """
+        if self._lazy_lock is None:
+            if self._data_slice is None:
+                raise ValueError("Data slice is not initialized")
+            return self._data_slice
+        else:
+            if self._data_slice_creator is None:
+                raise ValueError("Data slice creator is not initialized")
+            with self._lazy_lock:
+                if self._data_slice is None:
+                    self._data_slice = self._data_slice_creator(None)
+                return self._data_slice
+
+    async def engine_data_slice_async(self) -> _engine.DataSlice:
+        """
+        Get the internal DataSlice.
+        This can be blocking.
+        """
+        return await asyncio.to_thread(lambda: self.engine_data_slice)
+
+    def attach_to_scope(self, scope: _engine.DataScopeRef, field_name: str) -> None:
+        """
+        Attach the current data slice (if not yet attached) to the given scope.
+        """
+        if self._lazy_lock is not None:
+            with self._lazy_lock:
+                if self._data_slice_creator is None:
+                    raise ValueError("Data slice creator is not initialized")
+                if self._data_slice is None:
+                    self._data_slice = self._data_slice_creator((scope, field_name))
+                    return
+        # TODO: We'll support this by an identity transformer or "aliasing" in the future.
+        raise ValueError("DataSlice is already attached to a field")
+
+
+class DataSlice(Generic[T]):
+    """A data slice represents a slice of data in a flow. It's readonly."""
+
+    _state: _DataSliceState
+
+    def __init__(self, state: _DataSliceState):
+        self._state = state
+
+    def __str__(self) -> str:
+        return str(self._state.engine_data_slice)
+
+    def __repr__(self) -> str:
+        return repr(self._state.engine_data_slice)
+
+    def __getitem__(self, field_name: str) -> DataSlice[T]:
+        field_slice = self._state.engine_data_slice.field(field_name)
+        if field_slice is None:
+            raise KeyError(field_name)
+        return DataSlice(_DataSliceState(self._state.flow_builder_state, field_slice))
+
+    def row(
+        self,
+        /,
+        *,
+        max_inflight_rows: int | None = None,
+        max_inflight_bytes: int | None = None,
+    ) -> DataScope:
+        """
+        Return a scope representing each row of the table.
+        """
+        row_scope = self._state.flow_builder_state.engine_flow_builder.for_each(
+            self._state.engine_data_slice,
+            execution_options=dump_engine_object(
+                _ExecutionOptions(
+                    max_inflight_rows=max_inflight_rows,
+                    max_inflight_bytes=max_inflight_bytes,
+                ),
+            ),
+        )
+        return DataScope(self._state.flow_builder_state, row_scope)
+
+    def for_each(
+        self,
+        f: Callable[[DataScope], None],
+        /,
+        *,
+        max_inflight_rows: int | None = None,
+        max_inflight_bytes: int | None = None,
+    ) -> None:
+        """
+        Apply a function to each row of the collection.
+        """
+        with self.row(
+            max_inflight_rows=max_inflight_rows,
+            max_inflight_bytes=max_inflight_bytes,
+        ) as scope:
+            f(scope)
+
+    def transform(
+        self, fn_spec: op.FunctionSpec | Callable[..., Any], *args: Any, **kwargs: Any
+    ) -> DataSlice[Any]:
+        """
+        Apply a function to the data slice.
+        """
+        transform_args: list[tuple[Any, str | None]] = [
+            (self._state.engine_data_slice, None)
+        ]
+        transform_args += [
+            (self._state.flow_builder_state.get_data_slice(v), None) for v in args
+        ]
+        transform_args += [
+            (self._state.flow_builder_state.get_data_slice(v), k)
+            for k, v in kwargs.items()
+        ]
+
+        return _transform_helper(
+            self._state.flow_builder_state, fn_spec, transform_args
+        )
+
+    def call(self, func: Callable[..., S], *args: Any, **kwargs: Any) -> S:
+        """
+        Call a function with the data slice.
+        """
+        return func(self, *args, **kwargs)
+
+
+def _data_slice_state(data_slice: DataSlice[T]) -> _DataSliceState:
+    return data_slice._state  # pylint: disable=protected-access
+
+
+class DataScope:
+    """
+    A data scope in a flow.
