ic

apps/normativa/api/agent/__init__.py

@@ -0,0 +1,3 @@
+from .main import Agent
+
+__all__ = ["Agent"]

apps/normativa/api/agent/main.py

@@ -0,0 +1,373 @@
+import logging
+from pathlib import Path
+from typing import Annotated, List, Sequence, Literal, Any, AsyncGenerator
+from typing_extensions import TypedDict
+
+from langchain_core.messages import BaseMessage, HumanMessage, AIMessage, AIMessageChunk
+from langchain_azure_ai.chat_models import AzureAIChatCompletionsModel
+from langchain_azure_ai.embeddings import AzureAIEmbeddingsModel
+from langgraph.graph.message import add_messages
+from langgraph.graph import StateGraph, START, END
+from langgraph.checkpoint.memory import MemorySaver
+from pydantic import BaseModel, Field
+
+from banortegpt.storage.azure_storage import AzureStorage
+from banortegpt.vector.qdrant import AsyncQdrant
+
+import api.context as ctx
+from api.config import config
+
+logger = logging.getLogger(__name__)
+
+parent = Path(__file__).parent
+SYSTEM_PROMPT = (parent / "system_prompt.md").read_text()
+AZURE_AI_URI = "https://eastus2.api.cognitive.microsoft.com"
+
+
+class get_information(BaseModel):
+    """Search a private repository for information."""
+    question: str = Field(..., description="The user question")
+
+class MayaNormativaState(TypedDict):
+    messages: Annotated[Sequence[BaseMessage], add_messages]
+    query: str
+    search_results: List[dict]
+    iteration_count: int
+    max_iterations: int
+    final_response: str
+
+
+class MayaNormativa:
+    system_prompt = SYSTEM_PROMPT
+    generation_config = {
+        "temperature": config.model_temperature,
+    }
+    message_limit = config.message_limit
+    index = config.vector_index
+    limit = config.search_limit
+    bucket = config.storage_bucket
+
+    search = AsyncQdrant.from_config(config)
+    llm = AzureAIChatCompletionsModel(
+        endpoint=f"{AZURE_AI_URI}/openai/deployments/{config.model}",
+        credential=config.openai_api_key,
+    ).bind_tools([get_information])
+    embedder = AzureAIEmbeddingsModel(
+        endpoint=f"{AZURE_AI_URI}/openai/deployments/{config.embedding_model}",
+        credential=config.openai_api_key,
+    )
+    storage = AzureStorage.from_config(config)
+
+    def __init__(self) -> None:
+        self.tool_map = {"get_information": self.get_information}
+        self.memory = MemorySaver()
+        self.graph = self._build_complete_langgraph()
+
+    def build_response(self, payloads):
+        """Mejorado con más info que OCP original"""
+        preface = ["Recuerda citar las referencias en el formato: texto[1]."]
+        template = "------ REFERENCIA {index} ----- \n\n{content}\n\n**Fuente:** {source_info}"
+
+        filled_templates = []
+        for idx, payload in enumerate(payloads):
+            content = payload.get("content", "") or payload.get("page_content", "")
+            metadata = payload.get("metadata", {})
+            
+            source_info = ""
+            if metadata:
+                file_name = metadata.get("file_name", "") or metadata.get("file", "")
+                page = metadata.get("page", "")
+                
+                if file_name and page:
+                    source_info = f"{file_name} - Página {page}"
+                elif file_name:
+                    source_info = file_name
+                else:
+                    source_info = "Documento interno"
+            
+            if not source_info:
+                source_info = "No disponible"
+            
+            filled_template = template.format(
+                index=idx + 1,  
+                content=content,
+                source_info=source_info
+            )
+            filled_templates.append(filled_template)
+
+        return "\n".join(preface + filled_templates)
+
+    async def get_information(self, question: str):
+        logger.info(f"Embedding question: {question} with model {self.embedder.model_name}")
+        embedding = await self.embedder.aembed_query(question)
+
+        results = await self.search.semantic_search(
+            embedding=embedding, collection=self.index, limit=self.limit
+        )
+
+        tool_response = self.build_response(results)
+        return tool_response, results
+
+    async def get_shareable_urls(self, metadatas: list):
+        reference_urls = []
+        image_urls = []
+
+        for metadata in metadatas:
+            if file := metadata.get("file"):
+                reference_url = await self.storage.get_file_url(
+                    filename=file,
+                    bucket=self.bucket,
+                    minute_duration=20,
+                    image=False,
+                )
+                reference_urls.append(reference_url)
+            if image_file := metadata.get("image"):
+                image_url = await self.storage.get_file_url(
+                    filename=image_file,
+                    bucket=self.bucket,
