perf: avoid all heavy operations for cocoindex ls (#415)

* fix(cli): `ls` only needs flow names, flows don't need to be built

* perf(ops): lazily import `sentence_transformer` as it's slow

Author: badmonster0

python/cocoindex/cli.py

@@ -1,4 +1,3 @@
-import asyncio
 import click
 import datetime
 from rich.console import Console
@@ -21,7 +20,7 @@ def ls(show_all: bool):
     """
     List all flows.
     """
-    current_flow_names = [fl.name for fl in flow.flows()]
+    current_flow_names = flow.flow_names()
     persisted_flow_names = flow_names_with_setup()
     remaining_persisted_flow_names = set(persisted_flow_names)
 

python/cocoindex/functions.py

@@ -1,10 +1,13 @@
 """All builtin functions."""
-from typing import Annotated, Any
+from typing import Annotated, Any, TYPE_CHECKING
 
-import sentence_transformers
 from .typing import Float32, Vector, TypeAttr
 from . import op, llm
 
+# Libraries that are heavy to import. Lazily import them later.
+if TYPE_CHECKING:
+    import sentence_transformers
+
 class ParseJson(op.FunctionSpec):
     """Parse a text into a JSON object."""
 
@@ -35,9 +38,10 @@ class SentenceTransformerEmbedExecutor:
     """Executor for SentenceTransformerEmbed."""
 
     spec: SentenceTransformerEmbed
-    _model: sentence_transformers.SentenceTransformer
+    _model: "sentence_transformers.SentenceTransformer"
 
     def analyze(self, text):
+        import sentence_transformers # pylint: disable=import-outside-toplevel
         args = self.spec.args or {}
         self._model = sentence_transformers.SentenceTransformer(self.spec.model, **args)
         dim = self._model.get_sentence_embedding_dimension()

python/cocoindex/op.py

@@ -5,10 +5,10 @@
 import dataclasses
 import inspect
 
-from typing import get_type_hints, Protocol, Any, Callable, Awaitable, dataclass_transform
+from typing import Protocol, Any, Callable, Awaitable, dataclass_transform
 from enum import Enum
 
-from .typing import encode_enriched_type
+from .typing import encode_enriched_type, resolve_forward_ref
 from .convert import encode_engine_value, make_engine_value_decoder
 from . import _engine
 
@@ -214,10 +214,11 @@ def _inner(cls: type[Executor]) -> type:
         """
         Decorate a class to provide an executor for an op.
         """
-        type_hints = get_type_hints(cls)
+        # Use `__annotations__` instead of `get_type_hints`, to avoid resolving forward references.
+        type_hints = cls.__annotations__
         if 'spec' not in type_hints:
             raise TypeError("Expect a `spec` field with type hint")
-        spec_cls = type_hints['spec']
+        spec_cls = resolve_forward_ref(type_hints['spec'])
         sig = inspect.signature(cls.__call__)
         return _register_op_factory(
             category=spec_cls._op_category,

python/cocoindex/typing.py

@@ -245,3 +245,8 @@ def encode_enriched_type(t) -> dict[str, Any] | None:
         return None
 
     return encode_enriched_type_info(analyze_type_info(t))
+
+def resolve_forward_ref(t):
+    if t is str:
+        return eval(t) # pylint: disable=eval-used
+    return t