+    It has multple fields and collectors, and allow users to add new fields and collectors.
+    """
+
+    _flow_builder_state: _FlowBuilderState
+    _engine_data_scope: _engine.DataScopeRef
+
+    def __init__(
+        self, flow_builder_state: _FlowBuilderState, data_scope: _engine.DataScopeRef
+    ):
+        self._flow_builder_state = flow_builder_state
+        self._engine_data_scope = data_scope
+
+    def __str__(self) -> str:
+        return str(self._engine_data_scope)
+
+    def __repr__(self) -> str:
+        return repr(self._engine_data_scope)
+
+    def __getitem__(self, field_name: str) -> DataSlice[T]:
+        return DataSlice(
+            _DataSliceState(
+                self._flow_builder_state,
+                self._flow_builder_state.engine_flow_builder.scope_field(
+                    self._engine_data_scope, field_name
+                ),
+            )
+        )
+
+    def __setitem__(self, field_name: str, value: DataSlice[T]) -> None:
+        from .validation import validate_field_name
+
+        validate_field_name(field_name)
+        value._state.attach_to_scope(self._engine_data_scope, field_name)
+
+    def __enter__(self) -> DataScope:
+        return self
+
+    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        del self._engine_data_scope
+
+    def add_collector(self, name: str | None = None) -> DataCollector:
+        """
+        Add a collector to the flow.
+        """
+        return DataCollector(
+            self._flow_builder_state,
+            self._engine_data_scope.add_collector(
+                self._flow_builder_state.field_name_builder.build_name(
+                    name, prefix="_collector_"
+                )
+            ),
+        )
+
+
+class GeneratedField(Enum):
+    """
+    A generated field is automatically set by the engine.
+    """
+
+    UUID = "Uuid"
+
+
+class DataCollector:
+    """A data collector is used to collect data into a collector."""
+
+    _flow_builder_state: _FlowBuilderState
+    _engine_data_collector: _engine.DataCollector
+
+    def __init__(
+        self,
+        flow_builder_state: _FlowBuilderState,
+        data_collector: _engine.DataCollector,
+    ):
+        self._flow_builder_state = flow_builder_state
+        self._engine_data_collector = data_collector
+
+    def collect(self, **kwargs: Any) -> None:
+        """
+        Collect data into the collector.
+        """
+        regular_kwargs = []
+        auto_uuid_field = None
+        for k, v in kwargs.items():
+            if isinstance(v, GeneratedField):
+                if v == GeneratedField.UUID:
+                    if auto_uuid_field is not None:
+                        raise ValueError("Only one generated UUID field is allowed")
+                    auto_uuid_field = k
+                else:
+                    raise ValueError(f"Unexpected generated field: {v}")
+            else:
+                regular_kwargs.append((k, self._flow_builder_state.get_data_slice(v)))
+
+        self._flow_builder_state.engine_flow_builder.collect(
+            self._engine_data_collector, regular_kwargs, auto_uuid_field
+        )
+
+    def export(
+        self,
+        target_name: str,
+        target_spec: op.TargetSpec,
+        /,
+        *,
+        primary_key_fields: Sequence[str],
+        vector_indexes: Sequence[index.VectorIndexDef] = (),
+        vector_index: Sequence[tuple[str, index.VectorSimilarityMetric]] = (),
+        setup_by_user: bool = False,
+    ) -> None:
+        """
+        Export the collected data to the specified target.
+
+        `vector_index` is for backward compatibility only. Please use `vector_indexes` instead.
+        """
+
+        validate_target_name(target_name)
+        if not isinstance(target_spec, op.TargetSpec):
+            raise ValueError(
+                "export() can only be called on a CocoIndex target storage"
+            )
+
+        # For backward compatibility only.
