Dalel builds data infrastructure for superintelligent models and the agents beside them. Two systems. An object database, then semantic operators that read the vectors it stores. This page is the reason, and the order of work.
Agents keep many things in play and return to the same corpus all day. What they lack is shared memory, and a fast way to address it.
Objects
Vectors live in one collection, with their own metadata. An object is a subset of that collection. A paper, a person, a statute, and a period of time are objects. There are many object types.
One vector can belong to several objects. The embedding of a review is about the film and about the person who wrote it. A judicial opinion is about a case, a doctrine, a court, and a year. The year can be an object of its own. Membership in several objects is the ordinary case, so the data model allows it from the start.
Similar objects
Filtered vector search stays. The other query is which object resembles this one: which matter, which body of work, which period.
The nearest vector is a weak guide. It can sit inside the wrong object, while another object matches better because every part of the query matches inside it.
Scores that compare a query with one record already exist. They assume each vector has a single parent. A vector in this database does not. The score is a distance between sets. Earth mover’s distance and chamfer distance are two of them. An index approximates that distance so the search can name an object. Several index designs will be tried. They are alternatives, not one required layout. Exact set distance remains the final score. The same indexes have to run on accelerators and on disk, where large collections actually live.
Semantic operators
The second system runs on vectors that are already stored. A semantic filter asks whether a predicate holds. Which emails discuss this topic. Which papers concern this method. The same pattern covers operators beyond filters.
The stored vectors stay fixed. A new embedding for each predicate, or a stored embedding per operator, is the costly path. A model reads the query and a stored vector, or a stored vector group, and returns a decision. It is trained across the space of predicates. Large language models check those decisions during training. The output is a decision, not generated text.
Public memory
A private index does not fix a problem every agent will share. After the database works, we publish collections for science, law, and news, under rate limits. Science begins with the articles researchers already treat as the record, such as arXiv and PubMed. Law is legislation and case documents. News is the public account of events.
After those collections, stored key-value state for open models. A document that has been read once should not be recomputed on every request.
Order of work
- The object database. Shared membership. Search for a similar object, with set distance as the score and more than one index design on trial.
- Semantic operators over the vectors that database stores.
- Public collections for science, law, and news.
- Stored key-value state for open models.
Each stage is usable before the next one starts.
Who we are
Dalel AI is the company building this. This site describes the work. There is no account system here.
Correspondence: info@dalel.ai.