Data Nexus

Data · We run it ourselves

Retrieval-augmented generation

Answering from a body of documents the model retrieves at question time, rather than from what it absorbed in training.

A model that answers from training answers plausibly. A model that answers from your catalogue answers correctly, and the difference is entirely in the retrieval — which is where nearly all the engineering is, and nearly none of the attention.

en.wikipedia.org

01/Why this one

  1. 01

    It makes an answer traceable. When a reply can be pointed back to the passage it came from, a wrong answer becomes a fixable retrieval problem instead of an argument about the model.


  2. 02

    It changes what an update costs. Correcting a fact means editing a document, not retraining anything, so the people who own the knowledge can own the corrections.


  3. 03

    Grounding is what makes a sales conversation survivable. A buyer comparing a serious purchase asks specific questions, and an assistant that invents an accreditation has done more damage than one that stayed silent.


  4. 04

    Embeddings are only the index. Chunking, ranking, what gets shown to the model and what happens when nothing relevant is found are the parts that decide the answer, and they are ordinary engineering rather than magic.

02/What you can check

An entry that cannot point at something you can open, run or read does not compile. That is a property of the type, not a promise in a paragraph.

Case record
An agent whose knowledge is generated from the live fleet on every build, so a car that leaves the fleet leaves its answers in the same deploy.Read the record

Case record
A retrieval agent over a course catalogue — curriculum, hours, accreditation, prerequisites, price — configured to state the price early and to disqualify enquirers the course was wrong for.Read the record

In our code
mozg/contracts/content-db/src/embeddings.ts — An embedding and content layer held as a contract between services rather than inside one of them.

Product we run
An adaptive learning platform we run ourselves, which means we carry its wrong answers.CyberEGE
03/Where it stops

Invariant

Retrieval cannot fix a corpus that disagrees with itself. Where two documents state different prices, the system will confidently return one of them, and the work that follows is editorial rather than technical. We say this before the build, because it is the most common reason a grounded assistant disappoints.