Update app.py

app.py

@@ -29,7 +29,11 @@ app.add_middleware(
 API_KEY = os.environ.get("API_KEY", "change_me")
 logger.info(f"API key loaded: {API_KEY}")
 
-# NOTE: Ollama’s default endpoint is often "/generate", not "/api/generate"
+# IMPORTANT: Choose the correct endpoint for your Ollama version.
+#    - If your Ollama logs say "POST /generate" => "http://localhost:11434/generate"
+#    - If it says "POST /api/generate" => "http://localhost:11434/api/generate"
+# Double-check with a direct curl:  curl -X POST -d '{}' http://localhost:11434/generate
+# If that returns 404, try /api/generate instead.
 OLLAMA_SERVER_URL = "http://localhost:11434/generate"
 
 @app.post("/api/generate")
@@ -73,7 +77,7 @@ async def generate(request: Request):
                         json=payload,
                         headers={"Content-Type": "application/json"}
                     ) as response:
-                        # If we get 4xx or 5xx, raise_for_status() triggers HTTPStatusError
+                        # If we get a 4xx/5xx, raise_for_status() triggers HTTPStatusError
                         response.raise_for_status()
 
                         # Stream out the content as it arrives
@@ -85,12 +89,10 @@ async def generate(request: Request):
                 yield json.dumps({"error": "Unable to communicate with Ollama"})
 
             except httpx.HTTPStatusError as exc:
-                # Read the remaining body so we can log the error text
-                error_bytes = await exc.response.aread()
-                error_text = error_bytes.decode("utf-8", errors="replace")
-
-                logger.error(f"HTTP error from Ollama: {exc.response.status_code} - {error_text}")
-                yield json.dumps({"error": f"HTTP error: {error_text}"})
+                # The stream is already closed if we get 404 or another 4xx/5xx error
+                # so we can't call aread() safely. Let's just log the status code.
+                logger.error(f"HTTP error from Ollama: {exc.response.status_code}")
+                yield json.dumps({"error": f"HTTP error: {exc.response.status_code}"})
 
         return StreamingResponse(stream_response(), media_type="application/json")
 
@@ -108,4 +110,3 @@ if __name__ == "__main__":
     import uvicorn
     logger.info("Starting server on http://0.0.0.0:7860")
     uvicorn.run(app, host="0.0.0.0", port=7860)
-