Add model system prompt support and update prompt hierarchy to: default → model → user

Co-authored-by: owndev <69784886+owndev@users.noreply.github.com>

Author: Copilot

docs/google-gemini-integration.md

@@ -253,15 +253,19 @@ Native tool calling is enabled/disabled via the standard 'Function calling' Open
 
 ## System Prompt Hierarchy
 
-The Google Gemini pipeline supports a hierarchical system prompt configuration that combines multiple sources. This allows for flexible customization at different levels: global defaults, per-user personalization, and per-chat customization.
+The Google Gemini pipeline supports a hierarchical system prompt configuration that combines multiple sources. This allows for flexible customization at different levels: global defaults, model-specific settings, and user preferences.
 
 ### Prompt Sources (in order of combination)
 
 1. **Default System Prompt** (`GOOGLE_DEFAULT_SYSTEM_PROMPT`): Global default applied to all chats, configurable via environment variable or Admin UI valves.
 
-2. **Per-User System Prompt** (User Personalization): Each user can set their own system prompt in Open WebUI's Settings > Personalization. This is stored in `settings.ui.system` and is automatically included.
+2. **Model System Prompt**: The system prompt configured in the model settings (Admin > Models > Select Model > System Prompt). This is passed via `__metadata__["chat"]["params"]["system"]`.
 
-3. **Chat-Level System Prompt**: The system message defined in the model settings or passed with individual chat messages.
+3. **User System Prompt**: The user's personalized system prompt from either:
+   - **User Settings** (Settings > Personalization): Stored in `settings.ui.system`
+   - **Chat Controls**: The system message passed with individual chat messages
+   
+   Note: User settings take precedence over chat controls if both are set.
 
 ### How It Works
 
@@ -270,9 +274,9 @@ All available prompts are combined in order, separated by blank lines:
 ```
 {Default System Prompt}
 
-{Per-User Personalization Prompt}
+{Model System Prompt}
 
-{Chat-Level System Prompt}
+{User System Prompt}
 ```
 
 If only one prompt source is set, it is used as-is without any additional formatting.
@@ -283,20 +287,30 @@ If only one prompt source is set, it is used as-is without any additional format
 
 ```bash
 # Default system prompt applied to all chats
-# Combined with per-user and chat-level prompts if they exist
+# Combined with model and user prompts if they exist
 GOOGLE_DEFAULT_SYSTEM_PROMPT="You are a helpful AI assistant. Always be concise and accurate."
 ```
 
 Or configure through the pipeline valves in Open WebUI's Admin panel.
 
-**Per-User Personalization:**
+**Model System Prompt:**
+
+Configure the model's system prompt in Open WebUI:
+1. Go to Admin > Models
+2. Select the model you want to configure
+3. Enter the system prompt in the "System Prompt" field
+4. Save settings
+
+This prompt will be applied to all chats using this model.
+
+**User System Prompt:**
 
 Users can set their personalized system prompt in Open WebUI:
 1. Go to Settings > Personalization
 2. Enter your preferred system prompt in the "System Prompt" field
 3. Save settings
 
-This prompt will be automatically applied to all your Gemini chats, combined with any default and chat-level prompts.
+This prompt will be automatically applied to all your Gemini chats, combined with any default and model prompts.
 
 ### Example
 
@@ -307,23 +321,23 @@ If your configuration is:
 You are a helpful AI assistant.
 ```
 
-**Per-user personalization prompt (Settings > Personalization):**
+**Model system prompt (Admin > Models > System Prompt):**
 ```
-My name is John. I prefer detailed explanations.
+You specialize in Python programming.
 ```
 
-**Chat-level system prompt (model settings):**
+**User system prompt (Settings > Personalization OR chat controls):**
 ```
-Always respond in formal English.
+My name is John. I prefer detailed explanations.
 ```
 
 The combined system prompt sent to Gemini will be:
 ```
 You are a helpful AI assistant.
 
-My name is John. I prefer detailed explanations.
+You specialize in Python programming.
 