+                    minute_duration=20,
+                    image=True,
+                )
+                image_urls.append(image_url)
+
+        return reference_urls, image_urls
+
+    def _generation_config_overwrite(self, overwrites: dict | None) -> dict[str, Any]:
+        generation_config_copy = self.generation_config.copy()
+        if overwrites:
+            for k, v in overwrites.items():
+                generation_config_copy[k] = v
+        return generation_config_copy
+
+    async def retrieve_node(self, state: MayaNormativaState) -> dict:
+        query = state["query"]
+        logger.info(f"Retrieving information for: {query}")
+        
+        try:
+            _, results = await self.get_information(query)
+            logger.info(f"Retrieved {len(results)} results")
+            return {
+                "search_results": results,
+                "iteration_count": state["iteration_count"] + 1
+            }
+        except Exception as e:
+            logger.error(f"Error in retrieve_node: {e}")
+            return {
+                "search_results": [],
+                "iteration_count": state["iteration_count"] + 1
+            }
+
+    async def evaluate_node(self, state: MayaNormativaState) -> dict:
+        results = state["search_results"]
+        iteration = state["iteration_count"]
+        max_iter = state["max_iterations"]
+        
+        has_sufficient_results = len(results) >= 2
+        reached_max_iterations = iteration >= max_iter
+        
+        if has_sufficient_results or reached_max_iterations:
+            logger.info(f"Stopping search: {len(results)} results, iteration {iteration}")
+            return {"continue_search": False}
+        else:
+            original_query = state["query"]
+            new_query = f"circular artículo {original_query}"
+            logger.info(f"Continuing search with modified query: {new_query}")
+            return {
+                "continue_search": True,
+                "query": new_query
+            }
+
+    async def generate_node(self, state: MayaNormativaState) -> dict:  
+        results = state["search_results"]
+        query = state["query"]
+        messages = state.get("messages", [])
+        
+        logger.info(f"Generating response for query: {query}")
+        logger.info(f"Using {len(results)} search results")
+        logger.info(f"Message history length: {len(messages)}")
+        
+        if not results:
+            final_response = "No encontré información sobre este tema en la documentación actual."
+        else:
+            context_text = self.build_response(results)
+            
+            try:
+                history = [
+                    {"role": "system", "content": self.system_prompt}
+                ]
+                
+                for msg in messages[:-1]:
+                    if isinstance(msg, HumanMessage):
+                        history.append({"role": "user", "content": msg.content})
+                    elif isinstance(msg, AIMessage):
+                        history.append({"role": "assistant", "content": msg.content})
+          
+                current_prompt = f"""
+                Consulta del usuario: {query}
+
+                Información encontrada:
+                {context_text}
+
+                INSTRUCCIONES:
+                - Reproduce la información EXACTAMENTE como aparece en la documentación
+                - NO parafrasees ni interpretes
+                - Usa las palabras exactas del documento original
+                - Mantén los tiempos verbales originales
+                - Mejora el formato con emojis 
+                - Respuestas extensas y completas
+                - Siempre has referencia al articulo, ley o seccion de la pagina donde encontraste la informacion 
+                - Pregunta por informacion relacionada con la respuesta que requiera al final
+                - Considera el contexto de la infomacion anterior si existe
+                """
+                
+                history.append({"role": "user", "content": current_prompt})
+                
+                generation_config = self._generation_config_overwrite(None)
+                
+                response_chunks = []
+                async for delta in self.llm.astream(input=history, **generation_config):
+                    assert isinstance(delta, AIMessageChunk)
+                    if delta.content:
+                        response_chunks.append(delta.content)
+                
+                final_response = "".join(response_chunks)
+                logger.info(f"Generated response length: {len(final_response)}")
+                  
+            except Exception as e:
+                logger.error(f"ERROR generando respuesta: {e}")
+                final_response = f"Error generando respuesta: {str(e)}"
+            
+        return {
+            "final_response": final_response,
+            "messages": [AIMessage(content=final_response)]
