fix: Groq Content Parts, Fixups Refactor, and Gemini Caching (#103)

* Refactored fixups. Corrected behaviors for Groq content parts not being pure text.

* make compat flags a set

Author: monoxgas

rigging/generator/base.py

@@ -1,7 +1,7 @@
 import abc
+import functools
 import inspect
 import typing as t
-from dataclasses import dataclass, field
 from functools import lru_cache
 
 from loguru import logger
@@ -42,7 +42,7 @@ class Fixup(abc.ABC):
     """
 
     @abc.abstractmethod
-    def can_fix(self, exception: Exception) -> bool:
+    def can_fix(self, exception: Exception) -> bool | t.Literal["once"]:
         """
         Check if the fixup can resolve the given exception if made active.
 
@@ -68,10 +68,62 @@ def fix(self, messages: t.Sequence[Message]) -> t.Sequence[Message]:
         ...
 
 
-@dataclass
-class Fixups:
-    available: list[Fixup] = field(default_factory=list)
-    active: list[Fixup] = field(default_factory=list)
+FixupCompatibleFunc = t.Callable[
+    t.Concatenate[t.Any, t.Sequence[Message], P],
+    t.Awaitable[R],
+]
+
+
+def with_fixups(
+    *fixups: Fixup,
+) -> t.Callable[[FixupCompatibleFunc[P, R]], FixupCompatibleFunc[P, R]]:
+    """
+    Decorator that adds fixup retry logic with persistent state.
+
+    Args:
+        fixups: Sequence of fixups to try
+    """
+    available_fixups: list[Fixup] = list(fixups)
+    active_fixups: list[Fixup] = []
+    once_fixups: list[Fixup] = []
+
+    def decorator(func: FixupCompatibleFunc[P, R]) -> FixupCompatibleFunc[P, R]:
+        @functools.wraps(func)
+        async def wrapper(
+            self: t.Any,
+            messages: t.Sequence[Message],
+            *args: P.args,
+            **kwargs: P.kwargs,
+        ) -> R:
+            nonlocal available_fixups, active_fixups
+
+            for fixup in [*active_fixups, *once_fixups]:
+                messages = fixup.fix(messages)
+
+            try:
+                result = await func(self, messages, *args, **kwargs)
+                available_fixups = [*available_fixups, *once_fixups]
+                once_fixups.clear()
+            except Exception as e:
+                for fixup in list(available_fixups):
+                    if (can_fix := fixup.can_fix(e)) is False:
+                        continue
+
+                    if can_fix == "once":
+                        once_fixups.append(fixup)
+                    else:
+                        active_fixups.append(fixup)
+                    available_fixups.remove(fixup)
+
+                    return await wrapper(self, messages, *args, **kwargs)
+
+                raise
+
+            return result
+
+        return wrapper  # type: ignore[return-value]
+
+    return decorator
 
 
 # TODO: We also would like to support N-style
@@ -305,8 +357,6 @@ class Generator(BaseModel):
     _watch_callbacks: list["WatchChatCallback | WatchCompletionCallback"] = []
     _wrap: t.Callable[[CallableT], CallableT] | None = None
 
-    _fixups: Fixups = Fixups()
-
     def to_identifier(self, params: GenerateParams | None = None) -> str:
         """
         Converts the generator instance back into a rigging identifier string.
@@ -393,38 +443,6 @@ async def supports_function_calling(self) -> bool | None:
         """
         return None
 
-    def _check_fixups(self, error: Exception) -> bool:
-        """
-        Check if any fixer can handle this error.
-
-        Args:
-            error: The error to be checked.
-
-        Returns:
-            Whether a fixer was able to handle the error.
-        """
-        for fixup in self._fixups.available[:]:
-            if fixup.can_fix(error):
-                self._fixups.active.append(fixup)
-                self._fixups.available.remove(fixup)
-                return True
-        return False
-
-    async def _apply_fixups(self, messages: t.Sequence[Message]) -> t.Sequence[Message]:
-        """
-        Apply all active fixups to the messages.
-
-        Args:
-            messages: The messages to be fixed.
-
-        Returns:
-            The fixed messages.
-        """
-        current_messages = messages
-        for fixup in self._fixups.active:
-            current_messages = fixup.fix(current_messages)
-        return current_messages
-
     async def generate_messages(
         self,
         messages: t.Sequence[t.Sequence[Message]],

rigging/generator/litellm_.py

@@ -10,13 +10,13 @@
 
 from rigging.generator.base import (
     Fixup,
-    Fixups,
     GeneratedMessage,
     GeneratedText,
     GenerateParams,
     Generator,
     trace_messages,
     trace_str,
+    with_fixups,
 )
 from rigging.message import ContentAudioInput, ContentImageUrl, ContentText, Message
 from rigging.tool.api import ApiFunctionDefinition, ApiToolDefinition
@@ -74,11 +74,40 @@ class CacheTooSmallFixup(Fixup):
     # are below a certain threshold can result in a 400
     # error from APIs (Vertex/Gemini).
 
