diff --git a/ragna/assistants/_api.py b/ragna/assistants/_api.py
index 21be7557..207533ec 100644
--- a/ragna/assistants/_api.py
+++ b/ragna/assistants/_api.py
@@ -1,5 +1,10 @@
+from __future__ import annotations
+
 import abc
 import os
+from typing import Annotated
+
+import pydantic
 
 import ragna
 from ragna.core import Assistant, EnvVarRequirement, Requirement, Source
@@ -22,7 +27,22 @@ def __init__(self) -> None:
         self._api_key = os.environ[self._API_KEY_ENV_VAR]
 
     async def answer(
-        self, prompt: str, sources: list[Source], *, max_new_tokens: int = 256
+        self,
+        prompt: str,
+        sources: list[Source],
+        *,
+        max_new_tokens: Annotated[
+            int,
+            pydantic.Field(
+                title="Maximum new tokens",
+                description=(
+                    "Maximum number of new tokens to generate. "
+                    "If you experience truncated answers, increase this value. "
+                    "However, be aware that longer answers also incur a higher cost."
+                ),
+                gt=0,
+            ),
+        ] = 256,
     ) -> str:
         return await self._call_api(prompt, sources, max_new_tokens=max_new_tokens)
 
diff --git a/ragna/core/_components.py b/ragna/core/_components.py
index f2196439..9ebbb8d8 100644
--- a/ragna/core/_components.py
+++ b/ragna/core/_components.py
@@ -4,13 +4,13 @@
 import enum
 import functools
 import inspect
-from typing import Type
+from typing import Type, get_type_hints
 
 import pydantic
 import pydantic.utils
 
 from ._document import Document
-from ._utils import RequirementsMixin, merge_models
+from ._utils import RagnaException, RequirementsMixin, merge_models
 
 
 class Component(RequirementsMixin):
@@ -56,17 +56,30 @@ def _protocol_models(
             ).parameters
             extra_param_names = concrete_params.keys() - protocol_params.keys()
 
+            # We can't rely on inspect.Param.annotation, since that might be str. This
+            # happens if the annotation was either explicitly provided as str, or there
+            # is a top-level from __future__ import annotations in the module.
+            annotations = get_type_hints(method, include_extras=True)
+
+            field_definitions = {}
+            for param_name in extra_param_names:
+                annotation = annotations.get(param_name)
+                if annotation is None:
+                    # FIXME: can we live with typing.Any here?
+                    raise RagnaException("Missing annotation")
+
+                default = concrete_params[param_name].default
+                if default is inspect.Parameter.empty:
+                    default = ...
+
+                field_info = pydantic.fields.FieldInfo.from_annotated_attribute(
+                    annotation, default
+                )
+
+                field_definitions[param_name] = (annotation, field_info)
+
             models[(cls, method_name)] = pydantic.create_model(  # type: ignore[call-overload]
-                f"{cls.__name__}.{method_name}",
-                **{
-                    (param := concrete_params[param_name]).name: (
-                        param.annotation,
-                        param.default
-                        if param.default is not inspect.Parameter.empty
-                        else ...,
-                    )
-                    for param_name in extra_param_names
-                },
+                f"{cls.__name__}.{method_name}", **field_definitions
             )
         return models
 
diff --git a/ragna/core/_utils.py b/ragna/core/_utils.py
index 9c20184e..49414166 100644
--- a/ragna/core/_utils.py
+++ b/ragna/core/_utils.py
@@ -13,7 +13,6 @@
 
 import packaging.requirements
 import pydantic
-import pydantic_core
 
 from ragna._compat import importlib_metadata_package_distributions
 
@@ -140,24 +139,28 @@ def merge_models(
     *models: Type[pydantic.BaseModel],
     config: Optional[pydantic.ConfigDict] = None,
 ) -> Type[pydantic.BaseModel]:
-    raw_field_definitions = defaultdict(list)
+    field_infoss = defaultdict(list)
     for model_cls in models:
-        for name, field in model_cls.model_fields.items():
-            type_ = field.annotation
-
-            default: Any
-            if field.is_required():
-                default = ...
-            elif field.default is pydantic_core.PydanticUndefined:
-                default = field.default_factory()  # type: ignore[misc]
-            else:
-                default = field.default
-
-            raw_field_definitions[name].append((type_, default))
+        for name, field_info in model_cls.model_fields.items():
+            field_infoss[name].append(field_info)
 
     field_definitions = {}
-    for name, definitions in raw_field_definitions.items():
-        types, defaults = zip(*definitions)
+    for name, field_infos in field_infoss.items():
+        # data = defaultdict(set)
+        # for field_info in field_infos:
+        #     data["annotation"].add(field_info.annotation)
+        #     data["default"].add(field_info.get_default(call_default_factory=True))
+        #     data["title"].add()
+
+        types, defaults = zip(
+            *(
+                (
+                    field_info.annotation,
+                    field_info.get_default(call_default_factory=True),
+                )
+                for field_info in field_infos
+            )
+        )
 
         types = set(types)
         if len(types) > 1:
@@ -172,7 +175,12 @@ def merge_models(
         else:
             default = None
 
-        field_definitions[name] = (type_, default)
+        # FIXME: We need a way to also check / merge the additional metadata instead of
+        #  just taking the first at face value.
+        field_info = field_infos[0]
+        field_info.default = default
+
+        field_definitions[name] = (type_, field_info)
 
     return cast(
         Type[pydantic.BaseModel],