Sevilla FC's Scout Advisor puts a query box in front of 300,000 scouting reports

The interface is a question, the product is what happens before the model sees it

Sevilla FC's scouting problem was a search problem. The club holds over 300,000 scouting reports, and evaluating a single shortlist of players used to take recruiters 200 to 300 hours of manual reading. Scout Advisor, built by the club's data department with IBM on watsonx and using Meta's Llama 3.1 70B Instruct, replaces that with a text box: a recruiter asks a question and gets back a list of matching players plus AI-generated summaries.

From a frontend perspective, the surface is deliberately plain — natural language in, ranked results out. But the announcement is unusually candid that a raw query typed by a recruiter is not what gets sent to the model. There's a translation step in between, and that step is where the system earns its keep.

Prompt enrichment as a domain vocabulary layer

Meta calls the mechanism prompt enrichment: the system automatically rewrites a user's question to add soccer-specific context before it hits the analysis. The example given is concrete. A recruiter types "show me talented wings," and the system expands it internally to "A talented wing takes on defenders with dribbling, creating space and penetrating the opposition."

The failure case Meta names is telling — without enrichment, a general-purpose model might return a chicken wing recipe. That's the whole argument for building this layer: a soccer term like "wing" is ambiguous to a generic model, and the club can't retrain recruiters to write disambiguated prompts. So the disambiguation lives in the application, not in the user's head. The frontend stays as simple as a search bar precisely because the enrichment layer absorbs the domain knowledge.

Why the model choice was about Spanish and summarization, not raw capability

Chief Data Officer Elias Zamora is specific about why they picked this model rather than defaulting to the largest available option.

We selected Llama 3.1 70B for its text enrichment and summarization performance, particularly in the Spanish language.Montana Labs

That framing matters. The scouting reports carry qualitative judgments — attitude, tenacity, leadership — written in the recruiters' language. The model's job is to compress those into summaries a decision-maker can read fast, in Spanish. Sporting Director Victor Orta describes the payoff at the point of use: instead of reviewing 45 reports on a player, he says he can get what he needs to make a decision in roughly two minutes. The value shows up as reading time collapsed, not as a novel capability.

What Sevilla's build shows about wrapping a model for non-technical users

The specific implication here is that the hard part of a natural-language interface over proprietary data isn't the chat box — it's the invisible rewriting that makes casual questions map onto a specialized corpus. Sevilla FC didn't ask recruiters to learn prompt engineering; they moved the domain vocabulary into an enrichment step so the input could stay conversational.

Zamora's own advice frames the prerequisites plainly: a solid data foundation, deep business understanding, and a well-educated team. Those 300,000 reports were the asset; the interface only worked because someone encoded what "a talented wing" means before the query reached the model. Teams building similar tools should budget for that translation layer as a first-class piece of the product, not an afterthought bolted onto a prompt.