+        }
+
+    def _build_complete_langgraph(self) -> StateGraph:
+        workflow = StateGraph(MayaNormativaState)
+    
+        workflow.add_node("retrieve", self.retrieve_node)
+        workflow.add_node("evaluate", self.evaluate_node)
+        workflow.add_node("generate", self.generate_node)
+        
+        workflow.add_edge(START, "retrieve")
+        workflow.add_edge("retrieve", "evaluate")
+        
+        workflow.add_conditional_edges(
+            "evaluate",
+            self._decide_next_step,
+            {
+                "continue": "retrieve",  
+                "finish": "generate"     
+            }
+        )
+        
+        workflow.add_edge("generate", END)
+        
+        return workflow.compile(checkpointer=self.memory)
+    
+    def _decide_next_step(self, state: MayaNormativaState) -> Literal["continue", "finish"]:
+        if state.get("continue_search", False):
+            return "continue"
+        else:
+            return "finish"
+
+    async def stream(self, history, overwrites: dict | None = None, thread_id: str = "default"):
+        """Stream simplificado que mantiene memoria"""
+        
+        last_message = history[-1] if history else {"content": ""}
+        query = last_message.get("content", "")
+        
+        if not query:
+            yield "Error: No se encontró pregunta en el historial"
+            return
+        
+        logger.info(f"Processing query: {query}")
+        logger.info(f"Thread ID: {thread_id}")
+        
+        try:
+            config_with_thread = {
+                "configurable": {"thread_id": thread_id}
+            }
+
+            initial_state = {
+                "messages": [HumanMessage(content=query)],
+                "query": query,
+                "search_results": [],
+                "iteration_count": 0,
+                "max_iterations": 2,
+                "final_response": ""
+            }
+            
+            logger.info("Invoking LangGraph...")
+            
+            final_state = await self.graph.ainvoke(initial_state, config=config_with_thread)
+            
+            logger.info("LangGraph execution completed")
+            
+            self.last_search_results = final_state.get("search_results", [])
+            
+            # Extraer metadatos
+            if self.last_search_results:
+                try:
+                    metadatas = []
+                    for result in self.last_search_results:
+                        metadata = result.get("metadata", {})
+                        if metadata:
+                            metadatas.append(metadata)
+                    
+                    self.last_metadatas = metadatas
+                    logger.info(f"Extracted {len(metadatas)} metadata objects")
+                    
+                except Exception as e:
+                    logger.error(f"Error extrayendo metadatos: {e}")
+                    self.last_metadatas = []
+            else:
+                self.last_metadatas = []
+            
+            final_response = final_state.get("final_response", "Error: No se pudo generar respuesta")
+            
+            chunk_size = 50
+            for i in range(0, len(final_response), chunk_size):
+                chunk = final_response[i:i + chunk_size]
+                ctx.buffer.set(ctx.buffer.get() + chunk)
+                yield chunk
+                
+        except Exception as e:
+            error_msg = f"Error en stream: {str(e)}"
+            logger.error(error_msg)
+            yield error_msg
+
+    async def get_conversation_history(self, thread_id: str = "default") -> List[BaseMessage]:
+        try:
+            config_with_thread = {
+                "configurable": {"thread_id": thread_id}
+            }
+            
+            checkpoint = await self.graph.aget_state(config=config_with_thread)
+            
+            if checkpoint and checkpoint.values:
+                return checkpoint.values.get("messages", [])
+            else:
+                return []
+                
+        except Exception as e:
+            logger.error(f"Error obteniendo historial: {e}")
+            return []
+    
+    async def debug_memory(self, thread_id: str = "default"):
+
+        try:
+            history = await self.get_conversation_history(thread_id)
+            logger.info(f"MEMORY DEBUG (thread: {thread_id}) ===")
+            logger.info(f"Total messages: {len(history)}")
+            
+            for i, msg in enumerate(history):
+                msg_type = "USER" if isinstance(msg, HumanMessage) else "ASSISTANT"
+                content_preview = msg.content[:50] + "..." if len(msg.content) > 50 else msg.content
+                logger.info(f"{i+1}. {msg_type}: {content_preview}")
+                
+        except Exception as e:
+            logger.error(f"Error in debug_memory: {e}")
+
+    
+Agent = MayaNormativa

