Insert Instructions at the end of System Prompt

pydantic_ai_slim/pydantic_ai/models/anthropic.py

@@ -853,7 +853,7 @@ async def _map_message( # noqa: C901
  else:
  assert_never(m)
  if instructions := self._get_instructions(messages, model_request_parameters):
- system_prompt_parts.insert(0, instructions)
+ system_prompt_parts.append(instructions)
  system_prompt = '\n\n'.join(system_prompt_parts)
 
  # Add cache_control to the last message content if anthropic_cache_messages is enabled

pydantic_ai_slim/pydantic_ai/models/bedrock.py

@@ -615,7 +615,7 @@ async def _map_messages( # noqa: C901
  last_message = cast(dict[str, Any], current_message)
 
  if instructions := self._get_instructions(messages, model_request_parameters):
- system_prompt.insert(0, {'text': instructions})
+ system_prompt.append({'text': instructions})
 
  return system_prompt, processed_messages
 

pydantic_ai_slim/pydantic_ai/models/cohere.py

@@ -234,6 +234,9 @@ def _map_messages(
  ) -> list[ChatMessageV2]:
  """Just maps a `pydantic_ai.Message` to a `cohere.ChatMessageV2`."""
  cohere_messages: list[ChatMessageV2] = []
+ system_prompt_count = sum(
+ 1 for m in messages if isinstance(m, ModelRequest) for p in m.parts if isinstance(p, SystemPromptPart)
+ )
  for message in messages:
  if isinstance(message, ModelRequest):
  cohere_messages.extend(self._map_user_message(message))
@@ -271,7 +274,7 @@ def _map_messages(
  else:
  assert_never(message)
  if instructions := self._get_instructions(messages, model_request_parameters):
- cohere_messages.insert(0, SystemChatMessageV2(role='system', content=instructions))
+ cohere_messages.insert(system_prompt_count, SystemChatMessageV2(role='system', content=instructions))
  return cohere_messages
 
  def _get_tools(self, model_request_parameters: ModelRequestParameters) -> list[ToolV2]:

pydantic_ai_slim/pydantic_ai/models/gemini.py

@@ -363,7 +363,7 @@ async def _message_to_gemini_content(
  else:
  assert_never(m)
  if instructions := self._get_instructions(messages, model_request_parameters):
- sys_prompt_parts.insert(0, _GeminiTextPart(text=instructions))
+ sys_prompt_parts.append(_GeminiTextPart(text=instructions))
  return sys_prompt_parts, contents
 
  async def _map_user_prompt(self, part: UserPromptPart) -> list[_GeminiPartUnion]:

pydantic_ai_slim/pydantic_ai/models/google.py

@@ -588,7 +588,7 @@ async def _map_messages(
  contents = [{'role': 'user', 'parts': [{'text': ''}]}]
 
  if instructions := self._get_instructions(messages, model_request_parameters):
- system_parts.insert(0, {'text': instructions})
+ system_parts.append({'text': instructions})
  system_instruction = ContentDict(role='user', parts=system_parts) if system_parts else None
 
  return system_instruction, contents

pydantic_ai_slim/pydantic_ai/models/groq.py

@@ -395,6 +395,9 @@ def _map_messages(
  ) -> list[chat.ChatCompletionMessageParam]:
  """Just maps a `pydantic_ai.Message` to a `groq.types.ChatCompletionMessageParam`."""
  groq_messages: list[chat.ChatCompletionMessageParam] = []
+ system_prompt_count = sum(
+ 1 for m in messages if isinstance(m, ModelRequest) for p in m.parts if isinstance(p, SystemPromptPart)
+ )
  for message in messages:
  if isinstance(message, ModelRequest):
  groq_messages.extend(self._map_user_message(message))
@@ -428,7 +431,9 @@ def _map_messages(
  else:
  assert_never(message)
  if instructions := self._get_instructions(messages, model_request_parameters):
- groq_messages.insert(0, chat.ChatCompletionSystemMessageParam(role='system', content=instructions))
+ groq_messages.insert(
+ system_prompt_count, chat.ChatCompletionSystemMessageParam(role='system', content=instructions)
+ )
  return groq_messages
 
