Dalel

Dalel AI

Data infrastructure for superintelligent models.

Agents are limited by the memory they can find, compare, and decide over. Dalel builds that memory: an object database, and semantic operators that read the vectors it stores.

One vector belonging to three objects Overlapping regions labeled person, film, and year. Dots are vectors. Dots in the overlaps belong to more than one object. PERSON FILM YEAR
A review is one vector and three objects: the person, the film, and the year.
  1. 01

    Objects that share vectors

    Vectors live in one collection. An object is a subset: a paper, a person, a statute, a period. One vector can belong to several objects at once.

  2. 02

    Find a similar object

    Filtered vector search stays. The other query asks which object resembles this one. The nearest vector can belong to the wrong object. The score is a distance between sets, and the index approximates it. Several designs will be tried.

  3. 03

    Semantic operators

    Operators run on the vectors already stored. “Which of these discuss this topic” is a decision. The vectors stay fixed. The model returns a decision, not a new embedding and not generated text.

Public collections for science, law, and news come after the database, under rate limits. Stored model state comes after that. The argument is in the master plan.

Read the master plan