Merge pull request #368 from harry0703/dev

enhanced exception handling for llm and optimized video concatenation

app/services/llm.py

@@ -9,6 +9,8 @@ from openai.types.chat import ChatCompletion
 
 from app.config import config
 
+_max_retries = 5
+
 
 def _generate_response(prompt: str) -> str:
     content = ""
@@ -219,11 +221,8 @@ Generate a script for a video, depending on the subject of the video.
 
     final_script = ""
     logger.info(f"subject: {video_subject}")
-    # logger.debug(f"prompt: \n{prompt}")
-    response = _generate_response(prompt=prompt)
 
-    # Return the generated script
+    def format_response(response):
-    if response:
         # Clean the script
         # Remove asterisks, hashes
         response = response.replace("*", "")
@@ -240,19 +239,30 @@ Generate a script for a video, depending on the subject of the video.
         selected_paragraphs = paragraphs[:paragraph_number]
 
         # Join the selected paragraphs into a single string
-        final_script = "\n\n".join(selected_paragraphs)
+        return "\n\n".join(selected_paragraphs)
 
-        # Print to console the number of paragraphs used
+    for i in range(_max_retries):
-        # logger.info(f"number of paragraphs used: {len(selected_paragraphs)}")
+        try:
-    else:
+            response = _generate_response(prompt=prompt)
-        logging.error("gpt returned an empty response")
+            if response:
+                final_script = format_response(response)
+            else:
+                logging.error("gpt returned an empty response")
 
-    # g4f may return an error message
+            # g4f may return an error message
-    if final_script and "当日额度已消耗完" in final_script:
+            if final_script and "当日额度已消耗完" in final_script:
-        raise ValueError(final_script)
+                raise ValueError(final_script)
+
+            if final_script:
+                break
+        except Exception as e:
+            logger.error(f"failed to generate script: {e}")
+
+        if i < _max_retries:
+            logger.warning(f"failed to generate video script, trying again... {i + 1}")
 
     logger.success(f"completed: \n{final_script}")
-    return final_script
+    return final_script.strip()
 
 
 def generate_terms(video_subject: str, video_script: str, amount: int = 5) -> List[str]:
@@ -283,25 +293,28 @@ Please note that you must use English for generating video search terms; Chinese
 """.strip()
 
     logger.info(f"subject: {video_subject}")
-    # logger.debug(f"prompt: \n{prompt}")
+
-    response = _generate_response(prompt)
     search_terms = []
+    for i in range(_max_retries):
+        try:
+            response = _generate_response(prompt)
+            search_terms = json.loads(response)
+            if not isinstance(search_terms, list) or not all(isinstance(term, str) for term in search_terms):
+                logger.error("response is not a list of strings.")
+                continue
 
-    try:
+        except Exception as e:
-        search_terms = json.loads(response)
+            match = re.search(r'\[.*]', response)
-        if not isinstance(search_terms, list) or not all(isinstance(term, str) for term in search_terms):
+            if match:
-            raise ValueError("response is not a list of strings.")
+                try:
+                    search_terms = json.loads(match.group())
+                except json.JSONDecodeError:
+                    pass
 
-    except (json.JSONDecodeError, ValueError):
+        if search_terms and len(search_terms) > 0:
-        # logger.warning(f"gpt returned an unformatted response. attempting to clean...")
+            break
-        # Attempt to extract list-like string and convert to list
+        if i < _max_retries:
-        match = re.search(r'\["(?:[^"\\]|\\.)*"(?:,\s*"[^"\\]*")*\]', response)
+            logger.warning(f"failed to generate video terms, trying again... {i + 1}")
-        if match:
-            try:
-                search_terms = json.loads(match.group())
-            except json.JSONDecodeError:
-                logger.error(f"could not parse response: {response}")
-                return []
 
     logger.success(f"completed: \n{search_terms}")
     return search_terms
@@ -310,8 +323,8 @@ Please note that you must use English for generating video search terms; Chinese
 if __name__ == "__main__":
     video_subject = "生命的意义是什么"
     script = generate_script(video_subject=video_subject, language="zh-CN", paragraph_number=1)
-    # print("######################")
+    print("######################")
-    # print(script)
+    print(script)
-    # search_terms = generate_terms(video_subject=video_subject, video_script=script, amount=5)
+    search_terms = generate_terms(video_subject=video_subject, video_script=script, amount=5)
-    # print("######################")
+    print("######################")
-    # print(search_terms)
+    print(search_terms)

app/services/video.py

@@ -60,15 +60,17 @@ def combine_videos(combined_video_path: str,
             end_time = min(start_time + max_clip_duration, clip_duration)
             split_clip = clip.subclip(start_time, end_time)
             raw_clips.append(split_clip)
-            logger.info(f"splitting from {start_time:.2f} to {end_time:.2f}, clip duration {clip_duration:.2f}, split_clip duration {split_clip.duration:.2f}")
+            # logger.info(f"splitting from {start_time:.2f} to {end_time:.2f}, clip duration {clip_duration:.2f}, split_clip duration {split_clip.duration:.2f}")
             start_time = end_time
+            if video_concat_mode.value == VideoConcatMode.sequential.value:
+                break
+
+    # random video_paths order
+    if video_concat_mode.value == VideoConcatMode.random.value:
+        random.shuffle(raw_clips)
+
     # Add downloaded clips over and over until the duration of the audio (max_duration) has been reached
     while video_duration < audio_duration:
-        # random video_paths order
-        if video_concat_mode.value == VideoConcatMode.random.value:
-            random.shuffle(raw_clips)
-
-
         for clip in raw_clips:
             # Check if clip is longer than the remaining audio
             if (audio_duration - video_duration) < clip.duration: