Implementation of an Agentic Chat Interface for a Football Scouting Platform using Retrieval-Augmented Generation and Natural Language Database Querying

Kurzfassung

Football clubs store what they know about players partly as structured data (age, position, ratings) and partly as free text in scouting reports. This thesis designs a chat system that answers scouting questions needing both kinds of data. It is built for the scouting platform of SCOUTASTIC and evaluated on a test database containing 486 reports from one club.

A router splits each question into facts and traits. An agent answers facts by writing queries in MongoDB’s query language. A trait search ranks players using their reports and rating fields. A language model acts as judge, checking whether passages support or contradict the sought trait. A test set of 154 questions in English and German was built for the evaluation.

The agent answered 89.8% of fact questions correctly with GPT-6 Sol and 82.0% with GPT-6 Luna. The trait search reached an nDCG@10 of 0.78 (random ranking: 0.22), and the judge halved the number of players with contradicting reports among the top ten.