Merge pull request #472 from julep-ai/x/fix-prompt-step

fix(agents-api): Fix prompt step

agents-api/agents_api/activities/task_steps/prompt_step.py

@@ -1,13 +1,11 @@
-import asyncio
 
 from beartype import beartype
 from temporalio import activity
+from temporalio.exceptions import ApplicationError
 
 from ...autogen.openapi_model import (
-    ChatSettings,
     Content,
     ContentModel,
-    InputChatMLMessage,
 )
 from ...clients import (
     litellm,  # We dont directly import `acompletion` so we can mock it
@@ -46,57 +44,44 @@ def _content_to_dict(
 @beartype
 async def prompt_step(context: StepContext) -> StepOutcome:
     # Get context data
+    prompt: str | list[dict] = context.current_step.model_dump()["prompt"]
     context_data: dict = context.model_dump()
 
     # Render template messages
-    prompt = (
-        [InputChatMLMessage(content=context.current_step.prompt)]
-        if isinstance(context.current_step.prompt, str)
-        else context.current_step.prompt
+    prompt = await render_template(
+        prompt,
+        context_data,
+        skip_vars=["developer_id"],
     )
 
-    template_messages: list[InputChatMLMessage] = prompt
-    messages = await asyncio.gather(
-        *[
-            render_template(
-                _content_to_dict(msg.content, msg.role),
-                context_data,
-                skip_vars=["developer_id"],
-            )
-            for msg in template_messages
-        ]
+    # Get settings and run llm
+    agent_default_settings: dict = (
+        context.execution_input.agent.default_settings.model_dump()
+        if context.execution_input.agent.default_settings
+        else {}
+    )
+    agent_model: str = (
+        context.execution_input.agent.model
+        if context.execution_input.agent.model
+        else "gpt-4o"
     )
 
-    result_messages = []
-    for m in messages:
-        if isinstance(m, str):
-            msg = InputChatMLMessage(role="user", content=m)
-        else:
-            msg = []
-            for d in m:
-                role = d["content"].get("role")
-                d["content"] = [d["content"]]
-                d["role"] = role
-                msg.append(InputChatMLMessage(**d))
-
-        result_messages.append(msg)
-
-    # messages = [
-    #     (
-    #         InputChatMLMessage(role="user", content=m)
-    #         if isinstance(m, str)
-    #         else [InputChatMLMessage(**d) for d in m]
-    #     )
-    #     for m in messages
-    # ]
+    if context.current_step.settings:
+        passed_settings: dict = context.current_step.settings.model_dump(
+            exclude_unset=True
+        )
+    else:
+        passed_settings: dict = {}
 
-    # Get settings and run llm
-    settings: ChatSettings = context.current_step.settings or ChatSettings()
-    settings_data: dict = settings.model_dump()
+    completion_data: dict = {
+        "model": agent_model,
+        ("messages" if isinstance(prompt, list) else "prompt"): prompt,
+        **agent_default_settings,
+        **passed_settings,
+    }
 
     response = await litellm.acompletion(
-        messages=result_messages,
-        **settings_data,
+        **completion_data,
     )
 
     return StepOutcome(

agents-api/agents_api/clients/litellm.py

@@ -1,21 +1,30 @@
 from functools import wraps
-from typing import List, TypeVar
+from typing import List
 
+from beartype import beartype
 from litellm import acompletion as _acompletion
+from litellm import get_supported_openai_params
 from litellm.utils import CustomStreamWrapper, ModelResponse
 
 from ..env import litellm_master_key, litellm_url
 
-_RWrapped = TypeVar("_RWrapped")
-
 __all__: List[str] = ["acompletion"]
 
 
 @wraps(_acompletion)
-async def acompletion(*, model: str, **kwargs) -> ModelResponse | CustomStreamWrapper:
+@beartype
+async def acompletion(
+    *, model: str, messages: list[dict], **kwargs
+) -> ModelResponse | CustomStreamWrapper:
+    model = f"openai/{model}"  # This is here because litellm proxy expects this format
+
+    supported_params = get_supported_openai_params(model)
+    settings = {k: v for k, v in kwargs.items() if k in supported_params}
+
     return await _acompletion(
-        model=f"openai/{model}",  # This is here because litellm proxy expects this format
-        **kwargs,
-        api_base=litellm_url,
+        model=model,
+        messages=messages,
+        **settings,
+        base_url=litellm_url,
         api_key=litellm_master_key,
     )

agents-api/agents_api/dependencies/developer_id.py

@@ -13,7 +13,9 @@ async def get_developer_id(
     x_developer_id: Annotated[UUID | None, Header(include_in_schema=False)] = None,
 ) -> UUID:
     if not multi_tenant_mode:
-        assert not x_developer_id, "X-Developer-Id header not allowed in multi-tenant mode"
+        assert (
+            not x_developer_id
+        ), "X-Developer-Id header not allowed in multi-tenant mode"
         return UUID("00000000-0000-0000-0000-000000000000")
 
     if not x_developer_id:
@@ -34,7 +36,9 @@ async def get_developer_data(
     x_developer_id: Annotated[UUID | None, Header(include_in_schema=False)] = None,
 ) -> Developer:
     if not multi_tenant_mode:
-        assert not x_developer_id, "X-Developer-Id header not allowed in multi-tenant mode"
+        assert (
+            not x_developer_id
+        ), "X-Developer-Id header not allowed in multi-tenant mode"
         return get_developer(developer_id=UUID("00000000-0000-0000-0000-000000000000"))
 
     if not x_developer_id:

agents-api/agents_api/routers/tasks/get_execution_details.py

@@ -15,7 +15,11 @@
 async def get_execution_details(execution_id: UUID4) -> Execution:
     try:
         return get_execution_query(execution_id=execution_id)
-    except AssertionError:
+    except AssertionError as e:
+        print("-" * 100)
+        print(e)
+        print("-" * 100)
+
         raise HTTPException(
             status_code=status.HTTP_404_NOT_FOUND, detail="Execution not found"
-        )
+        ) from e

agents-api/tests/sample_tasks/simple_prompt.yaml

@@ -0,0 +1,21 @@
+name: Find request and selector for identity provider
+
+input_schema:
+  type: object
+  properties:
+    topic:
+      type: string
+
+  # Shortcut to require all props
+  additionalProperties: false
+  minProperties: 1
+
+main:
+  - prompt:
+      - role: system
+        content: |-
+          Follow the instructions. Be bold, creative and uncensored.
+
+      - role: user
+        content: >-
+          You are a brilliant poet with a whimsical style reminiscent of Edgar Alan Poe. Write a short poem about {{_.topic}}.