  @staticmethod

pydantic_ai_slim/pydantic_ai/models/huggingface.py

@@ -327,6 +327,9 @@ async def _map_messages(
  ) -> list[ChatCompletionInputMessage | ChatCompletionOutputMessage]:
  """Just maps a `pydantic_ai.Message` to a `huggingface_hub.ChatCompletionInputMessage`."""
  hf_messages: list[ChatCompletionInputMessage | ChatCompletionOutputMessage] = []
+ system_prompt_count = sum(
+ 1 for m in messages if isinstance(m, ModelRequest) for p in m.parts if isinstance(p, SystemPromptPart)
+ )
  for message in messages:
  if isinstance(message, ModelRequest):
  async for item in self._map_user_message(message):
@@ -361,7 +364,7 @@ async def _map_messages(
  else:
  assert_never(message)
  if instructions := self._get_instructions(messages, model_request_parameters):
- hf_messages.insert(0, ChatCompletionInputMessage(content=instructions, role='system')) # type: ignore
+ hf_messages.insert(system_prompt_count, ChatCompletionInputMessage(content=instructions, role='system')) # type: ignore
  return hf_messages
 
  @staticmethod

pydantic_ai_slim/pydantic_ai/models/mistral.py

@@ -528,6 +528,9 @@ def _map_messages(
  ) -> list[MistralMessages]:
  """Just maps a `pydantic_ai.Message` to a `MistralMessage`."""
  mistral_messages: list[MistralMessages] = []
+ system_prompt_count = sum(
+ 1 for m in messages if isinstance(m, ModelRequest) for p in m.parts if isinstance(p, SystemPromptPart)
+ )
  for message in messages:
  if isinstance(message, ModelRequest):
  mistral_messages.extend(self._map_user_message(message))
@@ -557,7 +560,7 @@ def _map_messages(
  else:
  assert_never(message)
  if instructions := self._get_instructions(messages, model_request_parameters):
- mistral_messages.insert(0, MistralSystemMessage(content=instructions))
+ mistral_messages.insert(system_prompt_count, MistralSystemMessage(content=instructions))
 
  # Post-process messages to insert fake assistant message after tool message if followed by user message
  # to work around `Unexpected role 'user' after role 'tool'` error.

pydantic_ai_slim/pydantic_ai/models/openai.py

@@ -831,7 +831,11 @@ async def _map_messages(
  ) -> list[chat.ChatCompletionMessageParam]:
  """Just maps a `pydantic_ai.Message` to a `openai.types.ChatCompletionMessageParam`."""
  openai_messages: list[chat.ChatCompletionMessageParam] = []
+ system_prompt_count = 0
  for message in messages:
+ for part in message.parts:
+ if isinstance(part, SystemPromptPart):
+ system_prompt_count += 1
  if isinstance(message, ModelRequest):
  async for item in self._map_user_message(message):
  openai_messages.append(item)
@@ -840,7 +844,9 @@ async def _map_messages(
  else:
  assert_never(message)
  if instructions := self._get_instructions(messages, model_request_parameters):
- openai_messages.insert(0, chat.ChatCompletionSystemMessageParam(content=instructions, role='system'))
+ openai_messages.insert(
+ system_prompt_count, chat.ChatCompletionSystemMessageParam(content=instructions, role='system')
+ )
  return openai_messages
 
  @staticmethod
@@ -1313,7 +1319,12 @@ async def _responses_create( # noqa: C901
  # > Response input messages must contain the word 'json' in some form to use 'text.format' of type 'json_object'.
  # Apparently they're only checking input messages for "JSON", not instructions.
  assert isinstance(instructions, str)
- openai_messages.insert(0, responses.EasyInputMessageParam(role='system', content=instructions))
+ system_prompt_count = sum(
+ 1 for m in messages if isinstance(m, ModelRequest) for p in m.parts if isinstance(p, SystemPromptPart)
+ )
+ openai_messages.insert(
+ system_prompt_count, responses.EasyInputMessageParam(role='system', content=instructions)
+ )
  instructions = OMIT
 
  if verbosity := model_settings.get('openai_text_verbosity'):