Update streamlit_app.py

streamlit_app.py

@@ -297,70 +297,56 @@ def get_runtime_api_key():
         return key
     return os.getenv("GOOGLE_API_KEY", "").strip() or None
 
-# Robust responses caller: SDK preferred, then attempt HTTP fallback via requests
-# --- REPLACE/ADD START ---
+# responses caller: prefer SDK responses, fallback to generativelanguage generate endpoints
 import json
 import requests
 
 def responses_generate(model, messages, files, max_output_tokens, api_key):
     if not api_key:
         raise RuntimeError("No API key for responses_generate")
-    # If genai SDK present and has responses.generate, prefer it
+    sdk_err = None
     if HAS_GENAI and genai is not None:
         try:
             genai.configure(api_key=api_key)
             if hasattr(genai, "responses") and getattr(genai, "responses") is not None:
                 return genai.responses.generate(model=model, messages=messages, files=files, max_output_tokens=max_output_tokens)
         except Exception as e:
-            # allow fallback to HTTP and surface SDK error in debug later
             sdk_err = str(e)
-    else:
-        sdk_err = "SDK not available"
 
-    # HTTP fallback using requests with retries
-    url = "https://gen-ai.googleapis.com/v1/responses"
-    payload = {"model": model, "messages": messages, "max_output_tokens": max_output_tokens}
+    host = "https://generativelanguage.googleapis.com"
+    candidates = [
+        f"{host}/v1/models/{model}:generate",
+        f"{host}/v1beta3/models/{model}:generate",
+        f"{host}/v1beta2/models/{model}:generate",
+    ]
+    # adapt messages to a simple prompt wrapper expected by generate
+    payload = {"prompt": {"messages": messages}, "maxOutputTokens": int(max_output_tokens or 512)}
     headers = {"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}
     last_exc = None
-    for attempt in range(3):
+    for url in candidates:
         try:
-            resp = requests.post(url, json=payload, headers=headers, timeout=30)
-            # Try to parse JSON on success
-            if resp.status_code == 200:
+            r = requests.post(url, json=payload, headers=headers, timeout=60)
+            if r.status_code == 200:
                 try:
-                    return resp.json()
+                    return r.json()
                 except Exception:
-                    return {"text": resp.text}
-            else:
-                last_exc = RuntimeError(f"HTTP {resp.status_code}: {resp.text}")
-                time.sleep(1 + attempt)
+                    return {"text": r.text}
+            last_exc = RuntimeError(f"HTTP {r.status_code}: {r.text}")
         except Exception as e:
             last_exc = e
-            time.sleep(1 + attempt)
-
-    # If we get here, both SDK and HTTP failed. Provide diagnostic info.
-    diag = {
-        "sdk_error": sdk_err if 'sdk_err' in locals() else None,
-        "http_error": str(last_exc),
-        "http_url": url,
-        "payload_sample": json.dumps(payload)[:1000],
-        "note": "See Streamlit 'Debug (compact)' output for response keys/body"
-    }
+    diag = {"sdk_error": sdk_err, "http_error": str(last_exc), "tried_urls": candidates}
     raise RuntimeError(f"genai.responses not available and HTTP fallback failed: {diag}")
 
-# call wrapper that returns response or raises; replace the previous call_responses_once with this:
 def call_responses_once(model_used, system_msg, user_msg, fname, max_tokens):
     files = [{"name": fname}] if fname else None
     for attempt in range(2):
         try:
             return responses_generate(model_used, [system_msg, user_msg], files, max_tokens, api_key=get_runtime_api_key())
-        except Exception as e:
+        except Exception:
             if attempt == 0:
                 time.sleep(1.0)
                 continue
             raise
-# --- REPLACE/ADD END ---
-
 
 if (st.session_state.get("busy") is False) and ('generate_now' in locals() and generate_now):
     if not st.session_state.get("videos"):
@@ -412,7 +398,6 @@ if (st.session_state.get("busy") is False) and ('generate_now' in locals() and g
                             st.session_state["last_loaded_path"] = current_path
                             st.session_state["file_hash"] = current_hash
                     else:
-                        # no SDK upload support: continue without upload and try HTTP responses with no file reference
                         uploaded = None
                         processed = None
                         st.session_state["uploaded_file"] = None
@@ -432,18 +417,6 @@ if (st.session_state.get("busy") is False) and ('generate_now' in locals() and g
                 system_msg = {"role": "system", "content": "You are a helpful assistant that summarizes videos concisely in vivid detail."}
                 user_msg = {"role": "user", "content": prompt_text}
 
-                def call_responses_once(model_used, system_msg, user_msg, fname, max_tokens):
-                    for attempt in range(2):
-                        try:
-                            files = [{"name": fname}] if fname else None
-                            resp = responses_generate(model_used, [system_msg, user_msg], files, max_tokens, api_key=runtime_key)
-                            return resp
-                        except Exception:
-                            if attempt == 0:
-                                time.sleep(1.0)
-                                continue
-                            raise
-
                 fname = file_name_or_id(processed) or file_name_or_id(uploaded)
                 response = call_responses_once(model_used, system_msg, user_msg, fname, max_tokens)