apps/normativa/api/agent/system_prompt.md

@@ -0,0 +1,23 @@
+Eres MayaNormativa, asistente virtual especializada en normativa y procesos internos de Banorte.
+Reglas Fundamentales
+OBLIGATORIO:
+
+Usar únicamente get_information para cada consulta
+Siempre has referencia al articulo, ley o seccion de la pagina donde encontraste la informacion 
+Reproducir información EXACTAMENTE como aparece en documentación
+Usar Markdown para respuestas claras y concisas
+
+PROHIBIDO:
+
+Agregar, interpretar, suponer o extrapolar información
+Modificar términos técnicos, números, fechas o procedimientos
+Parafrasear si altera el significado original
+Inventar o completar información faltante
+
+Respuestas Estándar
+Información no disponible: "No encontré información sobre [tema] en la documentación actual"
+Información parcial: "Esta es toda la información disponible en la documentación interna"
+Información contradictoria: Presenta ambas versiones indicando fuentes
+Pie de página: Información obtenida de: [referencia/documento específico]
+Objetivo
+Proporcionar información oficial precisa y confiable. Es mejor reconocer limitaciones que dar información inexacta.

apps/normativa/api/config.py

@@ -0,0 +1,55 @@
+from hvac import Client
+from pydantic import Field
+from pydantic_settings import BaseSettings
+
+client = Client(url="https://vault.ia-innovacion.work")
+
+if not client.is_authenticated():
+    raise Exception("Vault authentication failed")
+
+secret_map = client.secrets.kv.v2.read_secret_version(
+    path="banortegpt", mount_point="secret"
+)["data"]["data"]
+
+class Settings(BaseSettings):
+    # Config básico
+    model: str = "gpt-4o"
+    model_temperature: int = 0
+    message_limit: int = 10
+    host: str = "0.0.0.0"
+    port: int = 8000
+
+    # AGREGAR ESTAS LÍNEAS (igual que OCP):
+    embedding_model: str = "text-embedding-3-large"
+    storage_bucket: str = "normativa-bucket"  # Ajusta el nombre
+    vector_index: str = "MayaNormativaLLM"
+    search_limit: int = 3
+
+    # API Keys existentes
+    azure_endpoint: str = Field(default_factory=lambda: secret_map["azure_endpoint"])
+    openai_api_key: str = Field(default_factory=lambda: secret_map["openai_api_key"])
+    openai_api_version: str = Field(default_factory=lambda: secret_map["openai_api_version"])
+    mongodb_url: str = Field(default_factory=lambda: secret_map["cosmosdb_connection_string"])
+    
+    # AGREGAR ESTAS LÍNEAS (igual que OCP):
+    azure_blob_connection_string: str = Field(
+        default_factory=lambda: secret_map["azure_blob_connection_string"]
+    )
+    qdrant_url: str = Field(default_factory=lambda: secret_map["qdrant_api_url"])
+    qdrant_api_key: str | None = Field(
+        default_factory=lambda: secret_map["qdrant_api_key"]
+    )
+
+    async def init_mongo_db(self):
+        from banortegpt.database.mongo_memory.models import Conversation
+        from beanie import init_beanie
+        from motor.motor_asyncio import AsyncIOMotorClient
+
+        client = AsyncIOMotorClient(self.mongodb_url)
+
+        await init_beanie(
+            database=client.banortegptdos,
+            document_models=[Conversation],
+        )
+
+config = Settings()

apps/normativa/api/context.py

@@ -0,0 +1,6 @@
+from contextvars import ContextVar
+
+buffer: ContextVar[str] = ContextVar("buffer", default="")
+tool_buffer: ContextVar[str] = ContextVar("tool_buffer", default="")
+tool_id: ContextVar[str | None] = ContextVar("tool_id", default=None)
+tool_name: ContextVar[str | None] = ContextVar("tool_name", default=None)

