Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/geremyjampiersalasgarcia-eng/Caso_Practico_Semillero_IA/llms.txt

Use this file to discover all available pages before exploring further.

Mesa de Ayuda IA is built around five independent, specialized AI agents. Every agent is a Python subclass of the abstract BaseAgent class, owns a ChromaDB vector collection as its knowledge source, loads its behavior from a dedicated Markdown system-prompt file, and is registered at startup in the global AgentRegistry. The LangGraph StateGraph orchestrator classifies each incoming request and routes it to the appropriate agent — or to multiple agents in parallel for mixed-intent queries — before consolidating the answers into a single coherent response.

Catalog Agent

Answers questions about Patito S.A. products, technical specifications, list prices, and stock availability. Intent: catalogo_precios.

Policies Agent

Handles discount limits, authorization levels, credit terms, warranties, and return policies. Intent: politicas_comerciales.

Sales Process Agent

Guides sales reps through the CRM pipeline stages, data requirements, and opportunity-closing criteria. Intent: proceso_ventas.

Multimodal Agent

Identifies products from uploaded images using Gemini Vision, then cross-references the catalog for pricing and availability. Intent: multimodal.

Action Agent

The only write-capable agent: collects opportunity data, validates required fields, requests confirmation, then registers via LangChain function calling. Intent: accion_registro.

​
Agent comparison

Agent nameIntent categoryChromaDB collectionKnowledge document
agente_catalogocatalogo_precioscol_catalogo01_Catalogo_Productos_Precios.txt
agente_politicaspoliticas_comercialescol_politicas02_Politicas_Comerciales_Descuentos_Credito.txt
agente_proceso_ventasproceso_ventascol_proceso_ventas03_Proceso_Ventas_CRM.txt
agente_multimodalmultimodalcol_catalogo(shares catalog collection)
agente_accionaccion_registrocol_proceso_ventas03_Proceso_Ventas_CRM.txt
For the mixta intent, the orchestrator runs agente_catalogo, agente_politicas, and agente_proceso_ventas in parallel and passes all three partial answers to a consolidation step.

​
How to add a new agent

Every new agent follows the same four-step pattern:
  1. Extend BaseAgent — create a new file under backend/app/agents/.
  2. Implement the four abstract properties — name, description, collection_name, and system_prompt_path.
  3. Write a system prompt — add a Markdown file under backend/app/prompts/.
  4. Register the agent — import and instantiate it in backend/app/agents/__init__.py inside register_all_agents().
The registry pattern means the orchestrator never needs to be modified; it discovers agents through AgentRegistry at runtime.

​
BaseAgent abstract interface

The following excerpt shows the abstract property stubs and the process_query signature that every agent must satisfy. The concrete implementation in BaseAgent provides the full RAG→prompt→LLM→result pipeline as a default; subclasses that need a different flow (such as MultimodalAgent and AccionAgent) override process_query directly. 
from abc import ABC, abstractmethod
from app.schemas.agent import AgentResult
from app.rag.retriever import retrieve_context
from app.core.llm import get_llm
from app.utils.logger import logger
from langchain_core.messages import SystemMessage, HumanMessage

class BaseAgent(ABC):
    """
    Clase base para todos los agentes especializados.
    Define el flujo estándar RAG + LLM.
    """

    @property
    @abstractmethod
    def name(self) -> str:
        """Nombre del agente (ej. agente_arquitectura)"""
        pass

    @property
    @abstractmethod
    def description(self) -> str:
        """Descripción de lo que hace el agente"""
        pass

    @property
    @abstractmethod
    def collection_name(self) -> str:
        """Nombre de la colección ChromaDB de este agente"""
        pass

    @property
    @abstractmethod
    def system_prompt_path(self) -> str:
        """Ruta al archivo Markdown con el prompt del agente"""
        pass

    def process_query(self, question: str) -> AgentResult:
        """
        Flujo principal del agente:
        1. Busca en su colección ChromaDB
        2. Inyecta el contexto en su prompt
        3. Invoca al LLM
        4. Retorna el resultado con las fuentes
        """
        ...

​
AgentResult schema

Every agent returns an AgentResult Pydantic model. Downstream consumers (the consolidator and the API response) rely on this contract: 
from pydantic import BaseModel
from typing import List
from app.schemas.chat import SourceInfo

class AgentResult(BaseModel):
    """Resultado devuelto por un agente individual tras procesar una consulta"""
    agent_name: str
    answer: str
    sources: List[SourceInfo] = []
    tokens_input: int = 0
    tokens_output: int = 0
Each SourceInfo entry in sources carries document_name, content_snippet (first 200 characters of the matched chunk), and relevance_score from the ChromaDB cosine-similarity search.

Build docs developers (and LLMs) love

Get started for freeTalk to us