+        if len(vector_indexes) == 0 and len(vector_index) > 0:
+            vector_indexes = [
+                index.VectorIndexDef(field_name=field_name, metric=metric)
+                for field_name, metric in vector_index
+            ]
+
+        index_options = index.IndexOptions(
+            primary_key_fields=primary_key_fields,
+            vector_indexes=vector_indexes,
+        )
+        self._flow_builder_state.engine_flow_builder.export(
+            target_name,
+            _spec_kind(target_spec),
+            dump_engine_object(target_spec),
+            dump_engine_object(index_options),
+            self._engine_data_collector,
+            setup_by_user,
+        )
+
+
+_flow_name_builder = _NameBuilder()
+
+
+class _FlowBuilderState:
+    """
+    A flow builder is used to build a flow.
+    """
+
+    engine_flow_builder: _engine.FlowBuilder
+    field_name_builder: _NameBuilder
+
+    def __init__(self, full_name: str):
+        self.engine_flow_builder = _engine.FlowBuilder(full_name)
+        self.field_name_builder = _NameBuilder()
+
+    def get_data_slice(self, v: Any) -> _engine.DataSlice:
+        """
+        Return a data slice that represents the given value.
+        """
+        if isinstance(v, DataSlice):
+            return v._state.engine_data_slice
+        return self.engine_flow_builder.constant(encode_enriched_type(type(v)), v)
+
+
+@dataclass
+class _SourceRefreshOptions:
+    """
+    Options for refreshing a source.
+    """
+
+    refresh_interval: datetime.timedelta | None = None
+
+
+@dataclass
+class _ExecutionOptions:
+    max_inflight_rows: int | None = None
+    max_inflight_bytes: int | None = None
+
+
+class FlowBuilder:
+    """
+    A flow builder is used to build a flow.
+    """
+
+    _state: _FlowBuilderState
+
+    def __init__(self, state: _FlowBuilderState):
+        self._state = state
+
+    def __str__(self) -> str:
+        return str(self._state.engine_flow_builder)
+
+    def __repr__(self) -> str:
+        return repr(self._state.engine_flow_builder)
+
+    def add_source(
+        self,
+        spec: op.SourceSpec,
+        /,
+        *,
+        name: str | None = None,
+        refresh_interval: datetime.timedelta | None = None,
+        max_inflight_rows: int | None = None,
+        max_inflight_bytes: int | None = None,
+    ) -> DataSlice[T]:
+        """
+        Import a source to the flow.
+        """
+        if not isinstance(spec, op.SourceSpec):
+            raise ValueError("add_source() can only be called on a CocoIndex source")
+        return _create_data_slice(
+            self._state,
+            lambda target_scope, name: self._state.engine_flow_builder.add_source(
+                _spec_kind(spec),
+                dump_engine_object(spec),
+                target_scope,
+                self._state.field_name_builder.build_name(
+                    name, prefix=_to_snake_case(_spec_kind(spec)) + "_"
+                ),
+                refresh_options=dump_engine_object(
+                    _SourceRefreshOptions(refresh_interval=refresh_interval)
+                ),
+                execution_options=dump_engine_object(
+                    _ExecutionOptions(
+                        max_inflight_rows=max_inflight_rows,
+                        max_inflight_bytes=max_inflight_bytes,
+                    )
+                ),
+            ),
+            name,
+        )
+
+    def transform(
+        self, fn_spec: FunctionSpec | Callable[..., Any], *args: Any, **kwargs: Any
+    ) -> DataSlice[Any]:
+        """
+        Apply a function to inputs, returning a DataSlice.
+        """
+        transform_args: list[tuple[Any, str | None]] = [
+            (self._state.get_data_slice(v), None) for v in args
+        ]
+        transform_args += [
+            (self._state.get_data_slice(v), k) for k, v in kwargs.items()
+        ]
+
+        if not transform_args:
+            raise ValueError("At least one input is required for transformation")
+
+        return _transform_helper(self._state, fn_spec, transform_args)
+
+    def declare(self, spec: op.DeclarationSpec) -> None:
+        """
+        Add a declaration to the flow.
+        """
+        self._state.engine_flow_builder.declare(dump_engine_object(spec))
+
+
+@dataclass
+class FlowLiveUpdaterOptions:
+    """
+    Options for live updating a flow.
+    """
+
+    live_mode: bool = True
+    print_stats: bool = False
+
+
+class FlowUpdaterStatusUpdates(NamedTuple):
+    """
+    Status updates for a flow updater.
+    """
+
+    # Sources that are still active, i.e. not stopped processing.
