Cadence
An agentic, retrieval-augmented assistant over public-transit data: it answers GTFS specification questions with grounded, cited responses and queries a live STM Montréal feed, then exposes everything as an MCP server.
Timeline
2026-06 — 2026-06
Technologies
Overview
Talk to your transit data — spec knowledge and live schedules
Cadence is an agentic RAG system over public-transit data. It answers questions about the GTFS specification with grounded, cited responses, queries a real STM Montréal feed (7.1M stop_times rows) for live schedules, and exposes every capability as an MCP server any LLM client can use.
RAG + Agent
Core
Grounded spec RAG + tool-using agent
MCP
Protocol
Tools exposed via Model Context Protocol
7.1M
Live data
STM stop_times rows queried in SQLite
k=3
Tuned
Retrieval depth chosen by precision/recall eval
Retrieval
Grounded RAG with a self-contained chunking pipeline
The GTFS spec is chunked with a header-aware splitter that keeps each field definition intact and prefixes every chunk with its breadcrumb, so retrieved passages stay self-contained. Embeddings come from Azure OpenAI and are stored in Chroma; answers are grounded in retrieved context with an explicit "I don't know" guardrail.
- Header-aware Markdown chunking with token overlap, so field definitions are never cut mid-way.
- Azure OpenAI embeddings + Chroma cosine retrieval, behind a provider-agnostic client.
- Grounded, cited generation with an out-of-scope guardrail (faithfulness 1.00 on the eval set).
Evaluation
Decisions driven by measurement, not assumptions
An LLM-as-judge evaluation harness (faithfulness, answer relevancy, context precision, context sufficiency) was built before tuning, so every retrieval change is justified by numbers.
- 1Implemented a cross-encoder reranker, measured it, and disabled it — it lowered precision on this structured corpus.
- 2Swept retrieval depth k and selected k=3: full context sufficiency at the best precision.
- 3Kept the reranker in the codebase as an evidence-based, opt-in component.
Measurement over intuition
Two cross-encoder rerankers were tested and rejected on evidence (precision fell from 0.57 to 0.36–0.44), and k was tuned on the precision/recall tradeoff — rigor that separates shipping from guessing.
Agentic + MCP
One agent, real data, a standard protocol
A tool-calling agent decides per question whether to consult the GTFS spec (RAG) or query live STM data, and fuses multi-tool answers. The same tools are exposed through an MCP server, so external clients like Claude Desktop become the agent.
- Function-calling agent routes between spec RAG and structured transit tools.
- Transit tools run parameterized SQL over the STM feed — injection-safe, no schema hallucination.
- An MCP server (FastMCP) exposes the tools to any MCP-compatible LLM client.
Interoperability in action
The exact tools that power the in-app agent are callable from Claude Desktop via MCP — query Montréal departures or the GTFS spec directly inside a chat.
Stack & Delivery
Full-stack and reproducible
A FastAPI backend serves the agent; a frontend provides a modern chat UI that surfaces which tools each answer used.
More like this
Related work
ai
MrLabelling
Research platform for automatic 3D object segmentation and reconstruction from HoloLens 2 data. Combines YOLO, Mask R-CNN, and SAM achieving 80% precision for mixed reality applications.
ai
ClearSheet
Document scanning and OCR web app that detects a page inside a noisy phone photo, corrects perspective, generates a clean scan-style result, and extracts text in French, English, or both.
ai
ThoraxVision
AI-powered web application for detecting pulmonary diseases from chest X-rays using deep learning. YOLO-based detection with 80% accuracy on the NIH dataset.