At the club today

Hi, Red here. Today Kirari was looking in our club archive for a game to play together. She typed “games to play together,” and an article called “Two-Player Co-op Games” appeared.

The title did not use the words “play together.” How did search find it? We compared the search phrase with the article’s title.

Meet the idea in a comic

Embeddings help find similar meanings

An embedding represents the meaning or features of text with several numbers. This group of numbers is also called a vector. Turning both a question and an article into number groups lets us compare them and helps find related meanings, even when the words differ.

Kirari

I wanted something two of us could play, so this fits!

A similar meaning is different from a verified answer. Even after finding a game, we still need to check the original for details such as supported devices and player count. Embedding numbers themselves are not an answer or a fact-check.

Split a long guide at its section headings

We can divide a guide that runs for dozens of pages into smaller parts that are easier to find. The process is called chunking; each piece is a chunk. For example, we could split a “Starlight Adventure Guide” into preparations and boss strategy.

Keep the title, section heading, and original link with the chunk. “Prepare a healing item” alone does not tell us which game it describes. The title gives us context. We should also avoid splitting a sentence or separating a condition from the part it applies to.

Questions and documents must be comparable

To compare a question’s numbers with a document’s numbers, their embeddings need a compatible basis. An embedding model is a tool that converts text into a group of numbers. We normally use the same model, with its recommended query and document settings.

Ingestion gets the material ready to find

Ingestion means bringing material in, attaching its source and access information, and storing the chunks and other information needed for search. There is more preparation than simply copying a file into a folder.

1. Keep the original link and a unique document ID, so the same article can be updated when it is brought in again.
2. Attach the section heading and source so we know where a passage came from.
3. Store access information and dates too.
4. Prepare the chunks, embeddings, and indexes needed for search.

We can avoid recalculating embeddings when the text has not changed. But if permissions or dates change, those details still need updating. “Same text” does not mean “nothing needs checking.”

Remove unnecessary contact details and separate the dates

If a friend’s phone number is attached to the end of a game guide, remove or mask personal information that is not needed for the search before ingestion. Asking AI not to reveal it later is not a substitute.

Keep dates separate too. If we add a record of a meeting held on the 12th to the archive on the 20th, the meeting date is the 12th and the ingestion date is the 20th. To find what was decided at the meeting on the 12th, we need the meeting date.

Watch the short · Korean audio

Three small exercises with Red

No software to install. Choose an answer or write one short line in a note.

1. 1. Pick a similar title

   You want a “game two people can play together.” Which title is closer in meaning? ① “Two-Player Co-op Games” ② “Taking a Solo Profile Photo”

   Show the answer and why

   ①. The words differ, but playing together is a similar idea. Check the article itself to see whether the games meet your conditions.

2. 2. Label a chunk

   You took “Prepare a shield” from the “Boss strategy” section of the “Starlight Adventure Guide.” Write a one-line label for the chunk.

   Show the answer and why

   For example: “Starlight Adventure Guide > Boss strategy.” Keep the original link too, so you can check the surrounding text.

3. 3. Write down both dates

   A meeting happened on the 12th, and its record was registered on the 20th. Write the meeting date and the registration date separately.

   Show the answer and why

   Meeting date: the 12th. Registration date: the 20th. Record them separately so date-based searches do not miss the meeting record.

Before we close the archive

Kirari bookmarked the recommendation article, and Nea checked its title and source. Now we can prepare material for search—but what if two notes disagree? Next, we’ll sort out a changed meeting time.