RAG and knowledge systems

Answers from your own documents and data, with the source shown every time.

A retrieval system that pulls the figure from the document that actually contains it, cites it, and says what it does not hold rather than guessing. Built for your material, not pointed at a folder.

See it in our work

At a glance

Starts with
A corpus audit
Answers
Cited, or declined
Runs in
Your account and region
Your data
Never trains a shared model

What it covers

Built for your documents, not a demo.

  • Retrieval built for your documents

    Chunking, reranking and purpose built tools per document type. Factsheets, contracts, research notes and policies are not the same shape, so they are not retrieved the same way.

  • Figures locked in pictures

    Much real data lives inside images of charts and tables. We make those answerable, which is usually the difference between a demo and a system.

  • Public named, confidential protected

    Public sources are attributed by name. Confidential material is used and attributed to the process, never to a person.

  • Honest about gaps

    The agent confirms it can answer before it tries, and declines in plain words when the material is not there.

How we do it

From a corpus audit to cited answers.

  1. 01

    Corpus audit first

    What you have, what shape it is in, and where the figures your users ask about actually live. This is where most RAG projects go wrong, so it is where we start.

  2. 02

    A data layer, not a chatbot

    One tool per source or view, bound to what that source can answer. The agent cannot ask what the data cannot answer.

  3. 03

    Grounded, cited, calculated

    Every figure comes from a query or an approved document, carries its source, and every sum goes through a calculator rather than the model’s head.

  4. 04

    Marked by your team, then shipped

    Real questions, marked right or wrong with the reason. The marked set runs before every change and lets you swap models later.

What you get

A knowledge layer your team can check.

  • A retrieval layer built for your document types, in your cloud
  • An answering agent that cites every figure and declines what it cannot support
  • A client marked evaluation set, run before every change
  • Ingestion for your sources: drives, CMS, PDFs, databases
  • Source code, documentation and a handover session

Where we have done it

Real engagements, named.

Questions

Questions we get asked.

Our documents are a mess. Is that a problem?
It is the normal state. The corpus audit tells us what is usable as is, what needs a purpose built tool, and what is not worth ingesting yet.
Will it hallucinate?
It answers only from the retrieved material and cites it. When the material is not there, it says so. The evaluation set is how we prove that, on your questions.
Which vector database?
pgvector, Pinecone or Qdrant, chosen for your scale and where your data already sits. It is not the hard part.
Can it stay inside our network?
Yes. Retrieval and the knowledge base run in your account and region, and nothing you send trains a shared model.

What clients say

In their words.

01 / 04

Life Insurance Direct
Before Metricsense, checking calls was a manual process: we selected files for review. Now the platform analyses every call, and we review by risk. It has identified customer experiences we would not have picked up before, and it lets us scale our quality assurance program without increasing headcount, with our team still the human in the loop. It helps us protect our customers, our business and our agents.

Russell Cain

Founder and CEO, Life Insurance Direct

Also under AI engineering

Where it leads next.

All of AI engineering

Not sure this is the right place to start?