Francesco Bellingeri

I developed Mango, an AI agent designed to simplify interaction with MongoDB databases through natural language. The goal of the project is to reduce the complexity of advanced queries, allowing users to transform natural language requests directly into MongoDB queries (NL → MQL) through an LLM-based system.

The core of the system is an orchestrated architecture that combines language models with a dynamic tool execution system, enabling the agent to interpret requests, select the necessary operations, and execute them autonomously. The backend was developed asynchronously using FastAPI, with real-time streaming support via Server-Sent Events (SSE), which allows tracking the agent's decision-making process and tool calls step by step.

To progressively improve response quality, I implemented a persistent memory system based on a vector store (ChromaDB), organized into three distinct layers: tool-usage memory, which automatically captures every successful interaction to enrich future prompts as few-shot examples; text memory, which allows saving free-form domain knowledge, field meanings, business terminology, database-specific quirks; training memory, which holds gold-standard pairs explicitly loaded by the user as a privileged reference, separate from auto-generated memory. All three layers are queried on every request and injected into the system prompt. Benchmarks on 110 real-world questions, evaluated using the XMaNeR metric (weighted average of execution rate, output quality, and fuzzy correctness), show compelling results: Qwen3.6-35B (3B active parameters, MoE architecture) reaches 0.877/1.0, trailing Kimi-K2.6 (~1T parameters, comparable to Claude Opus) by just 0.7 points at 1/30 of the cost; DeepSeek-v4-flash (13B active parameters) scores 0.868/1.0. The gap between models of vastly different sizes closes thanks to the context pipeline, schema injection, memory-retrieved few-shot examples, and schema linking enforcement which operates independently of the underlying model.

The architecture is modular and extensible: it supports multiple LLM providers (OpenAI, Anthropic, Gemini) and, for those who prefer self-hosting, local models via Ollama. The tool system is easily expandable. Mango is read-only by design, write operations are rejected at the tool level — making it safe for production environments.