BreadBowl
NewIntroducing BreadBowl-Embed

Retrieval, rebuiltfrom first principles.

BreadBowl is building a new late-interaction architecture for embeddings. Every passage becomes sixteen routing–value slots, so a single index can both retrieve and rerank, without a second model re-reading the text.

One passage · 16 stored slots

The Eiffel Tower stands in Paris, France. Construction began in 1887 and was completed in 1889 for the World's Fair. The iron tower was designed by engineers at Gustave Eiffel's company. The Louvre museum is also in Paris, while the Colosseum is in Rome.

Question

When was the Eiffel Tower completed?

How it works

Encode once. Read many ways.

Read the technical report
  1. 01

    Sixteen slots per passage.

    Each passage is read a single time and stored as sixteen routing–value slots. Nothing downstream needs the text again.

  2. 02

    Find and rank from one index.

    The routing vectors that retrieve a candidate also decide which of its stored values a question reads.

  3. 03

    Open weights.

    BreadBowl-Embed is on Hugging Face under Apache-2.0, with an inference toolkit and the technical report.

Work with us

Building retrieval for agents, search or memory?

Alpha and preview access are available for model evaluations, design partnerships, and teams building retrieval-heavy products.

contact@breadbowl.ai