apps/normativa/api/server.py

@@ -0,0 +1,300 @@
+import uuid
+import os
+from contextlib import asynccontextmanager
+from pathlib import Path
+
+from fastapi import FastAPI, HTTPException
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import StreamingResponse, FileResponse, RedirectResponse
+from pydantic import BaseModel
+from langfuse import Langfuse
+
+from dotenv import load_dotenv  # ← Agregar este import
+
+from api import services
+from api.agent import Agent
+from api.config import config
+
+# Cargar variables de entorno
+load_dotenv()
+
+# Configurar Langfuse desde variables de entorno
+langfuse = Langfuse(
+    public_key=os.getenv("LANGFUSE_PUBLIC_KEY"),
+    secret_key=os.getenv("LANGFUSE_SECRET_KEY"), 
+    host=os.getenv("LANGFUSE_HOST")
+)
+
+
+# Mapeo completo de archivos a URLs públicas
+PDF_PUBLIC_URLS = {
+    # Disposiciones de CNBV
+    "Disposiciones de carácter general aplicables a las casas de bolsa.pdf": "https://www.cnbv.gob.mx/Normatividad/Disposiciones%20de%20car%C3%A1cter%20general%20aplicables%20a%20las%20casas%20de%20bolsa.pdf",
+    "Disposiciones de carácter general aplicables a las instituciones de crédito.pdf": "https://www.cnbv.gob.mx/Normatividad/Disposiciones%20de%20car%C3%A1cter%20general%20aplicables%20a%20las%20instituciones%20de%20cr%C3%A9dito.pdf",
+    "Disposiciones de carácter general aplicables a las sociedades controladoras de grupos financieros y subcontroladoras que regulan las materias que corresponden de manera conjunta a las Comisio.pdf": "https://www.cnbv.gob.mx/Normatividad/Disposiciones%20de%20car%C3%A1cter%20general%20aplicables%20a%20las%20sociedades%20controladoras%20de%20grupos%20financieros%20y%20subcontroladoras%20que%20regulan%20las%20materias%20que%20corresponden%20de%20manera%20conjunta%20a%20las%20Comisiones%20Nacionales%20Supervisoras.pdf",
+    "Disposiciones de carácter general aplicables a los fondos de inversión y a las personas que les prestan servicios.pdf": "https://www.cnbv.gob.mx/Normatividad/Disposiciones%20de%20car%C3%A1cter%20general%20aplicables%20a%20los%20fondos%20de%20inversi%C3%B3n%20y%20a%20las%20personas%20que%20les%20prestan%20servicios.pdf",
+    "Ley para la Transparencia y Ordenamiento de los Servicios Financieros.pdf": "https://www.cnbv.gob.mx/Normatividad/Ley%20para%20la%20Transparencia%20y%20Ordenamiento%20de%20los%20Servicios%20Financieros.pdf",
+    
+    # Circulares CNBV adicionales
+    "circular_servicios_de_inversion.pdf": "https://www.cnbv.gob.mx/Normatividad/Disposiciones%20de%20car%C3%A1cter%20general%20aplicables%20a%20las%20entidades%20financieras%20y%20dem%C3%A1s%20personas%20que%20proporcionen%20servicios%20de.pdf",
+    "circular_unica_de_auditores_externos.pdf": "https://www.cnbv.gob.mx/Normatividad/Disposiciones%20de%20car%C3%A1cter%20general%20que%20establecen%20los%20requisitos%20que%20deber%C3%A1n%20cumplir%20los%20auditores%20y%20otros%20profesionales%20que.pdf",
+    "ley_de_instituciones_de_Credito.pdf": "https://www.cnbv.gob.mx/Normatividad/Ley%20de%20Instituciones%20de%20Cr%C3%A9dito.pdf",
+    
+    # Circulares de Banxico
+    "circular_13_2007.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-13-2007/cobro-intereses-por-adelantad.html",
+    "circular_13_2011.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-13-2011/%7BBA4CBC28-A468-16C9-6F17-9EA9D7B03318%7D.pdf",
+    "circular_14_2007.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-14-2007/%7BFB726B6B-D523-56F5-F9B1-BE5B3B95A504%7D.pdf",
+    "circular_17_2014.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-17-2014/%7BF36CEF03-9441-2DBE-082C-0DF274903782%7D.pdf",
+    "circular_1_2005.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-1-2005/%7B5CA4BA75-FEA8-199C-F129-E8E6A73E84F3%7D.pdf",
+    "circular_21_2009.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-21-2009/%7B29285862-EDE0-567A-BAFB-D261406641A3%7D.pdf",
+    "circular_22_2008.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-22-2008/%7BF15C8A26-C92E-BE2B-9344-51EDAA3C9B68%7D.pdf",
+    "circular_22_2010.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-22-2010/%7B0D531F59-1001-4D67-D7B4-D5854DD07A58%7D.pdf",
+    "circular_27_2008.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-27-2008/%7BBC4333FE-070F-E727-199E-CA6BCF2CBA66%7D.pdf",
+    "circular_34_2010.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-34-2010/%7B0C55B906-6DB4-6B88-FED0-67987E9FB3CC%7D.pdf",
+    "circular_35_2010.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-35-2010/%7B74C5641C-ED98-53C7-F08B-A3C7BAE0D480%7D.pdf",
