Upload app.py

app.py

@@ -792,7 +792,8 @@ def render_user_message(question: str) -> str:
 # Remote API Generation (via vLLM-compatible endpoint)
 # ============================================================
 
-def generate_response(prompt: str, max_new_tokens: int = MAX_NEW_TOKENS) -> str:
+
+async def generate_response(prompt: str, max_new_tokens: int = MAX_NEW_TOKENS) -> str:
     """Generate response using vLLM OpenAI-compatible API."""
     # Use /completions endpoint for raw prompt
     url = f"{REMOTE_API_BASE}/completions"
@@ -809,7 +810,9 @@ def generate_response(prompt: str, max_new_tokens: int = MAX_NEW_TOKENS) -> str:
         "stop": ["\n<tool_response>", "<tool_response>"],
     }
 
-    response = httpx.post(url, json=payload, headers=headers, timeout=300.0)
+    async with httpx.AsyncClient() as client:
+        response = await client.post(url, json=payload, headers=headers, timeout=300.0)
+        
     if response.status_code != 200:
         raise Exception(f"vLLM API error {response.status_code}: {response.text}")
 
@@ -820,7 +823,7 @@ def generate_response(prompt: str, max_new_tokens: int = MAX_NEW_TOKENS) -> str:
 # ============================================================
 # Streaming Agent Runner
 # ============================================================
-def run_agent_streaming(
+async def run_agent_streaming(
     question: str,
     serper_key: str,
     max_rounds: int
@@ -884,7 +887,7 @@ def run_agent_streaming(
                 yield ''.join(html_parts)
 
                 # Call ZeroGPU function
-                generated = generate_response(prompt, max_new_tokens=MAX_NEW_TOKENS)
+                generated = await generate_response(prompt, max_new_tokens=MAX_NEW_TOKENS)
                 
                 # Remove placeholder
                 html_parts.pop()
@@ -944,9 +947,9 @@ def run_agent_streaming(
                 result = ""
                 try:
                     if actual_fn == "search":
-                        result = asyncio.run(browser.search(args.get("query", ""), args.get("topn", 10)))
+                        result = await browser.search(args.get("query", ""), args.get("topn", 10))
                     elif actual_fn == "open":
-                        result = asyncio.run(browser.open(**args))
+                        result = await browser.open(**args)
                     elif actual_fn == "find":
                         result = browser.find(args.get("pattern", ""), args.get("cursor", -1))
                     else:
@@ -2510,7 +2513,7 @@ def create_interface():
                     clear_btn = gr.Button("🗑 Clear", scale=1)
 
         # Function to hide welcome and show output
-        def start_research(question, serper_key, max_rounds):
+        async def start_research(question, serper_key, max_rounds):
             # Generator that first hides welcome, then streams results
             # Also clears the input box for the next question
             
@@ -2518,7 +2521,7 @@ def create_interface():
             # IMPORTANT: Don't use empty string for output, or JS will hide the output area!
             yield "", '<div style="text-align: center; padding: 2rem; color: #6b7280;">Delving into it...</div>', ""
             
-            for result in run_agent_streaming(question, serper_key, max_rounds):
+            async for result in run_agent_streaming(question, serper_key, max_rounds):
                 yield "", result, ""
 
         # Event handlers