Foundation Models & Retrieval

Why this is hard

Most of what goes wrong with a large model in an enterprise is not the model. It is that the passage it needed was never retrieved, or that nobody can say afterwards which passage it used. Our work here is on the retrieval side and the accounting side: parsing estates that were never meant to be queried, ranking over them, and binding every claim to a source that a reader can open.

Questions we are working on

1. How should a document be split when its structure, not its character count, carries the meaning?
2. When is a hybrid of lexical and vector retrieval genuinely better than either, and how do you tell without a benchmark of your own?
3. How does a system decline to answer, rather than improvise, when the retrieved evidence does not support a claim?
4. What is the smallest model that still does a narrow, high-volume task well enough to own outright?

What comes out of it

• Layout-aware parsing of scans, forms and technical drawings, with a confidence figure per extracted field
• Hybrid retrieval with cross-encoder re-ranking, and permission-aware indexes so retrieval cannot leak
• Distillation of narrow tasks into small models a client can run on their own hardware
• Citation and abstention behaviour treated as a measurable property, not a prompt instruction

Programmes from this field

Where it ships