vt

backend/app/app/api/api_v1/endpoints/text2zip.py

@@ -72,7 +72,7 @@ def gen_prompt(content:str):
                                           "Directly answer to me. "
                                           "If the input is too long, summarize it. "
                                           "If the input is too short, expand it with reasonable imagination."
-                                          "Keep the prompt within 200 tokens. "},
+                                          "Keep the prompt within 70 tokens. "},
             {
                 "role": "user",
                 "content": content

backend/app/app/core/video_utils.py

@@ -10,8 +10,18 @@ from translate import Translator
 from langdetect import detect
 from chardet.universaldetector import UniversalDetector
 import numpy as np
+from openai import OpenAI
+from iso639 import Lang
 
 DEFAULT_ENCODING = "utf-8"
+client = OpenAI(base_url="http://192.168.192.84:8080/v1", api_key='choozmo9')
+system_prompt = (
+     "You are a precise and literal translator. "
+     "Translate the user's input from {from_lang} to {to_lang} as faithfully and literally as possible. "
+     "Preserve the original structure and vocabulary. "
+     "Avoid paraphrasing, interpretation, or creative rewrites. "
+     "Do not add explanations. Just return the translated text only."
+)
 
 def guess_codec(filenames: list) -> str:
   codec_detector = UniversalDetector()
@@ -104,21 +114,41 @@ def update_zip(zip_path, lang, new_filename, voice):
     os.remove(zip_path)
     
 def translate_table(table, lang):
-    
     print(f"translate to {lang}")
     for i in range(len(table)):
-        if (text:=table.loc[i, ['大標']].item()):
+        if (not table.loc[i, ['大標']].isna().item()) and (text:=table.loc[i, ['字幕']].item()):
             print("大標:",text)
-            translator= Translator(to_lang=lang, from_lang=detect(text))
-            translation = translator.translate(text)
+            #translator= Translator(to_lang=lang, from_lang=detect(text))
+            #translation = translator.translate(text)
+            translation = translate(text, lang)
             print("大標翻譯:",translation)
             table.loc[i, ['大標']] = translation
-        if (text:=table.loc[i, ['字幕']].item()):
+        if (not table.loc[i, ['字幕']].isna().item()) and (text:=table.loc[i, ['字幕']].item()):
             print('字幕:',text)
-            translator= Translator(to_lang=lang, from_lang=detect(text))
-            translation = translator.translate(text)
+            #translator= Translator(to_lang=lang, from_lang=detect(text))
+            #translation = translator.translate(text)
+            translation = translate(text, lang)
             print('字幕翻譯:',translation)
             table.loc[i, ['字幕']] = translation
     return table
         
+def translate(text, to_lang:str):
+    from_lang = Lang(detect(text)).name
+    to_lang = Lang(to_lang.split("-")).name
+    if to_lang == "Chinese":
+        to_lang = "Traditional Chinese"
+    completion = client.chat.completions.create(
+        model="gemma",
+        messages=[
+            {
+                "role": "system",
+                "content": system_prompt.format(from_lang=from_lang, to_lang=to_lang)
+            },
+            {
+                "role": "user",
+                "content": text
+            }
+        ]
+    )
+    return completion.choices[0].message.content