+    active_sources: list[str]
+
+    # Sources with updates since last time.
+    updated_sources: list[str]
+
+
+class FlowLiveUpdater:
+    """
+    A live updater for a flow.
+    """
+
+    _flow: Flow
+    _options: FlowLiveUpdaterOptions
+    _engine_live_updater: _engine.FlowLiveUpdater | None = None
+
+    def __init__(self, fl: Flow, options: FlowLiveUpdaterOptions | None = None):
+        self._flow = fl
+        self._options = options or FlowLiveUpdaterOptions()
+
+    def __enter__(self) -> FlowLiveUpdater:
+        self.start()
+        return self
+
+    def __exit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        self.abort()
+        self.wait()
+
+    async def __aenter__(self) -> FlowLiveUpdater:
+        await self.start_async()
+        return self
+
+    async def __aexit__(self, exc_type: Any, exc_value: Any, traceback: Any) -> None:
+        self.abort()
+        await self.wait_async()
+
+    def start(self) -> None:
+        """
+        Start the live updater.
+        """
+        execution_context.run(self.start_async())
+
+    async def start_async(self) -> None:
+        """
+        Start the live updater.
+        """
+        self._engine_live_updater = await _engine.FlowLiveUpdater.create(
+            await self._flow.internal_flow_async(), dump_engine_object(self._options)
+        )
+
+    def wait(self) -> None:
+        """
+        Wait for the live updater to finish.
+        """
+        execution_context.run(self.wait_async())
+
+    async def wait_async(self) -> None:
+        """
+        Wait for the live updater to finish. Async version.
+        """
+        await self._get_engine_live_updater().wait_async()
+
+    def next_status_updates(self) -> FlowUpdaterStatusUpdates:
+        """
+        Get the next status updates.
+
+        It blocks until there's a new status updates, including the processing finishes for a bunch of source updates,
+        and live updater stops (aborted, or no more sources to process).
+        """
+        return execution_context.run(self.next_status_updates_async())
+
+    async def next_status_updates_async(self) -> FlowUpdaterStatusUpdates:
+        """
+        Get the next status updates. Async version.
+        """
+        updates = await self._get_engine_live_updater().next_status_updates_async()
+        return FlowUpdaterStatusUpdates(
+            active_sources=updates.active_sources,
+            updated_sources=updates.updated_sources,
+        )
+
+    def abort(self) -> None:
+        """
+        Abort the live updater.
+        """
+        self._get_engine_live_updater().abort()
+
+    def update_stats(self) -> _engine.IndexUpdateInfo:
+        """
+        Get the index update info.
+        """
+        return self._get_engine_live_updater().index_update_info()
+
+    def _get_engine_live_updater(self) -> _engine.FlowLiveUpdater:
+        if self._engine_live_updater is None:
+            raise RuntimeError("Live updater is not started")
+        return self._engine_live_updater
+
+
+@dataclass
+class EvaluateAndDumpOptions:
+    """
+    Options for evaluating and dumping a flow.
+    """
+
+    output_dir: str
+    use_cache: bool = True
+
+
+class Flow:
+    """
+    A flow describes an indexing pipeline.
+    """
+
+    _name: str
+    _full_name: str
+    _lazy_engine_flow: Callable[[], _engine.Flow] | None
+
+    def __init__(
+        self, name: str, full_name: str, engine_flow_creator: Callable[[], _engine.Flow]
+    ):
+        validate_flow_name(name)
+        validate_full_flow_name(full_name)
+        self._name = name
+        self._full_name = full_name
+        engine_flow = None
+        lock = Lock()
+
+        def _lazy_engine_flow() -> _engine.Flow:
+            nonlocal engine_flow, lock
+            if engine_flow is None:
+                with lock:
+                    if engine_flow is None:
+                        engine_flow = engine_flow_creator()
+            return engine_flow
+
+        self._lazy_engine_flow = _lazy_engine_flow
+
+    def _render_spec(self, verbose: bool = False) -> Tree:
+        """
+        Render the flow spec as a styled rich Tree with hierarchical structure.
