Answers from your own documents, with the sources attached
Your team asks a question in plain language and gets an answer built only from your documents, with a link to the paragraph it came from. When the answer is not in your documents, it says so instead of inventing one.
Talk to an engineerThe problem this solves
- The answer exists, in a PDF, in a folder, in a version somebody renamed final_v3.
- The same question reaches the same three people every week, and answering it is not their job.
- Somebody acts on an outdated procedure because the current one was impossible to find.
- A general chatbot answers confidently using something it read on the internet instead of your rules.
- New staff take months to become useful, because the knowledge is in people rather than in a system.
How it works
1. Ingestion that can be re-run safely
Point it at your documents. Each one is fingerprinted by content hash, so re-running only reprocesses what actually changed. No duplicates, no full reindex every time.
2. Retrieval that finds the right paragraph
Two searches run in parallel: semantic, for meaning, and keyword, for exact terms like article numbers and product codes. The rankings are merged, then near-identical passages are filtered out so the context is evidence rather than five copies of one paragraph.
3. An answer bounded by the evidence
The model is given the retrieved passages and instructed to answer only from them. Every claim carries a citation the reader can open. If the evidence is not there, the honest answer is that it is not there.
4. Quality that is measured, not assumed
A fixed set of your real questions with expected answers, scored on every change. When a document set, a prompt or a model changes, you see whether quality moved before your users do.
Who it is for
Professional associations and chambers
Members asking the same questions about regulations, fees, procedures and deadlines. The rules are public, written down, and still generate constant phone calls.
Firms working against changing regulation
Accounting, legal and compliance teams, where the cost of citing a superseded rule is measured in money rather than embarrassment.
Companies with dense internal documentation
Operations manuals, technical specifications, safety procedures. Anywhere the answer exists but finding it takes an expert.
Support and onboarding teams
Turning an accumulated body of tickets, wikis and manuals into something a new hire can query on day one.
What you get
- The system deployed in your infrastructure or ours, your choice
- Ingestion connected to your document sources, safe to re-run as they change
- An evaluation set built from your real questions, so quality has a number
- A web interface, an API, or both, depending on where your team already works
- Documentation covering how it works, what it costs to run, and what it will not do
Questions we get asked
How is this different from just using ChatGPT?
A general assistant answers from what it learned on the internet, and it is confident whether or not it is right. This answers only from your documents, cites the paragraph it used, and says when it does not know. That difference is the entire point: for a regulation, a procedure or a contract, an answer without a source is not usable.
What happens when our documents change?
Re-run the ingestion. Documents are fingerprinted by content, so only what changed is reprocessed and the rest is left alone. That is a design decision rather than an optimisation: it means keeping the index current is cheap enough that it actually gets done.
Do our documents get used to train someone's model?
No. That is a deployment decision made explicitly rather than inherited from a default, and it is documented before the first request. If the requirement is that nothing leaves your infrastructure at all, that changes which models are viable, and we say so up front rather than after.
How do we know the answers are actually good?
Because it is measured. We build an evaluation set from your real questions with the answers you expect, and score against it on every change. You get a number and its history, not an impression from a demo.
What does it cost to run?
Two parts: the build, which is scoped and quoted before starting, and the running cost, which is mostly model usage and scales with how much it is used. We size the running cost during the build and show you the levers, because an AI feature with an unpredictable bill gets switched off within a quarter.
Can we start small?
That is the recommended path. A pilot on one document set and one team answers the only question that matters, which is whether it is good enough on your material. Everything else is speculation until then.
Start with a pilot
The fastest way to find out whether this works on your documents is to try it on your documents. Send us a note describing what your team keeps looking up, and we will tell you honestly whether this is a good fit before anyone commits to anything.