ic

apps/bursatil/api/agent/__init__.py

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

apps/bursatil/api/agent/main.py

@@ -0,0 +1,130 @@
+from pathlib import Path
+from typing import Any
+
+from langchain_core.messages import AIMessageChunk
+from pydantic import BaseModel, Field
+from langchain_azure_ai.chat_models import AzureAIChatCompletionsModel
+from langchain_azure_ai.embeddings import AzureAIEmbeddingsModel
+from banortegpt.storage.azure_storage import AzureStorage
+from banortegpt.vector.qdrant import AsyncQdrant
+
+from api import context
+from api.config import config
+
+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 MayaBursatil:
+    system_prompt = SYSTEM_PROMPT
+    generation_config = {
+        "temperature": config.model_temperature,
+    }
+    embedding_model = config.embedding_model
+    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, fallback):
+        template = "<FAQ {index}>\n\n{content}\n\n</FAQ {index}>"
+
+        filled_templates = [
+            template.format(index=idx, content=payload["content"])
+            for idx, payload in enumerate(payloads)
+        ]
+
+        filled_templates.append(f"<FALLBACK>\n{fallback}\n</FALLBACK>")
+
+        return "\n".join(filled_templates)
+
+    async def get_information(self, question: str):
+        embedding = await self.embedder.aembed_query(question)
+
+        payloads = await self.search.semantic_search(embedding=embedding, collection=self.index, limit=self.limit)
+
+        fallback_messages: dict[str, int] = {}
+
+        for payload in payloads:
+            fallback_message = payload.get("fallback_message", "None")
+            if fallback_message not in fallback_messages:
+                fallback_messages[fallback_message] = 1
+            else:
+                fallback_messages[fallback_message] += 1
+
+        fallback = max(fallback_messages, key=fallback_messages.get)  # type: ignore
+
+        tool_response = self.build_response(payloads, fallback)
+
+        return tool_response, payloads
+
+    async def get_shareable_urls(self, payloads: list):
+        reference_urls = []
+        image_urls = []
+
+        for payload in payloads:
+            if imagen := payload.get("imagen"):
+                image_url = await self.storage.get_file_url(
+                    filename=imagen,
+                    bucket=self.bucket,
+                    minute_duration=20,
+                    image=True,
+                )
+
+                if image_url:
+                    image_urls.append(image_url)
+                else:
+                    print("Image not found")
+
+        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 delta in self.llm.astream(input=history, **generation_config):
+            assert isinstance(delta, AIMessageChunk)
+            if call := delta.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 := delta.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/bursatil/api/agent/system_prompt.md

@@ -0,0 +1,6 @@
+Eres MayaBursatil, una muy amigable y símpatica asistente virtual del departamento de contraloria bursatil de Banorte.
+Tu objetivo es responder preguntas de usuarios de manera informativa y empatica.
+Para cada pregunta, utiliza la herramienta 'get_information' para obtener informacion de nuestro FAQ.
+Utiliza la informacion para responder la pregunta del usuario.
+Utiliza emojis.
+Si no puedes responder la pregunta basado en la informacion del FAQ, responde con el contenido en el FALLBACK.

apps/bursatil/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
+    model: str = "gpt-4o"
+    model_temperature: int = 0
+    embedding_model: str = "text-embedding-3-large"
+    message_limit: int = 10
+    storage_bucket: str = "bursatilreferences"
+    vector_index: str = "MayaBursatil"
+    search_limit: int = 3
+    host: str = "0.0.0.0"
+    port: int = 8000
+
+    # API Keys
+    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"]
+    )
+
+    async def init_mongo_db(self):
+        from beanie import init_beanie
+        from motor.motor_asyncio import AsyncIOMotorClient
+
+        from banortegpt.database.mongo_memory.models import Conversation
+
+        await init_beanie(
+            database=AsyncIOMotorClient(self.mongodb_url).banortegptdos,
+            document_models=[Conversation],
+        )
+
+
+config = Settings()

apps/bursatil/api/context.py

@@ -0,0 +1,7 @@
+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/bursatil/api/server.py

@@ -0,0 +1,55 @@
+import uuid
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI
+from fastapi.responses import StreamingResponse
+from pydantic import BaseModel
+
+from api import services
+from api.agent import MayaBursatil
+from api.config import config
+
+
+@asynccontextmanager
+async def lifespan(_: FastAPI):
+    await config.init_mongo_db()
+    yield
+
+
+app = FastAPI(lifespan=lifespan)
+agent = MayaBursatil()
+
+
+class Message(BaseModel):
+    conversation_id: uuid.UUID
+    prompt: str
+
+
+@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}
+
+
+@app.post("/api/v1/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

apps/bursatil/api/services/__init__.py

@@ -0,0 +1,10 @@
+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/bursatil/api/services/generate_response.py

@@ -0,0 +1,89 @@
+import json
+from typing import Any
+from uuid import UUID
+
+from langfuse.decorators import langfuse_context, observe
+from pydantic import BaseModel
+
+from api import context as ctx
+from api.agent import MayaBursatil
+from banortegpt.database.mongo_memory import crud
+
+
+class Response(BaseModel):
+    content: str
+    urls: list[str]
+
+
+@observe(capture_input=False, capture_output=False)
+async def generate(
+    agent: MayaBursatil,
+    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=str(conversation_id),
+        session_id=str(conversation_id),
+        input=prompt,
+        output=buffer,
+    )
+
+    return Response(content=buffer, urls=reference_urls + image_urls)

apps/bursatil/api/services/stream_response.py

@@ -0,0 +1,100 @@
+import json
+from enum import StrEnum
+from typing import TypeAlias
+from uuid import UUID
+
+from langfuse.decorators import langfuse_context, observe
+from pydantic import BaseModel
+
+from api import context as ctx
+from api.agent import MayaBursatil
+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
+
+
+class ResponseChunk(BaseModel):
+    type: ChunkType
+    content: ContentType | list[ContentType] | None
+
+
+@observe(capture_input=False, capture_output=False)
+async def stream(agent: MayaBursatil, 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)
+
+    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.replace()
+
+    yield ResponseChunk(type=ChunkType.END, content="")
+
+    langfuse_context.update_current_trace(
+        name=agent.__class__.__name__,
+        session_id=str(conversation_id),
+        input=prompt,
+        output=buffer,
+    )