ic

apps/riesgos/api/__init__.py

@@ -0,0 +1,89 @@
+from contextvars import ContextVar
+from types import SimpleNamespace
+
+from hvac import Client
+from pydantic import Field
+from langfuse import Langfuse
+from langfuse.decorators import langfuse_context
+from pydantic_settings import BaseSettings
+
+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)
+
+context = SimpleNamespace(
+    buffer=buffer,
+    tool_buffer=tool_buffer,
+    tool_id=tool_id,
+    tool_name=tool_name,
+)
+
+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):
+    model: str = "gpt-4o"
+    model_temperature: int = 0
+    embedding_model: str = "text-embedding-3-large"
+    message_limit: int = 10
+    storage_bucket: str = "riesgosreferences2"
+    vector_index: str = "MayaRiesgos2"
+    search_limit: int = 5
+    host: str = "0.0.0.0"
+    port: int = 8000
+
+    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"]
+    )
+    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"]
+    )
+    mongodb_url: str = Field(
+        default_factory=lambda: secret_map["cosmosdb_connection_string"]
+    )
+    langfuse_host: str | None = None
+    langfuse_public_key: str | None = None
+    langfuse_secret_key: str | None = None
+
+    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],
+        )
+
+    def init_langfuse(self):
+        langfuse_context.configure(
+            host=self.langfuse_host,
+            public_key=self.langfuse_public_key,
+            secret_key=self.langfuse_secret_key,
+        )
+
+    def get_langfuse(self):
+        return Langfuse(
+            host=self.langfuse_host,
+            public_key=self.langfuse_public_key,
+            secret_key=self.langfuse_secret_key,
+        )
+
+
+config = Settings()

apps/riesgos/api/agent.py

@@ -0,0 +1,131 @@
+from pathlib import Path
+from typing import Any, Literal
+
+from banortegpt.storage.azure_storage import AzureStorage
+from banortegpt.vector.qdrant import AsyncQdrant
+from langchain_core.messages import AIMessageChunk
+from qdrant_client import models
+from langchain_azure_ai.chat_models import AzureAIChatCompletionsModel
+from langchain_azure_ai.embeddings import AzureAIEmbeddingsModel
+from pydantic import BaseModel, Field
+
+from . import config, context
+
+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):
+    """Buscar informacion relevante de documentos de Banorte sobre un sistema en particular."""
+
+    question: str = Field(..., description="La pregunta del usuario, reescrita para ser comprensible fuera de contexto.")
+    system: str = Field(..., description="El sistema del cual se buscara informacion. Puede ser uno de los siguientes: ['ML','SACS','ED','CARATULA', 'SICRED']. 'ML' es 'Master de lineas', 'ED' es 'Expediente Digital'")
+
+class MayaRiesgos:
+    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}
+
+    def build_response(self, payloads):
+        template = "------ REFERENCIA {index} ----- \n\n{content}"
+
+        filled_templates = [
+            template.format(index=idx, content=payload["content"])
+            for idx, payload in enumerate(payloads)
+        ]
+
+        return "\n".join(filled_templates)
+
+    async def get_information(
+        self, question: str, system: Literal["ML", "SACS", "ED", "7.1", "SICRED"]
+    ):
+        embedding = await self.embedder.aembed_query(question)
+
+        conditions = models.Filter(
+            must=[
+                models.FieldCondition(
+                    key="system",
+                    match=models.MatchAny(any=["ALL", system]),
+                )
+            ]
+        )
+
+        payloads = await self.search.semantic_search(
+            collection=self.index,
+            embedding=embedding,
+            limit=self.limit,
+            conditions=conditions,
+        )
+
+        tool_response = self.build_response(payloads)
+
+        return tool_response, payloads
+
+    async def get_shareable_urls(self, metadatas: list):
+        reference_urls = []
+        image_urls = []
+
+        for metadata in metadatas:
+            if (pagina := metadata.get("pagina")) and (
+                archivo := metadata.get("archivo")
+            ):
+                image_file = f"{pagina}_{archivo}.png"
+
+                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 stream(self, history, overwrites: dict | None = None):
+        generation_config = self._generation_config_overwrite(overwrites)
+
+        async for chunk in self.llm.astream(input=history, **generation_config):
+            assert isinstance(chunk, AIMessageChunk)
+            if call := chunk.tool_call_chunks:
+                if tool_id := call[0].get("id"):
+                    context.tool_id.set(tool_id)
+                if name := call[0].get("name"):
+                    context.tool_name.set(name)
+                if args := call[0].get("args"):
+                    context.tool_buffer.set(context.tool_buffer.get() + args)
+            else:
+                if buffer := chunk.content:
+                    assert isinstance(buffer, str)
+                    context.buffer.set(context.buffer.get() + buffer)
+                    yield buffer
+
+    async def generate(self, history, overwrites: dict | None = None):
+        generation_config = self._generation_config_overwrite(overwrites)
+        return await self.llm.ainvoke(input=history, **generation_config)