+    def can_fix(self, exception: Exception) -> bool | t.Literal["once"]:
+        return "once" if "Cached content is too small." in str(exception) else False
+
+    def fix(self, messages: t.Sequence[Message]) -> t.Sequence[Message]:
+        return [message.cache(False) for message in messages]
+
+
+class GroqAssistantContentFixup(Fixup):
+    # Groq can complain if we try to send fully
+    # structured content parts when working with
+    # the assistant role.
+    #
+    # Compatibility flags are a poor workaround for the
+    # fact that we don't have direct control over the
+    # conversion to the OpenAI spec.
+
     def can_fix(self, exception: Exception) -> bool:
-        return "Cached content is too small." in str(exception)
+        return "Groq" in str(exception) and "content' : value must be a string" in str(exception)
+
+    def fix(self, messages: t.Sequence[Message]) -> t.Sequence[Message]:
+        updated_messages: list[Message] = []
+        for message in messages:
+            if message.role == "assistant":
+                message = message.clone()  # noqa: PLW2901
+                message._compability_flags.add("content_as_str")  # noqa: SLF001
+            updated_messages.append(message)
+        return updated_messages
 
-    def fix(self, items: t.Sequence[Message]) -> t.Sequence[Message]:
-        return [message.cache(False) for message in items]
+
+g_fixups = [
+    OpenAIToolsWithImageURLsFixup(),
+    CacheTooSmallFixup(),
+    GroqAssistantContentFixup(),
+]
 
 
 class LiteLLMGenerator(Generator):
@@ -123,8 +152,6 @@ class LiteLLMGenerator(Generator):
     _last_request_time: datetime.datetime | None = None
     _supports_function_calling: bool | None = None
 
-    _fixups = Fixups(available=[OpenAIToolsWithImageURLsFixup(), CacheTooSmallFixup()])
-
     @property
     def semaphore(self) -> asyncio.Semaphore:
         if self._semaphore is None:
@@ -299,6 +326,7 @@ def _parse_text_completion_response(
             extra={"response_id": response.id},
         )
 
+    @with_fixups(*g_fixups)
     async def _generate_message(
         self,
         messages: t.Sequence[Message],
@@ -313,20 +341,12 @@ async def _generate_message(
             if self._wrap is not None:
                 acompletion = self._wrap(acompletion)
 
-            # Prepare messages for specific providers
-            messages = await self._apply_fixups(messages)
-
-            try:
-                response = await acompletion(
-                    model=self.model,
-                    messages=[message.to_openai_spec() for message in messages],
-                    api_key=self.api_key,
-                    **self.params.merge_with(params).to_dict(),
-                )
-            except Exception as e:
-                if self._check_fixups(e):
-                    return await self._generate_message(messages, params)
-                raise
+            response = await acompletion(
+                model=self.model,
+                messages=[message.to_openai_spec() for message in messages],
+                api_key=self.api_key,
+                **self.params.merge_with(params).to_dict(),
+            )
 
             self._last_request_time = datetime.datetime.now(tz=datetime.timezone.utc)
             return self._parse_model_response(response)

rigging/message.py

@@ -336,6 +336,8 @@ def save(self, path: Path | str) -> None:
 """The types of content that can be included in a message."""
 ContentTypes = (ContentText, ContentImageUrl, ContentAudioInput)
 
+CompatibilityFlag = t.Literal["content_as_str"]
+
 
 class Message(BaseModel):
     """
@@ -364,6 +366,8 @@ class Message(BaseModel):
     tool_call_id: str | None = Field(None)
     """Associated call id if this message is a response to a tool call."""
 
+    _compability_flags: set[CompatibilityFlag] = set()
+
     def __init__(
         self,
         role: Role,
@@ -546,7 +550,7 @@ def to_openai_spec(self) -> dict[str, t.Any]:
                 isinstance(current, dict)
                 and current.get("type") == "text"
                 and next_.get("type") == "text"
-                and not current.get("text", "").endswith("\n")
+                and not str(current.get("text", "")).endswith("\n")
             ):
                 current["text"] += "\n"
 
@@ -556,6 +560,17 @@ def to_openai_spec(self) -> dict[str, t.Any]:
             if isinstance(part, dict) and part.get("type") == "input_audio":
                 part.get("input_audio", {}).pop("transcript", None)
 
+        # If enabled, we need to convert our content to a flat
+        # string for API compatibility. Groq is an example of an API
+        # which will complain for some roles if we send a list of content parts.
+
+        if "content_as_str" in self._compability_flags:
+            obj["content"] = "".join(
+                part["text"]
+                for part in obj["content"]
+                if isinstance(part, dict) and part.get("type") == "text"
+            )
+
         return obj
 
     # TODO: In general the add/remove/sync_part methods are