fix(vllm): addressed minor comments

Author: guicho271828

mellea/backends/utils.py

@@ -2,7 +2,7 @@
 
 import inspect
 from collections.abc import Callable
-from typing import Literal
+from typing import Literal, Any
 
 from mellea.backends.aloras import Alora
 from mellea.backends.formatter import Formatter
@@ -13,6 +13,7 @@
 from mellea.stdlib.requirement import ALoraRequirement, LLMaJRequirement, Requirement
 
 # Chat = dict[Literal["role", "content"], str] # external apply_chat_template type hint is weaker
+# Chat = dict[str, str | list[dict[str, Any]] ] # for multi-modal models
 Chat = dict[str, str]
 
 
@@ -22,6 +23,11 @@ def to_chat(
     formatter: Formatter,
     system_prompt: str | None,
 ) -> list[Chat]:
+    """
+    Converts a context and an action into a series of dicts to be passed to apply_chat_template .
+
+    This function is used by local inference backends.
+    """
     assert ctx.is_chat_context
 
     linearized_ctx = ctx.view_for_generation()
@@ -46,7 +52,7 @@ def to_chat(
 
     # handle custom system prompts. It's important that we do this before the _parse_and_**clean**_model_options step.
     if system_prompt is not None:
-        system_msg: dict[str, str] = {"role": "system", "content": system_prompt}
+        system_msg: Chat = {"role": "system", "content": system_prompt}
         ctx_as_conversation.insert(0, system_msg)
 
     return ctx_as_conversation