+        """
+        spec = self._get_spec(verbose=verbose)
+        tree = Tree(f"Flow: {self.full_name}", style="cyan")
+
+        def build_tree(label: str, lines: list[Any]) -> Tree:
+            node = Tree(label=label if lines else label + " None", style="cyan")
+            for line in lines:
+                child_node = node.add(Text(line.content, style="yellow"))
+                child_node.children = build_tree("", line.children).children
+            return node
+
+        for section, lines in spec.sections:
+            section_node = build_tree(f"{section}:", lines)
+            tree.children.append(section_node)
+        return tree
+
+    def _get_spec(self, verbose: bool = False) -> _engine.RenderedSpec:
+        return self.internal_flow().get_spec(
+            output_mode="verbose" if verbose else "concise"
+        )
+
+    def _get_schema(self) -> list[tuple[str, str, str]]:
+        return cast(list[tuple[str, str, str]], self.internal_flow().get_schema())
+
+    def __str__(self) -> str:
+        return str(self._get_spec())
+
+    def __repr__(self) -> str:
+        return repr(self.internal_flow())
+
+    @property
+    def name(self) -> str:
+        """
+        Get the name of the flow.
+        """
+        return self._name
+
+    @property
+    def full_name(self) -> str:
+        """
+        Get the full name of the flow.
+        """
+        return self._full_name
+
+    def update(self) -> _engine.IndexUpdateInfo:
+        """
+        Update the index defined by the flow.
+        Once the function returns, the index is fresh up to the moment when the function is called.
+        """
+        return execution_context.run(self.update_async())
+
+    async def update_async(self) -> _engine.IndexUpdateInfo:
+        """
+        Update the index defined by the flow.
+        Once the function returns, the index is fresh up to the moment when the function is called.
+        """
+        async with FlowLiveUpdater(
+            self, FlowLiveUpdaterOptions(live_mode=False)
+        ) as updater:
+            await updater.wait_async()
+        return updater.update_stats()
+
+    def evaluate_and_dump(
+        self, options: EvaluateAndDumpOptions
+    ) -> _engine.IndexUpdateInfo:
+        """
+        Evaluate the flow and dump flow outputs to files.
+        """
+        return self.internal_flow().evaluate_and_dump(dump_engine_object(options))
+
+    def internal_flow(self) -> _engine.Flow:
+        """
+        Get the engine flow.
+        """
+        if self._lazy_engine_flow is None:
+            raise RuntimeError(f"Flow {self.full_name} is already removed")
+        return self._lazy_engine_flow()
+
+    async def internal_flow_async(self) -> _engine.Flow:
+        """
+        Get the engine flow. The async version.
+        """
+        return await asyncio.to_thread(self.internal_flow)
+
+    def setup(self, report_to_stdout: bool = False) -> None:
+        """
+        Setup persistent backends of the flow.
+        """
+        execution_context.run(self.setup_async(report_to_stdout=report_to_stdout))
+
+    async def setup_async(self, report_to_stdout: bool = False) -> None:
+        """
+        Setup persistent backends of the flow. The async version.
+        """
+        bundle = await make_setup_bundle_async([self])
+        await bundle.describe_and_apply_async(report_to_stdout=report_to_stdout)
+
+    def drop(self, report_to_stdout: bool = False) -> None:
+        """
+        Drop persistent backends of the flow.
+
+        The current instance is still valid after it's called.
+        For example, you can still call `setup()` after it, to setup the persistent backends again.
+
+        Call `close()` if you want to remove the flow from the current process.
+        """
+        execution_context.run(self.drop_async(report_to_stdout=report_to_stdout))
+
+    async def drop_async(self, report_to_stdout: bool = False) -> None:
+        """
+        Drop persistent backends of the flow. The async version.
+        """
+        bundle = await make_drop_bundle_async([self])
+        await bundle.describe_and_apply_async(report_to_stdout=report_to_stdout)
+
+    def close(self) -> None:
+        """
+        Close the flow. It will remove the flow from the current process to free up resources.
+        After it's called, methods of the flow should no longer be called.
+
+        This will NOT touch the persistent backends of the flow.
