updated examples with gemini 2.5_pro

Author: staru09

examples/example_4_gemini_memory.py

@@ -6,7 +6,7 @@
 
 Usage:
     export GEMINI_API_KEY=your_api_key
-    python examples/example_5_gemini_memory.py
+    python examples/example_4_gemini_memory.py
 """
 
 import asyncio
@@ -53,7 +53,7 @@ async def main():
     3. Extracts memory categories from conversations
     4. Outputs the categories to files
     """
-    print("Example 5: Conversation Memory Processing (Google Gemini)")
+    print("Example 4: Conversation Memory Processing (Google Gemini)")
     print("-" * 55)
 
     api_key = os.getenv("GEMINI_API_KEY")
@@ -69,7 +69,7 @@ async def main():
                 "client_backend": "httpx",
                 "base_url": "https://generativelanguage.googleapis.com/v1beta",
                 "api_key": api_key,
-                "chat_model": "gemini-2.5-flash",  # Fast and capable model
+                "chat_model": "gemini-2.5-pro",  # Fast and capable model
                 "embed_model": "text-embedding-004",  # Gemini's embedding model
             },
         },

src/memu/llm/backends/gemini.py

@@ -7,7 +7,7 @@
 
 class GeminiLLMBackend(LLMBackend):
     """Backend for Google Gemini LLM API.
-    
+
     Gemini uses a different API format than OpenAI-compatible APIs:
     - Endpoint: /models/{model}:generateContent
     - Auth: x-goog-api-key header
@@ -22,36 +22,33 @@ def build_summary_payload(
     ) -> dict[str, Any]:
         """Build payload for Gemini generateContent API."""
         contents: list[dict[str, Any]] = []
-        
+
         # Add user message
         contents.append({
             "role": "user",
             "parts": [{"text": text}]
         })
-        
+
         payload: dict[str, Any] = {
             "contents": contents,
         }
-        
+
         # Add system instruction if provided
+        # Note: When system_prompt is None, we don't set a default to allow the user prompt
+        # to fully control the output format (e.g., for JSON responses)
         if system_prompt:
             payload["system_instruction"] = {
                 "parts": [{"text": system_prompt}]
             }
-        else:
-            # Default system prompt for summarization
-            payload["system_instruction"] = {
-                "parts": [{"text": "Summarize the text in one short paragraph."}]
-            }
-        
+
         # Add generation config
         generation_config: dict[str, Any] = {
             "temperature": 1.0,  # Gemini recommends keeping at 1.0
         }
         if max_tokens is not None:
             generation_config["maxOutputTokens"] = max_tokens
         payload["generationConfig"] = generation_config
-        
+
         return payload
 
     def parse_summary_response(self, data: dict[str, Any]) -> str:
@@ -83,30 +80,30 @@ def build_vision_payload(
                 }
             },
         ]
-        
+
         contents: list[dict[str, Any]] = [
             {
                 "role": "user",
                 "parts": user_parts,
             }
         ]
-        
+
         payload: dict[str, Any] = {
             "contents": contents,
         }
-        
+
         # Add system instruction if provided
         if system_prompt:
             payload["system_instruction"] = {
                 "parts": [{"text": system_prompt}]
             }
-        
+
         # Add generation config
         generation_config: dict[str, Any] = {
             "temperature": 1.0,
         }
         if max_tokens is not None:
             generation_config["maxOutputTokens"] = max_tokens
         payload["generationConfig"] = generation_config
-        
+
         return payload

src/memu/llm/http_client.py

@@ -49,7 +49,7 @@ def parse_embedding_response(self, data: dict[str, Any]) -> list[list[float]]:
 
 class _GeminiEmbeddingBackend(_EmbeddingBackend):
     """Gemini embedding API backend.
-    
+
     Gemini uses a different format:
     - Endpoint: /models/{model}:embedContent
     - Single text per request (batch supported via batchEmbedContents)