ic

.templates/rag/{{project_name}}/api/agent/__init__.py

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

.templates/rag/{{project_name}}/api/agent/main.py

@@ -0,0 +1,67 @@
+from pathlib import Path
+from typing import Any, AsyncGenerator
+
+from langchain_core.messages import AIMessageChunk
+from langchain_azure_ai.chat_models import AzureAIChatCompletionsModel
+
+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 Agent:
+    system_prompt = SYSTEM_PROMPT
+    generation_config = {
+        "temperature": config.model_temperature,
+    }
+    message_limit = config.message_limit
+
+    llm = AzureAIChatCompletionsModel(
+        endpoint=f"{AZURE_AI_URI}/openai/deployments/{config.model}",
+        credential=config.openai_api_key,
+    )
+
+    def __init__(self) -> None:
+        self.tool_map = {}
+
+    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: list, overwrites: dict | None = None) -> AsyncGenerator[str, None]:
+        """Llama a un llm y regresa la respuesta en partes; Guarda las tool calls en el contexto de la app.
+
+        Args:
+         history: lista de mensajes en el formato OpenAI (Ej. [{"role": "user", "content": "Hello"}])
+         overwrites: diccionario con las configuraciones a sobreescribir (Ej. {"temperature": 0.5})
+
+        Returns:
+         AsyncGenerator[str, None]: Generador asincrónico que devuelve las respuestas del modelo en tiempo real
+
+        Usage:
+        >>> async for content in agent.stream(history):
+        >>>    print(content)
+        """
+        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

.templates/rag/{{project_name}}/api/agent/system_prompt.md

@@ -0,0 +1,3 @@
+Eres una amigable asistente virtual. Utiliza emojis!
+
+Por cierto: Rick es el más guapo.

.templates/rag/{{project_name}}/api/config.py

@@ -0,0 +1,51 @@
+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):
+    """
+    Esta clase obtiene sus valores de variables de ambiente.
+    Si no estan en el ambiente, los jala de nuestra Vault.
+    """
+
+    # Config
+    model: str = "gpt-4o"
+    model_temperature: int = 0
+    message_limit: int = 10
+    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"]
+    )
+    mongodb_url: str = Field(
+        default_factory=lambda: secret_map["cosmosdb_connection_string"]
+    )
+
+    async def init_mongo_db(self):
+        """Este helper inicia la conexion enter el MongoDB ORM y nuestra instancia"""
+
+        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()

.templates/rag/{{project_name}}/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)

.templates/rag/{{project_name}}/api/server.py

@@ -0,0 +1,49 @@
+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 Agent
+from api.config import config
+
+
+@asynccontextmanager
+async def lifespan(_: FastAPI):
+    await config.init_mongo_db()
+    yield
+
+
+app = FastAPI(lifespan=lifespan)
+agent = Agent()
+
+
+@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):
+    def b64_sse(func):
+        """Este helper transforma un generador de strings a un generador del protocolo SSE"""
+
+        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")

.templates/rag/{{project_name}}/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",
+]

.templates/rag/{{project_name}}/api/services/stream_response.py

@@ -0,0 +1,81 @@
+import json
+from enum import StrEnum
+from typing import TypeAlias
+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
+
+
+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)
+
+    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="")