fix summary language respects user primary language setting (#6470)

## Summary
- Summaries were generated in the conversation language instead of the
user's primary language setting
- Added `output_language_code` param to `get_transcript_structure`,
`get_reprocess_transcript_structure`, `extract_action_items`, and
`get_message_structure`
- `_get_structured` now fetches the user's language preference and
passes it as the output language — falls back to conversation language
if no preference is set

## Demo


https://github.com/user-attachments/assets/3a9b793f-0285-43b6-b45a-3625a3f66b79

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Author: mdmohsin7

app/pubspec.lock

@@ -253,10 +253,10 @@ packages:
     dependency: transitive
     description:
       name: characters
-      sha256: faf38497bda5ead2a8c7615f4f7939df04333478bf32e4173fcb06d428b5716b
+      sha256: f71061c654a3380576a52b451dd5532377954cf9dbd272a78fc8479606670803
       url: "https://pub.dev"
     source: hosted
-    version: "1.4.1"
+    version: "1.4.0"
   checked_yaml:
     dependency: transitive
     description:
@@ -1457,18 +1457,18 @@ packages:
     dependency: transitive
     description:
       name: matcher
-      sha256: "12956d0ad8390bbcc63ca2e1469c0619946ccb52809807067a7020d57e647aa6"
+      sha256: dc58c723c3c24bf8d3e2d3ad3f2f9d7bd9cf43ec6feaa64181775e60190153f2
       url: "https://pub.dev"
     source: hosted
-    version: "0.12.18"
+    version: "0.12.17"
   material_color_utilities:
     dependency: transitive
     description:
       name: material_color_utilities
-      sha256: "9c337007e82b1889149c82ed242ed1cb24a66044e30979c44912381e9be4c48b"
+      sha256: f7142bb1154231d7ea5f96bc7bde4bda2a0945d2806bb11670e30b850d56bdec
       url: "https://pub.dev"
     source: hosted
-    version: "0.13.0"
+    version: "0.11.1"
   mcumgr_flutter:
     dependency: "direct main"
     description:
@@ -2219,10 +2219,10 @@ packages:
     dependency: transitive
     description:
       name: test_api
-      sha256: "19a78f63e83d3a61f00826d09bc2f60e191bf3504183c001262be6ac75589fb8"
+      sha256: ab2726c1a94d3176a45960b6234466ec367179b87dd74f1611adb1f3b5fb9d55
       url: "https://pub.dev"
     source: hosted
-    version: "0.7.8"
+    version: "0.7.7"
   time:
     dependency: transitive
     description:

backend/utils/conversations/process_conversation.py

@@ -85,6 +85,7 @@ def _get_structured(
 ) -> Tuple[Structured, bool]:
     try:
         tz = notification_db.get_user_time_zone(uid)
+        user_language = users_db.get_user_language_preference(uid) or language_code
 
         # Fetch existing action items from past 2 days for deduplication
         existing_action_items = None
@@ -113,6 +114,7 @@ def _get_structured(
                         language_code,
                         tz,
                         calendar_meeting_context=calendar_context,
+                        output_language_code=user_language,
                     )
                 with track_usage(uid, Features.CONVERSATION_ACTION_ITEMS):
                     structured.action_items = extract_action_items(
@@ -122,13 +124,19 @@ def _get_structured(
                         tz,
                         existing_action_items=existing_action_items,
                         calendar_meeting_context=calendar_context,
+                        output_language_code=user_language,
                     )
                 return structured, False
 
             if conversation.text_source == ExternalIntegrationConversationSource.message:
                 with track_usage(uid, Features.CONVERSATION_STRUCTURE):
                     structured = get_message_structure(
-                        conversation.text, conversation.started_at, language_code, tz, conversation.text_source_spec
+                        conversation.text,
+                        conversation.started_at,
+                        language_code,
+                        tz,
+                        conversation.text_source_spec,
+                        output_language_code=user_language,
                     )
                 return structured, False
 
@@ -153,6 +161,7 @@ def _get_structured(
                     tz,
                     conversation.structured.title,
                     photos=conversation.photos,
+                    output_language_code=user_language,
                 )
             with track_usage(uid, Features.CONVERSATION_ACTION_ITEMS):
                 structured.action_items = extract_action_items(
@@ -162,6 +171,7 @@ def _get_structured(
                     tz,
                     photos=conversation.photos,
                     existing_action_items=existing_action_items,
+                    output_language_code=user_language,
                 )
             return structured, False
 
@@ -185,6 +195,7 @@ def _get_structured(
                 tz,
                 photos=conversation.photos,
                 calendar_meeting_context=calendar_context,
+                output_language_code=user_language,
             )
         with track_usage(uid, Features.CONVERSATION_ACTION_ITEMS):
             structured.action_items = extract_action_items(
@@ -195,6 +206,7 @@ def _get_structured(
                 photos=conversation.photos,
                 existing_action_items=existing_action_items,
                 calendar_meeting_context=calendar_context,
+                output_language_code=user_language,
             )
         return structured, False
     except Exception as e:

backend/utils/llm/conversation_processing.py

@@ -306,6 +306,7 @@ def extract_action_items(
     photos: List[ConversationPhoto] = None,
     existing_action_items: List[dict] = None,
     calendar_meeting_context: 'CalendarMeetingContext' = None,
+    output_language_code: str = None,
 ) -> List[ActionItem]:
     """
     Dedicated function to extract action items from conversation content.
@@ -538,9 +539,10 @@ def extract_action_items(
         '    ', ''
     ).strip()
 
