The AI Assistant answers League questions by writing queries against Winrate's game data and running them. Open it with the AI Assistant button at the top right of any page, on web or desktop.
One question, start to finish
On Ahri's champion page we asked: How does this champion's win rate change with game length?
The default model answered in about 35 seconds, in five steps:
- It loaded its SQL skill, a guide to querying Winrate's data, and read the page. The URL told it that "this champion" is Ahri.
- It asked for the skill's champion reference with a wrong path, and the request failed.
- It asked again and got a set of tested query templates.
- It wrote and ran one ClickHouse query: Ahri mid in ranked solo games, Gold and above, on patch 16.20, in five game-length buckets beside all mid laners. It drew the rows as a bar chart.
- It wrote the answer from the rows.
Over 171,793 Ahri games, she won 48.32% of the ones that ended before 20 minutes, 52.45% at 25 to 30 minutes and 50.94% past 35. Mid laners as a whole stayed between 50.24% and 50.42%.
We never asked for mid lane, Gold and above or a single patch. The model picked that scope and stated it. It also explained the early dip with "Ahri needs items/levels before she's online", which no query tested. The percentages come from the data. The reason is the model's guess.
What it can do
The model works by writing JavaScript against a small API, and that code runs in a sandbox. On every page the API can read the page you are on, run read-only ClickHouse SQL over ranked games, and turn rows into a bar, line, area or pie chart. Three kinds of page add more:
- Player pages: find players by name and run SQL over one player's games.
- Match pages: load a match and run SQL over its events, per-minute frames and win probability.
- Draft tool: rank champions for the open slot by predicted win probability, and explain the factors behind one.
When the ask box shows a chip for the page you are on, that page's address and text go with your question. Remove the chip to ask without them.
An answer that ran SQL shows the query. The SQL Console and a player's SQL tab have their own ask box with a narrower job: a question becomes exactly one query, written into the editor.
What it remembers
Signed in, you have one thread, the same on every page, on web and on desktop. A guest's thread belongs to their browser. When a thread passes about 60,000 tokens, its oldest messages are condensed into a summary.
Signed in with a linked League account, you are known to the assistant before you ask: Riot ID, region, rank, and your last 20 games with their win rate, KDA and three most-played champions.
It also keeps a memory of stable facts you tell it, such as your role, your champions or how you like answers. It is told to say so in its reply when it saves one. Settings → Account shows the memory and has a Clear memory button. Guests get no memory.
What it costs
The default model is free for everyone, signed in or not.
Every other model in the picker is paid for with AI credits, which come with the Plus and Pro plans (pricing). Plus includes 3,000 credits a month and Pro 10,000. One credit is $0.001 of model usage. Models charge for each token of text they read and write, at the rates in the table. Each step of an answer is charged on its own and rounded up to a whole credit.
Credits per million tokens the model reads (input) and writes (output), from the app's own price list. The $ marks are the model picker's. They rank the cost of three input tokens to one output token.
Without credits, the picker disables the paid models. If one is still selected when your credits run out, the answer comes from the free model, with a note saying why. Used credits reset to zero each time a subscription invoice is paid, and Settings → Billing shows the count.
Limits
- An answer stops after 20 steps. A code run gets 60 seconds and a ClickHouse query gets 25.
- The model writes the SQL itself. Open the query before you rely on a number.
