AI enablement · Knowledge and retrieval
Answers from your own documents, with the source attached
We build systems that answer questions from your own documents, and show the source of every answer so a person can check it. The quality of the search is what matters most. The AI model itself is rarely what holds these systems back.
Where this starts
The demonstration always works
A document search system built in an afternoon gives excellent answers from the ten documents it was built with. At ten thousand documents, covering several years, including three outdated versions of the same policy and plenty of material nobody has organised, it starts giving answers that sound convincing but are wrong. The AI model has not changed. The search has simply stopped finding the right passage.
What we usually find
- Confident answers taken from an outdated version of a document
- A convincing answer, when the correct response was that the documents do not cover the question
- No source given, so nothing can be checked without redoing the work by hand
- Documents in formats the system quietly skipped
- Tables, clauses or procedures split in half when the documents were processed
- No measurement, so quality depends on the last person who tried it
Our position
Almost every disappointing document search system has a search problem, even though it looks like a problem with the AI model.
What the work covers
Processing, searching and grounding answers in sources
Most of the work happens before the AI model is involved. It is not glamorous, but it is what decides whether people trust the system or give up on it.
Bringing documents in
We bring in your documents, wikis, tickets, shared drives and scanned files. Layout, tables and structure are preserved, rather than flattened into one long block of text.
Splitting documents sensibly
We split documents along their own natural sections, not at a fixed number of characters. A clause, table or procedure stays together as one piece, rather than becoming two useless halves.
Combined search
We combine keyword search with meaning-based search, then rank the results. Meaning-based search alone often misses exact terms such as a policy number, part code or case reference, which are exactly what people search for.
Current and outdated versions
The system knows which version of a document is current, and keeps withdrawn versions out of its answers. In a regulated business, this is the most damaging mistake the system could make.
Sources and honest answers
Every answer shows its source. The system can also reply that the documents do not cover the question, which is far better than a convincing answer that has been made up.
Measuring search quality
We build a set of test questions with known correct passages. Any change to the system can then be judged on evidence, rather than on whether it seemed better after three attempts.
How quality is maintained
Search quality that is measured, not guessed
Everything here exists so that an improvement can be proved, rather than argued for.
Test questions
Real questions are paired with the passages that should be found, and the set is kept up to date. It is the only fair way to compare two versions of the search.
Finding the passage
We measure whether the correct passage was found at all, separately from whether the answer read well. These fail for different reasons and need different fixes.
Supported answers
Automated checks confirm that each answer is supported by the passages it cites. An unsupported claim is caught before release, not by a reader.
Refresh
Documents are reloaded on a schedule, and outdated versions are removed. The collection stays current without anyone having to maintain it by hand.
Gaps
Questions that the documents cannot answer are logged and reported. This is one of the most useful lists the owners of your documents will receive.
What you are left with
- A process for bringing in documents that handles the formats you actually have
- Combined keyword and meaning-based search, tuned against your own questions
- Answers that cite their source, and can say when the documents do not cover a question
- A set of test questions that makes every change measurable
- A regular report of questions your documents cannot answer
Tell us what people keep asking and cannot find
We will tell you whether your documents can answer it at all. That is the question that decides the project.