+    response_language = output_language_code or language_code
     action_items_parser = PydanticOutputParser(pydantic_object=ActionItemsExtraction)
     # Second system message: conversation context + existing items (dynamic, per-conversation)
-    context_message = 'The content language is {language_code}. Use the same language {language_code} for your response.\n\nContent:\n{conversation_context}{existing_items_context}'
+    context_message = 'The content language is {language_code}. You MUST respond entirely in {response_language}.\n\nContent:\n{conversation_context}{existing_items_context}'
     prompt = ChatPromptTemplate.from_messages([('system', instructions_text), ('system', context_message)])
     chain = prompt | llm_medium_experiment.bind(prompt_cache_key="omi-extract-actions") | action_items_parser
 
@@ -552,6 +554,7 @@ def extract_action_items(
                 'conversation_context': conversation_context,
                 'format_instructions': action_items_parser.get_format_instructions(),
                 'language_code': language_code,
+                'response_language': response_language,
                 'started_at': started_at.isoformat(),
                 'current_time': current_time.isoformat(),
                 'tz': tz,
@@ -589,14 +592,17 @@ def get_transcript_structure(
     tz: str,
     photos: List[ConversationPhoto] = None,
     calendar_meeting_context: 'CalendarMeetingContext' = None,
+    output_language_code: str = None,
 ) -> Structured:
     conversation_context = _build_conversation_context(transcript, photos, calendar_meeting_context)
     if not conversation_context:
         return Structured()  # Should be caught by discard logic, but as a safeguard.
 
+    response_language = output_language_code or language_code
+
     # First system message: task-specific instructions (static prefix enables cross-conversation caching)
+    # NOTE: language instructions are in context_message (second message) to keep this prefix fully static.
     instructions_text = '''You are an expert content analyzer. Your task is to analyze the provided content (which could be a transcript, a series of photo descriptions from a wearable camera, or both) and provide structure and clarity.
-    The content language is {language_code}. Use the same language {language_code} for your response.
 
     CRITICAL: If CALENDAR MEETING CONTEXT is provided with participant names, you MUST use those names:
     - The conversation DEFINITELY happened between the named participants
@@ -641,7 +647,7 @@ def get_transcript_structure(
     ).strip()
 
     # Second system message: conversation context (dynamic, per-conversation)
-    context_message = 'Content:\n{conversation_context}'
+    context_message = 'The content language is {language_code}. You MUST respond entirely in {response_language}.\n\nContent:\n{conversation_context}'
     prompt = ChatPromptTemplate.from_messages([('system', instructions_text), ('system', context_message)])
     chain = prompt | llm_medium_experiment.bind(prompt_cache_key="omi-transcript-structure") | parser
 
@@ -650,6 +656,7 @@ def get_transcript_structure(
             'conversation_context': conversation_context,
             'format_instructions': parser.get_format_instructions(),
             'language_code': language_code,
+            'response_language': response_language,
             'started_at': started_at.isoformat(),
             'tz': tz,
         }
@@ -670,6 +677,7 @@ def get_reprocess_transcript_structure(
     tz: str,
     title: str,
     photos: List[ConversationPhoto] = None,
+    output_language_code: str = None,
 ) -> Structured:
     context_parts = []
     if transcript and transcript.strip():
@@ -684,9 +692,10 @@ def get_reprocess_transcript_structure(
         return Structured()
 
     full_context = "\n\n".join(context_parts)
+    response_language = output_language_code or language_code
 
     prompt_text = '''You are an expert content analyzer. Your task is to analyze the provided content (which could be a transcript, a series of photo descriptions from a wearable camera, or both) and provide structure and clarity.
-    The content language is {language_code}. Use the same language {language_code} for your response.
+    The content language is {language_code}. You MUST respond entirely in {response_language}.
 
     For the title, use ```{title}```, if it is empty, use the main topic of the content.
     For the overview, condense the content into a summary with the main topics discussed or scenes observed, making sure to capture the key points and important details.
@@ -732,6 +741,7 @@ def get_reprocess_transcript_structure(
             'title': title,
             'format_instructions': parser.get_format_instructions(),
             'language_code': language_code,
+            'response_language': response_language,
             'started_at': started_at.isoformat(),
             'tz': tz,
         }

backend/utils/llm/external_integrations.py

@@ -15,11 +15,17 @@
 
 
 def get_message_structure(
-    text: str, started_at: datetime, language_code: str, tz: str, text_source_spec: str = None
+    text: str,
+    started_at: datetime,
+    language_code: str,
+    tz: str,
+    text_source_spec: str = None,
+    output_language_code: str = None,
 ) -> Structured:
+    response_language = output_language_code or language_code
     prompt_text = '''
     You are an expert message analyzer. Your task is to analyze the message content and provide structure and clarity.
-    The message language is {language_code}. Use the same language {language_code} for your response.
+    The message language is {language_code}. You MUST respond entirely in {response_language}.
 
     For the title, create a concise title that captures the main topic of the message.
     For the overview, summarize the message with the main points discussed, make sure to capture the key information and important details.
@@ -38,6 +44,7 @@ def get_message_structure(
     response = chain.invoke(
         {
             'language_code': language_code,
+            'response_language': response_language,
             'started_at': started_at.isoformat(),
             'tz': tz,
             'text': text,