-Always respond in formal English.
+My name is John. I prefer detailed explanations.
 ```
 
 ## Thinking Configuration

pipelines/google/google_gemini.py

@@ -4,7 +4,7 @@
 author_url: https://github.com/owndev/
 project_url: https://github.com/owndev/Open-WebUI-Functions
 funding_url: https://github.com/sponsors/owndev
-version: 1.8.2
+version: 1.8.3
 required_open_webui_version: 0.6.26
 license: Apache License 2.0
 description: Highly optimized Google Gemini pipeline with advanced image generation capabilities, intelligent compression, and streamlined processing workflows.
@@ -34,7 +34,7 @@
   - Flexible upload fallback options and optimization controls
   - Configurable thinking levels (low/high) for Gemini 3 models
   - Configurable thinking budgets (0-32768 tokens) for Gemini 2.5 models
-  - Hierarchical system prompts (default, per-user personalization, per-chat)
+  - Hierarchical system prompts (default, model, user)
 """
 
 import os
@@ -209,8 +209,8 @@ class Valves(BaseModel):
         )
         DEFAULT_SYSTEM_PROMPT: str = Field(
             default=os.getenv("GOOGLE_DEFAULT_SYSTEM_PROMPT", ""),
-            description="Default system prompt applied to all chats. Combined with per-user personalization "
-            "and chat-level prompts in a hierarchy: default → per-user → per-chat. Leave empty to disable.",
+            description="Default system prompt applied to all chats. Combined with model and user prompts "
+            "in a hierarchy: default → model → user. Leave empty to disable.",
         )
 
         # Image Processing Configuration
@@ -320,28 +320,67 @@ def _get_user_personalization_prompt(
 
         return None
 
+    def _get_model_system_prompt(
+        self, __metadata__: Optional[dict] = None
+    ) -> Optional[str]:
+        """Get the model's system prompt from model settings.
+
+        In Open WebUI, each model can have its own system prompt configured
+        in Admin > Models > Select Model > System Prompt. This is typically
+        passed via __metadata__["chat"]["params"]["system"].
+
+        Args:
+            __metadata__: The metadata dict passed to the pipe method
+
+        Returns:
+            The model's system prompt or None if not set
+        """
+        if __metadata__ is None:
+            return None
+
+        try:
+            chat = __metadata__.get("chat")
+            if chat and isinstance(chat, dict):
+                params = chat.get("params")
+                if params and isinstance(params, dict):
+                    system_prompt = params.get("system")
+                    if system_prompt and isinstance(system_prompt, str):
+                        return system_prompt.strip() or None
+        except Exception as e:
+            self.log.debug(f"Could not retrieve model system prompt: {e}")
+
+        return None
+
     def _combine_system_prompts(
-        self, chat_system_prompt: Optional[str], __user__: Optional[dict] = None
+        self,
+        chat_system_prompt: Optional[str],
+        __user__: Optional[dict] = None,
+        __metadata__: Optional[dict] = None,
     ) -> Optional[str]:
-        """Combine default, per-user, and chat-level system prompts.
+        """Combine default, model, and user system prompts.
 
         Prompt hierarchy (all prompts are combined if set):
         1. DEFAULT_SYSTEM_PROMPT (environment/valve setting)
-        2. Per-user system prompt (from __user__["settings"]["ui"]["system"] - Personalization settings)
-        3. Chat-level system prompt (from messages in this request)
+        2. Model system prompt (from model settings - __metadata__["chat"]["params"]["system"])
+        3. User system prompt (from user settings OR chat controls/messages)
 
         Args:
             chat_system_prompt: The chat-level system prompt from messages (may be None)
             __user__: The user dict passed to the pipe method
+            __metadata__: The metadata dict passed to the pipe method
 
         Returns:
             Combined system prompt or None if none are set
         """
         default_prompt = self.valves.DEFAULT_SYSTEM_PROMPT.strip() or None
+        model_prompt = self._get_model_system_prompt(__metadata__)
         user_personalization = self._get_user_personalization_prompt(__user__)
         chat_prompt = chat_system_prompt.strip() if chat_system_prompt else None
 
-        prompts = [p for p in [default_prompt, user_personalization, chat_prompt] if p]
+        # User prompt = user personalization OR chat prompt (user settings take precedence)
+        user_prompt = user_personalization or chat_prompt
+
+        prompts = [p for p in [default_prompt, model_prompt, user_prompt] if p]
 
         if not prompts:
             return None
@@ -354,8 +393,8 @@ def _combine_system_prompts(
         self.log.debug(
             f"Combined system prompts: "
             f"default={len(default_prompt) if default_prompt else 0}, "
-            f"user_personalization={len(user_personalization) if user_personalization else 0}, "
-            f"chat={len(chat_prompt) if chat_prompt else 0} -> "
+            f"model={len(model_prompt) if model_prompt else 0}, "
+            f"user={len(user_prompt) if user_prompt else 0} -> "
             f"total={len(combined)} chars"
         )
         return combined
@@ -446,6 +485,7 @@ async def _build_image_generation_contents(
         messages: List[Dict[str, Any]],
         __event_emitter__: Callable,
         __user__: Optional[dict] = None,
+        __metadata__: Optional[dict] = None,
     ) -> Tuple[List[Dict[str, Any]], Optional[str]]:
         """Construct the contents payload for image-capable models.
 
@@ -459,7 +499,7 @@ async def _build_image_generation_contents(
 
         # Combine with default system prompt if configured
         system_instruction = self._combine_system_prompts(
-            user_system_instruction, __user__
+            user_system_instruction, __user__, __metadata__
         )
 
         last_user_msg = next(
@@ -919,14 +959,18 @@ def _prepare_model_id(self, model_id: str) -> str:
         return model_id
 
     def _prepare_content(
-        self, messages: List[Dict[str, Any]], __user__: Optional[dict] = None
+        self,
+        messages: List[Dict[str, Any]],
+        __user__: Optional[dict] = None,
+        __metadata__: Optional[dict] = None,
     ) -> Tuple[List[Dict[str, Any]], Optional[str]]:
         """
         Prepare messages content for the API and extract system message if present.
 
         Args:
             messages: List of message objects from the request
             __user__: The user dict passed to the pipe method
+            __metadata__: The metadata dict passed to the pipe method
 
         Returns:
             Tuple of (prepared content list, system message string or None)
@@ -938,7 +982,9 @@ def _prepare_content(
         )
 
         # Combine with default system prompt if configured
-        system_message = self._combine_system_prompts(user_system_message, __user__)
+        system_message = self._combine_system_prompts(
+            user_system_message, __user__, __metadata__
+        )
 
         # Prepare contents for the API
         contents = []
@@ -2197,7 +2243,7 @@ async def pipe(
                         contents,
                         system_instruction,
                     ) = await self._build_image_generation_contents(
-                        messages, __event_emitter__, __user__
+                        messages, __event_emitter__, __user__, __metadata__
                     )
                     # For image generation, system_instruction is integrated into the prompt
                     # so it will be None here (this is expected and correct)
@@ -2209,7 +2255,9 @@ async def pipe(
             else:
                 # For non-image generation models, use the full conversation history
                 # Prepare content and extract system message normally
-                contents, system_instruction = self._prepare_content(messages, __user__)
+                contents, system_instruction = self._prepare_content(
+                    messages, __user__, __metadata__
+                )
                 if not contents:
                     return "Error: No valid message content found"
                 self.log.debug(