+        """
+        _engine.remove_flow_context(self.full_name)
+        self._lazy_engine_flow = None
+        with _flows_lock:
+            del _flows[self.name]
+
+
+def _create_lazy_flow(
+    name: str | None, fl_def: Callable[[FlowBuilder, DataScope], None]
+) -> Flow:
+    """
+    Create a flow without really building it yet.
+    The flow will be built the first time when it's really needed.
+    """
+    flow_name = _flow_name_builder.build_name(name, prefix="_flow_")
+    flow_full_name = get_flow_full_name(flow_name)
+
+    def _create_engine_flow() -> _engine.Flow:
+        flow_builder_state = _FlowBuilderState(flow_full_name)
+        root_scope = DataScope(
+            flow_builder_state, flow_builder_state.engine_flow_builder.root_scope()
+        )
+        fl_def(FlowBuilder(flow_builder_state), root_scope)
+        return flow_builder_state.engine_flow_builder.build_flow(
+            execution_context.event_loop
+        )
+
+    return Flow(flow_name, flow_full_name, _create_engine_flow)
+
+
+_flows_lock = Lock()
+_flows: dict[str, Flow] = {}
+
+
+def get_flow_full_name(name: str) -> str:
+    """
+    Get the full name of a flow.
+    """
+    return f"{setting.get_app_namespace(trailing_delimiter='.')}{name}"
+
+
+def open_flow(name: str, fl_def: Callable[[FlowBuilder, DataScope], None]) -> Flow:
+    """
+    Open a flow, with the given name and definition.
+    """
+    with _flows_lock:
+        if name in _flows:
+            raise KeyError(f"Flow with name {name} already exists")
+        fl = _flows[name] = _create_lazy_flow(name, fl_def)
+    return fl
+
+
+def add_flow_def(name: str, fl_def: Callable[[FlowBuilder, DataScope], None]) -> Flow:
+    """
+    DEPRECATED: Use `open_flow()` instead.
+    """
+    return open_flow(name, fl_def)
+
+
+def remove_flow(fl: Flow) -> None:
+    """
+    DEPRECATED: Use `Flow.close()` instead.
+    """
+    fl.close()
+
+
+def flow_def(
+    name: str | None = None,
+) -> Callable[[Callable[[FlowBuilder, DataScope], None]], Flow]:
+    """
+    A decorator to wrap the flow definition.
+    """
+    return lambda fl_def: open_flow(name or fl_def.__name__, fl_def)
+
+
+def flow_names() -> list[str]:
+    """
+    Get the names of all flows.
+    """
+    with _flows_lock:
+        return list(_flows.keys())
+
+
+def flows() -> dict[str, Flow]:
+    """
+    Get all flows.
+    """
+    with _flows_lock:
+        return dict(_flows)
+
+
+def flow_by_name(name: str) -> Flow:
+    """
+    Get a flow by name.
+    """
+    with _flows_lock:
+        return _flows[name]
+
+
+def ensure_all_flows_built() -> None:
+    """
+    Ensure all flows are built.
+    """
+    execution_context.run(ensure_all_flows_built_async())
+
+
+async def ensure_all_flows_built_async() -> None:
+    """
+    Ensure all flows are built.
+    """
+    for fl in flows().values():
+        await fl.internal_flow_async()
+
+
+def update_all_flows(
+    options: FlowLiveUpdaterOptions,
+) -> dict[str, _engine.IndexUpdateInfo]:
+    """
+    Update all flows.
+    """
+    return execution_context.run(update_all_flows_async(options))
+
+
+async def update_all_flows_async(
+    options: FlowLiveUpdaterOptions,
+) -> dict[str, _engine.IndexUpdateInfo]:
+    """
+    Update all flows.
