feat: conectar Langfuse al get_llm() central via LangChain CallbackHandler

backend/requirements.txt

@@ -93,6 +93,7 @@ jsonpointer==3.0.0
 jsonschema==4.23.0
 jsonschema-specifications==2023.12.1
 kiwisolver==1.4.5
+langfuse==4.14.1
 langchain==1.2.9
 langchain-community==0.4.1
 langchain-core==1.3.3

backend/utils/believe_brain.py

@@ -38,6 +38,7 @@
 import database.memories as memories_db
 from models.conversation import Conversation
 from utils.conversations.render import conversation_to_dict
+from utils.llm.clients import generate_embedding
 
 logger = logging.getLogger(__name__)
 
@@ -102,6 +103,19 @@ def _build_episode(uid: str, conversation: Conversation) -> Optional[dict]:
     action_items = structured.get("action_items", []) or []
     events = structured.get("events", []) or []
 
+    # believe: embedding 768d (Ollama nomic-embed-text) para búsqueda semántica
+    # en el Company Brain (mc-brain-search). Fail-soft: si falla, el episodio
+    # entra sin vector y sigue siendo consultable por filtros.
+    embedding = None
+    try:
+        _embed_text = (overview or title or transcript_ref or "").strip()
+        if _embed_text:
+            _vec = generate_embedding(_embed_text[:4000])
+            if isinstance(_vec, list) and len(_vec) == 768:
+                embedding = _vec
+    except Exception as _e:
+        logger.warning(f"believe_brain: embedding failed for {conversation.id}: {_e}")
+
     return {
         "source": "omi",
         "external_id": conversation.id,
@@ -110,6 +124,7 @@ def _build_episode(uid: str, conversation: Conversation) -> Optional[dict]:
         "summary": overview or title or None,
         "transcript_ref": transcript_ref,
         "channel": "omi",
+        "embedding": embedding,
         "created_at": conversation.created_at.isoformat() if conversation.created_at else None,
         "metadata": {
             "omi_uid": uid,

backend/utils/llm/clients.py

@@ -21,6 +21,39 @@
 
 _usage_callback = get_usage_callback()
 