apps/riesgos/api/server/__init__.py

@@ -0,0 +1,23 @@
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI
+
+from api import config
+
+from .v1 import router
+
+
+@asynccontextmanager
+async def lifespan(_: FastAPI):
+    await config.init_mongo_db()
+    config.init_langfuse()
+    yield
+
+
+app = FastAPI(
+    lifespan=lifespan,
+    docs_url="/api/docs",
+    openapi_url="/api/openapi.json",
+)
+
+app.include_router(router)

apps/riesgos/api/server/v1.py

@@ -0,0 +1,64 @@
+import uuid
+from typing import Literal
+
+from fastapi import APIRouter
+from fastapi.responses import StreamingResponse
+from pydantic import BaseModel
+
+from api import config, services
+from api.agent import MayaRiesgos
+
+router = APIRouter(prefix="/api/v1")
+agent = MayaRiesgos()
+
+
+class Message(BaseModel):
+    conversation_id: uuid.UUID
+    prompt: str
+
+
+@router.post("/conversation")
+async def create_conversation():
+    conversation_id = uuid.uuid4()
+    await services.create_conversation(conversation_id, agent.system_prompt)
+    return {"conversation_id": conversation_id}
+
+
+@router.post("/message")
+async def send(message: Message, stream: bool = False):
+    if stream is True:
+
+        def b64_sse(func):
+            async def wrapper(*args, **kwargs):
+                async for chunk in func(*args, **kwargs):
+                    content = chunk.model_dump_json()
+                    data = f"data: {content}\n\n"
+                    yield data
+
+            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")
+    else:
+        response = await services.generate(
+            agent, message.prompt, message.conversation_id
+        )
+        return response
+
+
+class Feedback(BaseModel):
+    key: str
+    rating: Literal["Good", "Bad", "None"]
+
+
+@router.post("/feedback")
+async def register_feedback(data: Feedback):
+    if data.rating:
+        langfuse = config.get_langfuse()
+        langfuse.score(
+            id=data.key + "-rating",
+            trace_id=data.key,
+            name="Rating",
+            value=data.rating,
+        )

apps/riesgos/api/services/__init__.py

@@ -0,0 +1,6 @@
+from banortegpt.database.mongo_memory.crud import create_conversation
+
+from .generate_response import generate
+from .stream_response import stream
+
+__all__ = ["stream", "generate", "create_conversation"]

apps/riesgos/api/services/generate_response.py

@@ -0,0 +1,89 @@
+import json
+from typing import Any
+from uuid import UUID
+
+from banortegpt.database.mongo_memory import crud
+from langfuse.decorators import langfuse_context, observe
+from pydantic import BaseModel
+
+from api import context as ctx
+from api.agent import MayaRiesgos
+
+
+class Response(BaseModel):
+    content: str
+    urls: list[str]
+
+
+@observe(capture_input=False, capture_output=False)
+async def generate(
+    agent: MayaRiesgos,
+    prompt: str,
+    conversation_id: UUID,
+) -> Response:
+    conversation = await crud.get_conversation(conversation_id)
+
+    if conversation is None:
+        raise ValueError(f"Conversation with id {conversation_id} not found")
+
+    conversation.add(role="user", content=prompt)
+
+    response = await agent.generate(conversation.to_openai_format(agent.message_limit))
+
+    reference_urls, image_urls = [], []
+
+    if call := response.tool_calls:
+        if id := call[0].id:
+            ctx.tool_id.set(id)
+        if name := call[0].function.name:
+            ctx.tool_name.set(name)
+        ctx.tool_buffer.set(call[0].function.arguments)
+    else:
+        ctx.buffer.set(response.content)
+
+    buffer = ctx.buffer.get()
+    tool_buffer = ctx.tool_buffer.get()
+    tool_id = ctx.tool_id.get()
+    tool_name = ctx.tool_name.get()
+
+    if tool_id is not None:
+        # Si tool_buffer es un string JSON, lo convertimos a diccionario
+        if isinstance(tool_buffer, str):
+            try:
+                tool_args = json.loads(tool_buffer)
+            except json.JSONDecodeError:
+                tool_args = {"question": tool_buffer}
+        else:
+            tool_args = tool_buffer
+
+        response, payloads = await agent.tool_map[tool_name](**tool_args)  # type: ignore
+
+        tool_call: dict[str, Any] = agent.llm.build_tool_call(
+            tool_id, tool_name, tool_buffer
+        )
+        tool_call_id: dict[str, Any] = agent.llm.build_tool_call_id(tool_id)
+
+        conversation.add("assistant", **tool_call)
+        conversation.add("tool", content=response, **tool_call_id)
+
+        response = await agent.generate(
+            conversation.to_openai_format(agent.message_limit), {"tools": None}
+        )
+        ctx.buffer.set(response.content)
+
+        reference_urls, image_urls = await agent.get_shareable_urls(payloads)  # type: ignore
+
+    buffer = ctx.buffer.get()
+    if buffer is None:
+        raise ValueError("No buffer found")
+
+    conversation.add(role="assistant", content=buffer)
+
+    langfuse_context.update_current_trace(
+        name=agent.__class__.__name__,
+        session_id=str(conversation_id),
+        input=prompt,
+        output=buffer,
+    )
+
+    return Response(content=buffer, urls=reference_urls + image_urls)