+    "circular_36_2010.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-36-2010/%7B26C55DE6-CC3A-3368-34FC-1A6C50B11130%7D.pdf",
+    "circular_3_2012.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-3-2012/%7B4E0281A4-7AD8-1462-BC79-7F2925F3171D%7D.pdf",
+    "circular_4_2012.pdf": "https://www.banxico.org.mx/marco-normativo/normativa-emitida-por-el-banco-de-mexico/circular-4-2012/%7B97C62974-1C94-19AE-AB5A-D0D949A36247%7D.pdf",
+    
+    # CONDUSEF
+    "circular_unica_de_condusef.pdf": "https://www.condusef.gob.mx/documentos/marco_legal/disposiciones-transparencia-if-sofom.pdf",
+    "ley_para_regular_las_sociedades_de_informacion_crediticia.pdf": "https://www.condusef.gob.mx/documentos/marco_legal/disposiciones-transparencia-if-sofom.pdf",
+    
+    # Leyes federales
+    "ley_federal_de_proteccion_de_datos_personales_en_posesion_de_los_particulares.pdf": "https://www.diputados.gob.mx/LeyesBiblio/pdf/LFPDPPP.pdf",
+    "reglamento_de_la_ley_federal_de_proteccion_de_datos_personales_en_posesion_de_los_particulares.pdf": "https://www.diputados.gob.mx/LeyesBiblio/regley/Reg_LFPDPPP.pdf",
+    
+    # SharePoint Banorte
+    "Modificaciones Recursos Procedencia Ilícita jul 25 PLD.pdf": "https://gfbanorte.sharepoint.com/:w:/r/sites/Formatosyplantillas/Documentos%20compartidos/Otros/Modificaciones%20Recursos%20Procedencia%20Il%C3%ADcita%20jul%2025%20PLD.docx?d=w6a941e9e2c26403ea41c12de35536516&csf=1&web=1&e=EHtc9b",
+}
+
+
+@asynccontextmanager
+async def lifespan(_: FastAPI):
+    await config.init_mongo_db()
+    yield
+
+
+app = FastAPI(lifespan=lifespan)
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"], 
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+    expose_headers=["*"]
+)
+
+agent = Agent()
+
+PDF_FOLDER = Path(__file__).parent / "agent" / "pdf"
+PDF_FOLDER.mkdir(parents=True, exist_ok=True) 
+
+
+@app.post("/api/v1/conversation")
+async def create_conversation():
+    conversation_id = uuid.uuid4()
+    await services.create_conversation(conversation_id, agent.system_prompt)
+    return {"conversation_id": conversation_id}
+
+
+class Message(BaseModel):
+    conversation_id: uuid.UUID
+    prompt: str
+
+
+@app.post("/api/v1/message")
+async def send(message: Message):
+    # Tracking básico del chat
+    trace = langfuse.trace(
+        name="rag_chat",
+        session_id=str(message.conversation_id),
+        input={"prompt": message.prompt}
+    )
+    
+    def b64_sse(func):
+        async def wrapper(*args, **kwargs):
+            response_parts = []
+            
+            async for chunk in func(*args, **kwargs):
+                if chunk.type == "text" and chunk.content:
+                    response_parts.append(str(chunk.content))
+                
+                content = chunk.model_dump_json()
+                data = f"data: {content}\n\n"
+                yield data
+            
+            # Solo registrar input y output
+            full_response = "".join(response_parts)
+            trace.update(output={"response": full_response})
+
+        return wrapper
+
+    sse_stream = b64_sse(services.stream)
+    generator = sse_stream(agent, message.prompt, message.conversation_id)
+    return StreamingResponse(generator, media_type="text/event-stream")
+
+
+@app.get("/api/pdf/{filename}")
+async def get_pdf(filename: str):
+    print(f"🔍 Solicitud PDF para: {filename}")
+    
+    if not filename.lower().endswith('.pdf'):
+        print(f"❌ Archivo no es PDF: {filename}")
+        raise HTTPException(status_code=400, detail="El archivo debe ser un PDF")
+    
+    if '..' in filename or ('/' in filename and not filename.startswith('http')) or '\\' in filename:
+        print(f"❌ Nombre de archivo inválido: {filename}")
+        raise HTTPException(status_code=400, detail="Nombre de archivo inválido")
+    
+    public_url = PDF_PUBLIC_URLS.get(filename)
+    
+    if public_url:
+        print(f"✅ Redirigiendo a URL pública: {public_url}")
+        return RedirectResponse(
+            url=public_url,
+            status_code=302,
+            headers={
+                "Cache-Control": "public, max-age=3600",
+                "Access-Control-Allow-Origin": "*",
+                "Access-Control-Allow-Methods": "GET, OPTIONS",
+                "Access-Control-Allow-Headers": "*"
+            }
+        )
+    
+    pdf_path = PDF_FOLDER / filename
+    
+    if not pdf_path.exists():
+        print(f"❌ PDF no encontrado: {pdf_path}")
