Your documents, answered — with the source on every reply.
Retrieval-augmented systems over your policies, contracts, manuals and case history: permission-aware, cited, evaluated on your own questions, and run inside your security boundary.
- Documents and questions
- Build and evaluate
- Pilot with real users
- Operate and extend
Runs where your data is allowed to be.
The knowledge an enterprise runs on is in documents nobody can find in time: the policy that answers the customer's question, the clause that decides the claim, the runbook that fixes the outage. Search returns forty files; the person who knows is on leave. A language model that can read all of it and answer in a sentence changes that — if the answer can be trusted.
Trust is the engineering problem. A demo that answers confidently from the wrong document is worse than no assistant at all. We build retrieval systems the way we build everything else that goes to production: with permissions enforced at retrieval time so nobody sees a document they could not open, with a citation on every answer, and with an evaluation set built from your own questions that the system has to pass before anyone relies on it.
Runs where your data is allowed to be: your cloud tenant, your region, or on your own hardware with open models. Nothing is used to train anything.
Four shapes of knowledge system.
The right one depends on who is asking and what a wrong answer costs. Most engagements start with one audience and one document set.
Internal knowledge assistant
Policies, procedures, runbooks and past cases made answerable for staff — with the user's own permissions applied to every retrieval, so the answer never leaks a document they could not open themselves.
Customer-facing answers
Product manuals, terms and FAQs behind a support channel or a portal. Grounded answers with citations, a confidence threshold below which the system hands over to a person, and a transcript your compliance team can review.
Contract and case analysis
Clause retrieval across thousands of agreements, claims matched against policy wording, precedent found by meaning rather than keyword. Built for legal, claims and procurement teams who need the exact passage, not a summary.
Retrieval inside your products
Knowledge retrieval as a component in the platforms you run — a help assistant inside Workflow & Procurement, policy lookup inside a loyalty programme — delivered as an API your engineers call. See also MCP servers.
How the work runs, day to day.
It goes live when it passes your questions, not ours.
Before any user sees the assistant, we collect the questions people actually ask and the answers a domain expert accepts. That set becomes the test the system must pass — and keeps running after launch, so a change to the documents, the model or the prompts cannot quietly make the answers worse.
- Permission-aware retrieval: the index knows who may see what, and enforces it on every query
- A citation on every answer, linking to the passage, not just the document
- Evaluation on your own question set, rerun on every change, with the pass rate reported
How the work runs.
Documents and questions
Which sources, who may see them, and the real questions collected from the people who will use it. The evaluation set is built here.
Build and evaluate
Ingestion, chunking, permissions, retrieval and the answer layer, measured against the question set until it passes the agreed bar.
Pilot with real users
A small group uses it for real work. Their questions extend the evaluation set; their feedback changes the system.
Operate and extend
Monitoring, drift checks and new document sets. We operate it, or your team does with the runbooks we hand over.
What you keep.
- A retrieval system running in your environment with permission-aware indexing
- An evaluation set from your own questions and the harness that runs it
- An API your engineers and products can call
- Monitoring for answer quality, latency and cost
- Runbooks and the ingestion pipeline for new sources
When this is the right call.
Choose this when people spend real time finding answers that exist in your documents, and a wrong answer has a cost you need to control — compliance, customer trust, or an outage that lasts longer than it should.
Look elsewhere if your documents are few and well-organised: a good search box may be enough. And if the answers do not exist in writing yet, the first job is writing them down.
Frequently asked questions
Services that pair with this.
Which answers do your people hunt for?
Tell us the documents and the audience. An engineer reads it and replies in writing with a view of what a trustworthy assistant over them would take.