+    """
+    await ensure_all_flows_built_async()
+
+    async def _update_flow(name: str, fl: Flow) -> tuple[str, _engine.IndexUpdateInfo]:
+        async with FlowLiveUpdater(fl, options) as updater:
+            await updater.wait_async()
+            return (name, updater.update_stats())
+
+    fls = flows()
+    all_stats = await asyncio.gather(
+        *(_update_flow(name, fl) for (name, fl) in fls.items())
+    )
+    return dict(all_stats)
+
+
+def _get_data_slice_annotation_type(
+    data_slice_type: type[DataSlice[T] | inspect._empty],
+) -> type[T] | None:
+    type_args = get_args(data_slice_type)
+    if data_slice_type is inspect.Parameter.empty or data_slice_type is DataSlice:
+        return None
+    if get_origin(data_slice_type) != DataSlice or len(type_args) != 1:
+        raise ValueError(f"Expect a DataSlice[T] type, but got {data_slice_type}")
+    return cast(type[T] | None, type_args[0])
+
+
+_transform_flow_name_builder = _NameBuilder()
+
+
+class TransformFlowInfo(NamedTuple):
+    engine_flow: _engine.TransientFlow
+    result_decoder: Callable[[Any], T]
+
+
+class FlowArgInfo(NamedTuple):
+    name: str
+    type_hint: Any
+    encoder: Callable[[Any], Any]
+
+
+class TransformFlow(Generic[T]):
+    """
+    A transient transformation flow that transforms in-memory data.
+    """
+
+    _flow_fn: Callable[..., DataSlice[T]]
+    _flow_name: str
+    _args_info: list[FlowArgInfo]
+
+    _lazy_lock: asyncio.Lock
+    _lazy_flow_info: TransformFlowInfo | None = None
+
+    def __init__(
+        self,
+        flow_fn: Callable[..., DataSlice[T]],
+        /,
+        name: str | None = None,
+    ):
+        self._flow_fn = flow_fn
+        self._flow_name = _transform_flow_name_builder.build_name(
+            name, prefix="_transform_flow_"
+        )
+        self._lazy_lock = asyncio.Lock()
+
+        sig = inspect.signature(flow_fn)
+        args_info = []
+        for param_name, param in sig.parameters.items():
+            if param.kind not in (
+                inspect.Parameter.POSITIONAL_OR_KEYWORD,
+                inspect.Parameter.KEYWORD_ONLY,
+            ):
+                raise ValueError(
+                    f"Parameter `{param_name}` is not a parameter can be passed by name"
+                )
+            value_type_annotation: type | None = _get_data_slice_annotation_type(
+                param.annotation
+            )
+            if value_type_annotation is None:
+                raise ValueError(
+                    f"Parameter `{param_name}` for {flow_fn} has no value type annotation. "
+                    "Please use `cocoindex.DataSlice[T]` where T is the type of the value."
+                )
+            encoder = make_engine_value_encoder(
+                analyze_type_info(value_type_annotation)
+            )
+            args_info.append(FlowArgInfo(param_name, value_type_annotation, encoder))
+        self._args_info = args_info
+
+    def __call__(self, *args: Any, **kwargs: Any) -> DataSlice[T]:
+        return self._flow_fn(*args, **kwargs)
+
+    @property
+    def _flow_info(self) -> TransformFlowInfo:
+        if self._lazy_flow_info is not None:
+            return self._lazy_flow_info
+        return execution_context.run(self._flow_info_async())
+
+    async def _flow_info_async(self) -> TransformFlowInfo:
+        if self._lazy_flow_info is not None:
+            return self._lazy_flow_info
+        async with self._lazy_lock:
+            if self._lazy_flow_info is None:
+                self._lazy_flow_info = await self._build_flow_info_async()
+            return self._lazy_flow_info
+
+    async def _build_flow_info_async(self) -> TransformFlowInfo:
+        flow_builder_state = _FlowBuilderState(self._flow_name)
+        kwargs: dict[str, DataSlice[T]] = {}
+        for arg_info in self._args_info:
+            encoded_type = encode_enriched_type(arg_info.type_hint)
+            if encoded_type is None:
+                raise ValueError(f"Parameter `{arg_info.name}` has no type annotation")
+            engine_ds = flow_builder_state.engine_flow_builder.add_direct_input(
+                arg_info.name, encoded_type
+            )
+            kwargs[arg_info.name] = DataSlice(
+                _DataSliceState(flow_builder_state, engine_ds)
+            )
+
+        output = await asyncio.to_thread(lambda: self._flow_fn(**kwargs))
+        output_data_slice = await _data_slice_state(output).engine_data_slice_async()
+
+        flow_builder_state.engine_flow_builder.set_direct_output(output_data_slice)
+        engine_flow = (
+            await flow_builder_state.engine_flow_builder.build_transient_flow_async(
+                execution_context.event_loop
+            )
+        )
+        engine_return_type = output_data_slice.data_type().schema()
+        python_return_type: type[T] | None = _get_data_slice_annotation_type(
+            inspect.signature(self._flow_fn).return_annotation
+        )
+        result_decoder = make_engine_value_decoder(
+            [], engine_return_type["type"], analyze_type_info(python_return_type)
+        )
+
+        return TransformFlowInfo(engine_flow, result_decoder)
+
+    def __str__(self) -> str:
+        return str(self._flow_info.engine_flow)
+
+    def __repr__(self) -> str:
+        return repr(self._flow_info.engine_flow)
+
+    def internal_flow(self) -> _engine.TransientFlow:
+        """
+        Get the internal flow.
