Research areas

Overview

Research areas

Retrieval and knowledge systems

Most of what a company knows lives in documents nobody can find. We build retrieval systems that find the right passage in large private collections, and answer questions with sources a person can check.

  1. sources

    docs · wikis

  2. parse

    layout · chunks

  3. index

    vectors · keywords

  4. retrieve

    search · rerank

  5. answer

    LLM · citations

Fig. 1A retrieval pipeline where answers are only as good as what reaches the model

Typical problems

  • Answers sound confident but cite the wrong document, or none
  • Search works on clean text but fails on PDFs, tables and scans
  • Permissions differ per document and must be respected in every answer
  • Nobody can tell whether a change to the pipeline made results better or worse

What you get

  • Ingestion that handles your real formats, including tables and scanned pages
  • Hybrid search with reranking, tuned on questions from your users
  • Answers with citations to the exact source passage
  • A retrieval evaluation set and metrics you can track over time

How we approach it

Retrieval quality decides answer quality, so that is where we spend most of the effort. We build a test set of real questions with known correct sources first, then measure each stage separately: whether parsing kept the content intact, whether search found the right passage, and whether the answer used it faithfully.

We combine keyword and vector search, add reranking where it pays for itself, and keep access control in the retrieval layer rather than relying on the model to hide what a user shouldn’t see.