VereskLearn

Module 4 · Retrieval

Lesson 4.7

Parent-child and compression

Parent-child retrieval searches small pieces, then returns the larger section that holds them.

Lesson 26 of 622 min read

The problem

Why it matters

Without it

A large section covers many topics, so one vector stands for all of them. A question about one detail can score low against it.

With it

The search compares the question with short pieces, which score sharply. The model then reads the whole section, so no detail is lost.

How it works

  • Cut the section into child pieces

    Each sentence of a section becomes one child piece. Every child stores the id of its parent, which is the section named reset. In a real system the parent id is saved with each child in the database.

    Child textParent id
    Hold the button for 10 secondsreset
    The light then blinks bluereset
    Wait one minute before you sign inreset
    The three children of the reset section. Each one points to the same parent.
    const section = { id: "reset" };const sentences = [  "Hold the button for 10 seconds",  "The light then blinks blue",  "Wait one minute before you sign in",];const children = sentences.map((text) => {  return { text, parent: section.id };});console.log(children[1]);

    Output

    { text: 'The light then blinks blue', parent: 'reset' }

  • Score each child against the question

    The question is compared with each child, not with the whole section. The question words share 1 word with the first child, 2 with the second, and 0 with the third. A short child gives a sharper score, since no other sentence dilutes it.

    Shared words between the question and each child.
    const query = ["what", "does", "the", "light", "do"];const children = [  "Hold the button for 10 seconds",  "The light then blinks blue",  "Wait one minute before you sign in",];for (const text of children) {  const words = text.toLowerCase().split(" ");  console.log(query.filter((w) => words.includes(w)).length, text);}

    Output

    1 Hold the button for 10 seconds
2 The light then blinks blue
0 Wait one minute before you sign in

  • The best child points to its parent

    The best child is "The light then blinks blue", with 2 shared words. The search returns its parent, the whole reset section, so the model reads the instructions around the matched sentence. Those neighbors often hold the warning or the next action.

    1. Matched child
    2. Parent id: reset
    3. Whole section
    The path from the matched child to the text the model reads.
    const sentences = [  "Hold the button for 10 seconds",  "The light then blinks blue",  "Wait one minute before you sign in",];const best = sentences[1];console.log("Best child:", best);console.log("Parent text:", sentences.join(". ") + ".");

    Output

    Best child: The light then blinks blue
Parent text: Hold the button for 10 seconds. The light then blinks blue. Wait one minute before you sign in.

  • Compression keeps the sentences that help

    A long parent can be cut down to the sentences that answer the question. The wait sentence does not help with the light question, so compression drops it. The model gets two sentences instead of three, which keeps the prompt short.

    Full parent

    • Hold the button for 10 seconds
    • The light then blinks blue
    • Wait one minute before you sign in

    Compressed

    • Hold the button for 10 seconds
    • The light then blinks blue
    The parent text before and after compression, for the light question.

Full code

Read the program

See the full code

Common mistakes

What goes wrong

  • Cutting children too small. A single word has no meaning, so it matches the wrong parents.
  • Returning the child alone. The model then lacks the context around the matched sentence.
  • Losing the parent pointer when the documents are loaded again. A child with no parent cannot return its section.

Use it when and skip it when

Is it the right tool?

Use it when

  • Sections are long, and each one covers several topics.
  • The answer needs the text around a matched sentence, such as a step and its warning.

Skip it when

  • Every chunk is already short, such as one-line FAQ entries.
  • The whole parent is too large for the model, and no compression step is planned.

Words to know

Terms in this lesson

Parent
The larger section that holds child pieces.
Child
A small piece of a parent, indexed for search.
Compression
Removing text that does not help answer the question.
Pointer
A stored id that links one record to another.
Full code