+        """
+        return self._flow_info.engine_flow
+
+    def eval(self, *args: Any, **kwargs: Any) -> T:
+        """
+        Evaluate the transform flow.
+        """
+        return execution_context.run(self.eval_async(*args, **kwargs))
+
+    async def eval_async(self, *args: Any, **kwargs: Any) -> T:
+        """
+        Evaluate the transform flow.
+        """
+        flow_info = await self._flow_info_async()
+        params = []
+        for i, arg_info in enumerate(self._args_info):
+            if i < len(args):
+                arg = args[i]
+            elif arg in kwargs:
+                arg = kwargs[arg]
+            else:
+                raise ValueError(f"Parameter {arg} is not provided")
+            params.append(arg_info.encoder(arg))
+        engine_result = await flow_info.engine_flow.evaluate_async(params)
+        return flow_info.result_decoder(engine_result)
+
+
+def transform_flow() -> Callable[[Callable[..., DataSlice[T]]], TransformFlow[T]]:
+    """
+    A decorator to wrap the transform function.
+    """
+
+    def _transform_flow_wrapper(fn: Callable[..., DataSlice[T]]) -> TransformFlow[T]:
+        _transform_flow = TransformFlow(fn)
+        functools.update_wrapper(_transform_flow, fn)
+        return _transform_flow
+
+    return _transform_flow_wrapper
+
+
+async def make_setup_bundle_async(flow_iter: Iterable[Flow]) -> SetupChangeBundle:
+    """
+    Make a bundle to setup flows with the given names.
+    """
+    full_names = []
+    for fl in flow_iter:
+        await fl.internal_flow_async()
+        full_names.append(fl.full_name)
+    return SetupChangeBundle(_engine.make_setup_bundle(full_names))
+
+
+def make_setup_bundle(flow_iter: Iterable[Flow]) -> SetupChangeBundle:
+    """
+    Make a bundle to setup flows with the given names.
+    """
+    return execution_context.run(make_setup_bundle_async(flow_iter))
+
+
+async def make_drop_bundle_async(flow_iter: Iterable[Flow]) -> SetupChangeBundle:
+    """
+    Make a bundle to drop flows with the given names.
+    """
+    full_names = []
+    for fl in flow_iter:
+        await fl.internal_flow_async()
+        full_names.append(fl.full_name)
+    return SetupChangeBundle(_engine.make_drop_bundle(full_names))
+
+
+def make_drop_bundle(flow_iter: Iterable[Flow]) -> SetupChangeBundle:
+    """
+    Make a bundle to drop flows with the given names.
+    """
+    return execution_context.run(make_drop_bundle_async(flow_iter))
+
+
+def setup_all_flows(report_to_stdout: bool = False) -> None:
+    """
+    Setup all flows registered in the current process.
+    """
+    with _flows_lock:
+        flow_list = list(_flows.values())
+    make_setup_bundle(flow_list).describe_and_apply(report_to_stdout=report_to_stdout)
+
+
+def drop_all_flows(report_to_stdout: bool = False) -> None:
+    """
+    Drop all flows registered in the current process.
+    """
+    with _flows_lock:
+        flow_list = list(_flows.values())
+    make_drop_bundle(flow_list).describe_and_apply(report_to_stdout=report_to_stdout)