apps/riesgos/api/services/stream_response.py

@@ -0,0 +1,102 @@
+import json
+from enum import StrEnum
+from typing import TypeAlias
+from uuid import UUID
+
+from banortegpt.database.mongo_memory import crud
+from langfuse.decorators import langfuse_context, observe
+from pydantic import BaseModel
+
+from api import context as ctx
+from api.agent import MayaRiesgos
+
+
+class ChunkType(StrEnum):
+    START = "start"
+    TEXT = "text"
+    REFERENCE = "reference"
+    IMAGE = "image"
+    TOOL = "tool"
+    END = "end"
+    ERROR = "error"
+
+
+ContentType: TypeAlias = str | int
+
+
+class ResponseChunk(BaseModel):
+    type: ChunkType
+    content: ContentType | list[ContentType] | None
+
+
+@observe(capture_input=False, capture_output=False)
+async def stream(agent: MayaRiesgos, prompt: str, conversation_id: UUID):
+        yield ResponseChunk(type=ChunkType.START, content="")
+
+        conversation = await crud.get_conversation(conversation_id)
+
+        if conversation is None:
+            raise ValueError(f"Conversation with id {conversation_id} 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)
+
+        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)
+
+            response, payloads = await agent.tool_map[tool_name](**buffer_dict)
+
+            conversation.add(
+                role="assistant",
+                tool_calls=[
+                    {
+                        "id": tool_id,
+                        "function": {
+                            "name": tool_name,
+                            "arguments": tool_buffer,
+                        },
+                        "type": "function",
+                    }
+                ],
+            )
+            conversation.add(role="tool", content=response, 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)
+
+            ref_urls, image_urls = await agent.get_shareable_urls(payloads)  # type: ignore
+
+            if len(ref_urls) > 0:
+                yield ResponseChunk(type=ChunkType.REFERENCE, content=ref_urls)
+
+            if len(image_urls) > 0:
+                yield ResponseChunk(type=ChunkType.IMAGE, content=image_urls)
+
+        buffer = ctx.buffer.get()
+
+        conversation.add(role="assistant", content=buffer)
+
+        await conversation.save()
+
+        langfuse_context.update_current_trace(
+            name=agent.__class__.__name__,
+            session_id=str(conversation_id),
+            input=prompt,
+            output=buffer,
+        )
+
+        yield ResponseChunk(
+            type=ChunkType.END, content=langfuse_context.get_current_trace_id()
+        )

apps/riesgos/api/system_prompt.md

@@ -0,0 +1,19 @@
+Eres MayaRiesgos, una asistente virtual amigable y profesional de Banorte, especializada en proporcionar información precisa a los usuarios.
+
+Tu función principal:
+1. Responder preguntas de los usuarios de manera informativa, empática y concisa.
+2. Utilizar la herramienta get_information antes de responder cualquier pregunta.
+3. Antes de usar la herramienta get_information, si la pregunta no menciona explícitamente un sistema, primero aclara con el usuario sobre qué sistema está consultando.
+Los posibles sistemas son:
+    - ML: "Máster de Líneas"
+    - ED: "Expediente Digital"
+    - CARATULA
+    - SACS
+    - SICRED
+5. Si el usuario realiza una pregunta de seguimiento, puedes asumir que se refiere al mismo sistema mencionado en la consulta anterior.
+
+Pautas para interactuar:
+    - Siempre basa tu respuesta únicamente en el resultado de get_information.
+    - Si la información obtenida mediante get_information no es suficiente para responder, informa al usuario que no cuentas con datos suficientes para brindar una respuesta.
+
+Tu objetivo es ofrecer una experiencia informativa y satisfactoria, reflejando los valores de confianza y excelencia de Banorte.

apps/riesgos/api/tools.json

@@ -0,0 +1,23 @@
+[
+    {
+        "type": "function",
+        "function": {
+            "name": "get_information",
+            "description": "Buscar informacion relevante de documentos de Banorte sobre un sistema en particular.",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "question": {
+                        "type": "string",
+                        "description": "La pregunta del usuario, reescrita para ser comprensible fuera de contexto."
+                    },
+                    "system": {
+                        "type": "string",
+                        "description": "El sistema del cual se buscara informacion. Puede ser uno de los siguientes: ['ML','SACS','ED','CARATULA', 'SICRED']. 'ML' es 'Master de lineas', 'ED' es 'Expediente Digital'"
+                    }
+                },
+                "required": ["question", "system"]
+            }
+        }
+    }
+]