Skip to content

Bug: Spanish responses returned in single-message mode when English is selected

Summary

When the application is configured for single-message mode (CONTEXT_OPTION=single-message), the LLM (Gemma) responds in Spanish even when the user has selected English as their language. Switching to conversation-with-context-labels mode resolves the issue.

Root Cause

The language preference was lost at two points in the single-message request chain:

1. Frontend does not send language in single-message payload

File: components/gov-chat-frontend/src/components/ChatBotComponent.vue (lines 783-791)

The single-message payload omitted the context object entirely:

queryData = {
  userId: ...,
  sessionId: ...,
  text: messageForBackend,
  contextOption: contextOption,
  timestamp: ...,
};

While conversation-with-labels mode explicitly includes language (lines 756-780):

context: {
  categoryLabel: categoryLabel,
  serviceLabels: serviceLabels,
  language: this.currentLocale.toUpperCase(),  // e.g. "EN"
}

this.currentLocale is always available (initialized to "en" at line 447) but was never included in the single-message payload.

2. Backend strips all context when building OPEA payload for single-message mode

File: components/gov-chat-backend/services/query-service.js (lines 406-409)

if (backendMode === 'single-message') {
  opeaPayload = {
    messages: queryText,
    stream: false
    // context: completely omitted
  };
}

Even if the frontend were fixed, the backend unconditionally discarded all context (including language) for single-message mode. Conversation mode includes it (lines 410-421).

3. ChatQnA service crashes on string-type messages (discovered during testing)

File: genie-ai-overlay/chatqna/genieai_chatqna.py

In single-message mode, messages is sent as a plain string (the extracted last message text), not an array of message objects. The ChatQnA Python code had three locations that assumed full_chat_history was always a list of dicts:

  • _get_translated_history_string() (line 1090): iterated with message["content"] — crashed with TypeError: string indices must be integers
  • Language auto-detection fallback (line 1285): iterated with msg.get("role") — same crash
  • English history flattening (line 1327): iterated with msg.get('role') — same crash

Resolution

Fix 1: Frontend — Include language in single-message payload

File: components/gov-chat-frontend/src/components/ChatBotComponent.vue

Added context.language to the single-message queryData object:

queryData = {
  userId: ...,
  sessionId: ...,
  text: messageForBackend,
  context: {
    language: this.currentLocale.toUpperCase(),
  },
  contextOption: contextOption,
  timestamp: ...,
};

Fix 2: Backend — Pass context to OPEA in single-message mode

File: components/gov-chat-backend/services/query-service.js

Include context.language in the single-message OPEA payload:

if (backendMode === 'single-message') {
  opeaPayload = {
    messages: queryText,
    stream: false,
    context: {
      language: queryData.context?.language,
    },
  };
}

Fix 3: ChatQnA — Handle string-type messages in single-message mode

File: genie-ai-overlay/chatqna/genieai_chatqna.py

Three locations updated to handle full_chat_history being either a str or List[Dict]:

  • _get_translated_history_string(): string input is translated directly via _translate_text_chunk() instead of iterating
  • Language auto-detection fallback: string input is used directly as last_user_content
  • English history flattening: string input is used as-is

Affected Files

  • components/gov-chat-frontend/src/components/ChatBotComponent.vue
  • components/gov-chat-backend/services/query-service.js
  • genie-ai-overlay/chatqna/genieai_chatqna.py

Testing

Smoke-tested on local build (RTX 4060, Gemma 3 1B) with CONTEXT_OPTION=single-message:

  • English quickhelp: LLM responded in English, MANDATORY: respond ONLY in English guard injected correctly
  • Spanish quickhelp: Query translated to English, response translated back to Spanish
  • Indonesian quickhelp: Query translated to English, response translated back to Indonesian
  • Conversation-with-labels mode: No regressions observed

Fixed In

Branch: sprint-21-bug-fixes (commit ec2777cd)

Edited by david-full