+        raise HTTPException(status_code=404, detail=f"PDF no encontrado. Archivo: {filename}")
+    
+    if not pdf_path.is_file():
+        print(f"❌ No es un archivo: {pdf_path}")
+        raise HTTPException(status_code=404, detail="El recurso no es un archivo")
+    
+    file_size = pdf_path.stat().st_size
+    print(f"📄 Sirviendo archivo local: {filename} ({file_size} bytes)")
+    
+    if file_size == 0:
+        print(f"❌ Archivo vacío: {pdf_path}")
+        raise HTTPException(status_code=500, detail="El archivo PDF está vacío")
+    
+    return FileResponse(
+        path=str(pdf_path),
+        media_type="application/pdf",
+        filename=filename,
+        headers={
+            "Content-Disposition": f"inline; filename={filename}", 
+            "Content-Type": "application/pdf",
+            "Cache-Control": "public, max-age=3600",
+            "X-Frame-Options": "ALLOWALL",  
+            "X-Content-Type-Options": "nosniff",
+            "Access-Control-Allow-Origin": "*"
+        }
+    )
+
+
+@app.get("/api/pdfs")
+async def list_pdfs():
+    try:
+        pdf_files = []
+        
+        for filename, url in PDF_PUBLIC_URLS.items():
+            pdf_files.append({
+                "filename": filename,
+                "size": "N/A (Público)",
+                "url": f"/api/pdf/{filename}",
+                "public_url": url,
+                "type": "public"
+            })
+        
+        local_files = []
+        for pattern in ["*.pdf", "*.PDF"]:
+            for file_path in PDF_FOLDER.glob(pattern):
+                if file_path.is_file() and file_path.name not in PDF_PUBLIC_URLS:
+                    local_files.append({
+                        "filename": file_path.name,
+                        "size": file_path.stat().st_size,
+                        "url": f"/api/pdf/{file_path.name}",
+                        "type": "local"
+                    })
+        
+        pdf_files.extend(local_files)
+        
+        debug_info = {
+            "current_working_directory": str(Path.cwd()),
+            "pdf_folder_path": str(PDF_FOLDER.absolute()),
+            "pdf_folder_exists": PDF_FOLDER.exists(),
+            "public_urls_count": len(PDF_PUBLIC_URLS),
+            "local_files_count": len(local_files),
+            "public_files": list(PDF_PUBLIC_URLS.keys()),
+        }
+        
+        return {
+            "pdfs": pdf_files,
+            "debug": debug_info,
+            "total_pdfs": len(pdf_files)
+        }
+    except Exception as e:
+        import traceback
+        return {
+            "error": str(e),
+            "traceback": traceback.format_exc(),
+            "debug": {
+                "current_working_directory": str(Path.cwd()),
+                "script_file_path": __file__ if '__file__' in globals() else "unknown"
+            }
+        }
+
+
+@app.get("/api/pdf/{filename}/info")
+async def get_pdf_info(filename: str):
+    if not filename.lower().endswith('.pdf'):
+        raise HTTPException(status_code=400, detail="El archivo debe ser un PDF")
+    
+    if '..' in filename or '/' in filename or '\\' in filename:
+        raise HTTPException(status_code=400, detail="Nombre de archivo inválido")
+    
+    public_url = PDF_PUBLIC_URLS.get(filename)
+    if public_url:
+        return {
+            "filename": filename,
+            "size": "N/A",
+            "size_mb": "N/A", 
+            "modified": "N/A",
+            "url": f"/api/pdf/{filename}",
+            "public_url": public_url,
+            "type": "public"
+        }
+    
+    pdf_path = PDF_FOLDER / filename
+    
+    if not pdf_path.exists():
+        raise HTTPException(status_code=404, detail="PDF no encontrado")
+    
+    if not pdf_path.is_file():
+        raise HTTPException(status_code=404, detail="El recurso no es un archivo")
+    
+    try:
+        file_stat = pdf_path.stat()
+        return {
+            "filename": filename,
+            "size": file_stat.st_size,
+            "size_mb": round(file_stat.st_size / (1024 * 1024), 2),
+            "modified": file_stat.st_mtime,
+            "url": f"/api/pdf/{filename}",
+            "type": "local"
+        }
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"Error al obtener información del PDF: {str(e)}")
+
+
+@app.get("/api/health")
+async def health_check():
+    return {
+        "status": "healthy",
+        "pdf_folder": str(PDF_FOLDER),
+        "pdf_folder_exists": PDF_FOLDER.exists(),
+        "public_urls_configured": len(PDF_PUBLIC_URLS)
+    }

