Hashtag AI  ·  graph retrieval  ·  MuSiQue multi-hop example

The Acornsoft Hop

A question that names one passage and depends on another. Watch the graph get built, then watch plain vector search stop one edge short of the answer.

Question What year did the publisher of Labyrinth end? Answer: 1986
Needs both supporting passages
PASSAGE Labyrinth (1984 video game) ✓✕ PASSAGE Acornsoft ✓✕ PASSAGE The Devil’s Labyrinth Labyrinth · published by · Acornsoft Labyrinth · released for · BBC Micro Acornsoft · ceased in · 1986 Acornsoft · relicensed to · Superior Devil’s Labyrinth · is a · film Labyrinth BBC Micro Acornsoft 1986 film QUESTION What year did the publisher of Labyrinth end?
Passages hold facts; facts name entities. The entity layer is what lets one passage reach another.
passage fact entity retrieval hop

Where the graph helped

The question names Labyrinth, so the Labyrinth passage is an easy semantic match. It never says Acornsoft — the word only exists inside the passage you are trying to find. Question-to-passage similarity has nothing to grip, and the second half of the evidence stays out of reach.

Graph retrieval scores facts, not just passages, then lets that score travel along the edges those facts sit on. One shared entity is all it takes.

Same embeddings on both sides. The only thing that changed is what the retriever is allowed to walk across.

  1. 01question → factMatches “Labyrinth · published by · Acornsoft”
  2. 02fact → passageLights the Labyrinth passage — rank 1
  3. 03fact → entityThe same fact names Acornsoft
  4. 04entity → factAcornsoft’s other fact inherits the activation
  5. 05fact → passageThe Acornsoft passage arrives at rank 2