+
+def _init_langfuse_callback() -> Optional[Any]:
+    """Create the Langfuse LangChain callback handler if Langfuse is configured.
+
+    No-op (returns None) when LANGFUSE_HOST/LANGFUSE_PUBLIC_KEY/LANGFUSE_SECRET_KEY
+    aren't all set in the environment — keeps local/test setups working without
+    Langfuse and without touching startup behavior.
+    """
+    if not (
+        os.environ.get('LANGFUSE_HOST')
+        and os.environ.get('LANGFUSE_PUBLIC_KEY')
+        and os.environ.get('LANGFUSE_SECRET_KEY')
+    ):
+        return None
+    try:
+        from langfuse.langchain import CallbackHandler
+
+        return CallbackHandler()
+    except Exception:
+        logger.exception('Failed to initialize Langfuse callback handler — continuing without it')
+        return None
+
+
+_langfuse_callback = _init_langfuse_callback()
+
+
+def _llm_callbacks() -> List[Any]:
+    """Base callback list for every LangChain LLM client: usage tracking + Langfuse (if configured)."""
+    if _langfuse_callback is not None:
+        return [_usage_callback, _langfuse_callback]
+    return [_usage_callback]
+
+
 # ---------------------------------------------------------------------------
 # BYOK (Bring Your Own Key)
 #
@@ -90,7 +123,7 @@ def _create_byok_client(
     model: str, provider: str, byok_key: str, streaming: bool = False, feature: str = ''
 ) -> Optional[ChatOpenAI]:
     """Create a ChatOpenAI using the user's BYOK key. Returns None if BYOK not supported for this provider."""
-    kwargs: Dict[str, Any] = {'callbacks': [_usage_callback], 'request_timeout': 120, 'max_retries': 1}
+    kwargs: Dict[str, Any] = {'callbacks': _llm_callbacks(), 'request_timeout': 120, 'max_retries': 1}
     if model == 'gpt-5.1':
         kwargs['extra_body'] = {"prompt_cache_retention": "24h"}
     if streaming:
@@ -397,7 +430,7 @@ def _get_or_create_openai_llm(model_name: str, streaming: bool = False) -> ChatO
     key = (model_name, streaming, 'openai')
     if key not in _llm_cache:
         kwargs: Dict[str, Any] = {
-            'callbacks': [_usage_callback],
+            'callbacks': _llm_callbacks(),
             'request_timeout': 120,
             'max_retries': 1,
         }
@@ -426,7 +459,7 @@ def _get_or_create_openrouter_llm(
             'api_key': os.environ.get('OPENROUTER_API_KEY'),
             'base_url': "https://openrouter.ai/api/v1",
             'default_headers': {"X-Title": "Omi Chat"},
-            'callbacks': [_usage_callback],
+            'callbacks': _llm_callbacks(),
             'request_timeout': 120,
             'max_retries': 1,
         }
@@ -465,7 +498,7 @@ def _get_or_create_gemini_llm(
         use_vertex = os.environ.get('USE_VERTEX_AI', '').lower() == 'true'
         gcp_project = os.environ.get('GOOGLE_CLOUD_PROJECT', '') if use_vertex else ''
         gemini_key = os.environ.get('GEMINI_API_KEY', '')
-        kwargs: Dict[str, Any] = {'callbacks': [_usage_callback], 'timeout': 120, 'max_retries': 1}
+        kwargs: Dict[str, Any] = {'callbacks': _llm_callbacks(), 'timeout': 120, 'max_retries': 1}
         if streaming:
             kwargs['streaming'] = True
         if thinking_budget is not None and model_name.startswith('gemini-2.5'):
@@ -616,7 +649,7 @@ def get_qos_info() -> Dict[str, Dict[str, str]]:
 # Legacy module-level alias (kept for test compatibility).
 # Production code should use get_llm(feature) exclusively.
 # ---------------------------------------------------------------------------
-llm_mini = ChatOpenAI(model='gpt-4.1-mini', callbacks=[_usage_callback], request_timeout=120, max_retries=1)
+llm_mini = ChatOpenAI(model='gpt-4.1-mini', callbacks=_llm_callbacks(), request_timeout=120, max_retries=1)
 
 # ---------------------------------------------------------------------------
 # Embeddings, parser, utilities

backend/utils/llm/conversation_processing.py

@@ -629,7 +629,9 @@ def get_transcript_structure(
     if not conversation_context:
         return Structured()  # Should be caught by discard logic, but as a safeguard.
 
-    response_language = output_language_code or language_code
+    # Believe: title/overview feeding the Company Brain MUST always be in Spanish, regardless of
+    # the detected conversation language or the user's output preference.
+    response_language = 'Spanish'
 
     # 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.
@@ -723,7 +725,9 @@ def get_reprocess_transcript_structure(
         return Structured()
 
     full_context = "\n\n".join(context_parts)
-    response_language = output_language_code or language_code
+    # Believe: title/overview feeding the Company Brain MUST always be in Spanish, regardless of
+    # the detected conversation language or the user's output preference.
+    response_language = 'Spanish'
 
     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}. You MUST respond entirely in {response_language}.

backend/utils/llm/external_integrations.py

@@ -27,7 +27,9 @@ def get_message_structure(
     text_source_spec: str = None,
     output_language_code: str = None,
 ) -> Structured:
-    response_language = output_language_code or language_code
+    # Believe: title/overview feeding the Company Brain MUST always be in Spanish, regardless of
+    # the detected message language or the user's output preference.
+    response_language = 'Spanish'
     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}. You MUST respond entirely in {response_language}.
@@ -74,6 +76,7 @@ def get_message_structure(
 def summarize_experience_text(text: str, text_source_spec: str = None) -> Structured:
     source_context = f"Source: {text_source_spec}" if text_source_spec else "their own experiences or thoughts"
     prompt = f'''The user sent a text of {source_context}, and wants to create a memory from it.
+      You MUST respond entirely in Spanish (every word of the title and overview).
       For the title, use the main topic of the experience or thought.
       For the overview, condense the descriptions into a brief summary with the main topics discussed, make sure to capture the key points and important details.
       For the category, classify the scenes into one of the available categories.