Retrieval: giving the AI your docs

Play this lesson

Look it up first

After the Winter Bloom update, the Garden Guide still describes last season's seeds, and Kai gets the complaints. A model only knows what it was trained on, not Shard Garden's latest update. Retrieval finds the right help pages for each question and adds them to the prompt.

What RAG means

Anthropic: "RAG is a method that retrieves relevant information from a knowledge base and appends it to the user's prompt". The model then answers from those pages.

Quick check

Why add help pages to the prompt?

  1. So answers match the current game
  2. To make the prompt shorter
  3. So the model can skip reading
Show the answer

So answers match the current game

The pages carry today's facts, which the model never saw in training.

Small? Send it all

Anthropic: "If your knowledge base is smaller than 200,000 tokens (about 500 pages of material)", "you can just include the entire knowledge base in the prompt". Retrieval matters when there is more.

Pick the plan

Remy counts Shard Garden's help pages: 40 of them. What is simplest?

  1. Put them all in the prompt
  2. Build a retrieval system
  3. Leave them out
Show the answer

Put them all in the prompt

40 pages is far under 500. With prompt caching, sending them all stays cheap.

Better search, fewer misses

With thousands of pages, the search must find the right ones. Anthropic reported that adding context to each piece of text before searching "reduced the top-20-chunk retrieval failure rate by 49%".

Keep the pages current

Retrieval is only as good as the pages. When an update changes the game, the help pages change in the same release, and the search index is rebuilt.

Quick check

Once retrieval works, the help pages never need updating.

Show the answer

False

Old pages give old answers. The pages change with every update.

What does each player cost?

Every question costs money. With a million players, it adds up.

Bolt says: Next: a budget per player.

Fact-checked 12 October 2026