apps/normativa/api/services/__init__.py

@@ -0,0 +1,8 @@
+from banortegpt.database.mongo_memory.crud import create_conversation
+
+from .stream_response import stream
+
+__all__ = [
+    "stream",
+    "create_conversation",
+]

apps/normativa/api/services/stream_response.py

@@ -0,0 +1,89 @@
+import json
+from enum import StrEnum
+from typing import TypeAlias, Any
+from uuid import UUID
+
+from pydantic import BaseModel
+
+import api.context as ctx
+from api.agent import Agent
+from banortegpt.database.mongo_memory import crud
+
+
+class ChunkType(StrEnum):
+    START = "start"
+    TEXT = "text"
+    REFERENCE = "reference"
+    IMAGE = "image"
+    TOOL = "tool"
+    END = "end"
+    ERROR = "error"
+
+
+ContentType: TypeAlias = str | int | dict | Any
+
+
+class ResponseChunk(BaseModel):
+    type: ChunkType
+    content: ContentType | list[ContentType] | None
+
+
+async def stream(agent: Agent, prompt: str, conversation_id: UUID):
+    yield ResponseChunk(type=ChunkType.START, content="")
+
+    conversation = await crud.get_conversation(conversation_id)
+
+    if conversation is None:
+        raise ValueError("Conversation not found")
+
+    conversation.add(role="user", content=prompt)
+
+    history = conversation.to_openai_format(agent.message_limit, langchain_compat=True)
+    
+    async for content in agent.stream(history):
+        yield ResponseChunk(type=ChunkType.TEXT, content=content)
+
+
+    metadatas = getattr(agent, 'last_metadatas', [])
+    
+
+    if metadatas:
+        yield ResponseChunk(type=ChunkType.REFERENCE, content=metadatas)
+
+    if (tool_id := ctx.tool_id.get()) is not None:
+        tool_buffer = ctx.tool_buffer.get()
+        assert tool_buffer is not None
+
+        tool_name = ctx.tool_name.get()
+        assert tool_name is not None
+
+        yield ResponseChunk(type=ChunkType.TOOL, content=None)
+
+        buffer_dict = json.loads(tool_buffer)
+
+        result = await agent.tool_map[tool_name](**buffer_dict)
+
+        conversation.add(
+            role="assistant",
+            tool_calls=[
+                {
+                    "id": tool_id,
+                    "type": "function",
+                    "function": {
+                        "name": tool_name,
+                        "arguments": tool_buffer,
+                    },
+                }
+            ],
+        )
+        conversation.add(role="tool", content=result, tool_call_id=tool_id)
+
+        history = conversation.to_openai_format(agent.message_limit, langchain_compat=True)
+        async for content in agent.stream(history, {"tools": None}):
+            yield ResponseChunk(type=ChunkType.TEXT, content=content)
+
+    conversation.add(role="assistant", content=ctx.buffer.get())
+
+    await conversation.replace()
+
+    yield ResponseChunk(type=ChunkType.END, content="")