Mesa de Ayuda IA is an intelligent help desk built for the sales team of Patito S.A. It uses a multi-agent architecture orchestrated by LangGraph — five specialized agents handle product catalog queries, commercial policy lookups, sales process guidance, multimodal image analysis, and CRM opportunity registration, all powered by Google Gemini and a RAG pipeline backed by ChromaDB.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.
Quickstart
Get the backend and frontend running locally in under 10 minutes.
Architecture
Understand the multi-agent StateGraph, RAG pipeline, and security layers.
Agents
Explore all five specialized agents and how they are routed.
API Reference
Full REST API reference with request/response schemas and examples.
How It Works
A sales rep sends a question (or uploads a product image) through the Next.js chat UI. The FastAPI backend passes it to the LangGraph orchestrator, which:Classifies the intent
Gemini (temperature 0) determines whether the query is about catalog pricing, commercial policies, the sales process, image analysis, a CRM registration action, or a mixed query.
Routes to the right agent(s)
The LangGraph StateGraph dispatches to one or more specialized agents. Mixed-intent queries fan out to three agents simultaneously.
Retrieves relevant context
Each agent searches its dedicated ChromaDB collection using semantic similarity (top-k=4 chunks) with Gemini embeddings.
Key Features
5 Specialized Agents
Catalog, Policies, Sales Process, Multimodal Vision, and CRM Action Registration agents — each with its own knowledge base.
RAG with ChromaDB
Three isolated ChromaDB collections with Google Gemini embeddings provide accurate, document-grounded answers.
Prompt Injection Security
Four security layers: input validation, prompt hardening, server-side sandboxing, and output validation.
Observability
Arize Phoenix traces every LangGraph span via OpenTelemetry OTLP. Token cost tracking per intent category.
LLM-as-Judge Evaluation
Offline batch evaluator scores responses on relevance, precision, and completeness using a Pydantic rubric.
Multimodal Vision
Upload a product image and the Gemini Vision agent identifies it, retrieves catalog data, and returns pricing.
Stack
| Layer | Technology |
|---|---|
| Backend framework | FastAPI + Uvicorn |
| AI orchestration | LangGraph (StateGraph) |
| LLM & embeddings | Google Gemini (gemini-flash-lite-latest, gemini-embedding-001) |
| Agent framework | LangChain |
| Vector store | ChromaDB (local persistence) |
| Database | PostgreSQL (SQLite fallback) |
| Frontend | Next.js 14 + TypeScript + Tailwind CSS |
| Observability | Arize Phoenix + OpenTelemetry |
This project was built for the Semillero de Inteligencia Artificial program by Geremy Jampier Salas Garcia. It is designed as a production-quality reference architecture for multi-agent RAG systems.