WHAT WE DO

Knowledge bases built to be asked hard questions

Most organisations already hold the material that answers their questions. It sits in policy documents, indices, reports and research, spread across formats and years, and the questions people actually ask cut across all of it at once.

We build the layer that turns that material into something you can interrogate: a knowledge base with a retrieval engine for what the documents say, a knowledge graph for how the things in them relate, and provenance on every answer so a reader can check it against the source.

THE PROBLEM

Search finds passages. Decisions need relationships.

A policy adviser rarely asks a question that one paragraph answers. They ask which of these twelve countries reimburse an intervention, and what changed in the ones that did. They ask where the evidence is thin enough to be worth commissioning. Those are questions about things and how they relate to each other, and a pile of documents with a search box cannot answer them however good the search is.

The question spans the documents
Comparing six countries means holding six answers side by side in the same shape, drawn from six sources that agree on neither structure nor vocabulary nor what counts as an answer. Retrieval alone returns six passages and leaves the hard part to the reader.
Half the evidence is not prose
Indicators, index scores, coverage rates and dates live in tables rather than paragraphs. A system that only reads text either ignores them or paraphrases them into something less precise than the number it started from.
Nobody can see what is missing
A document collection cannot tell you which country has nothing on a topic, because an absence looks exactly like something not yet found. A structured base can, and in policy work the gap is often the finding worth acting on.

HOW IT WORKS

Two things, doing different jobs.

These two techniques get talked about as alternatives. They are not. Each is weak exactly where the other is strong, which is why we build with both and why the interesting engineering is in how they are joined.

Retrieval-augmented generation

Answers in the words of the source. Ask a question, the system finds the passages that bear on it and writes an answer from them, with those passages attached. This is what makes an answer checkable: you read the paragraph it came from and judge it yourself. It handles nuance, caveats, and the careful hedging that policy documents are full of.

Where it falls short. It works a passage at a time. It has no idea that the programme named in this document and the one named differently in that document are the same programme, so it cannot reliably line up twelve countries and keep the columns straight.

The knowledge graph

The map underneath. Countries, policies, indicators, interventions and periods, held as things with explicit relationships rather than inferred each time from wording. It knows that a country introduced a policy in a given year, that an index score belongs to a particular country and period, and that two documents are discussing the same programme under two names.

Where it falls short. It holds structure, not nuance. It can tell you a policy exists and when it arrived. It cannot tell you how carefully the paragraph describing it was worded, which is often the part that matters.

Joined, the graph decides what to look at and the retrieval layer explains what it found. A question about twelve countries becomes twelve precise lookups rather than one broad search, and every cell of the answer still carries the passage behind it.

WHY IT MATTERS

What the combination makes possible

Comparison that holds its shape
Ask the same question of twenty countries and get twenty answers in the same form, because the graph guarantees you are comparing the same thing rather than whatever the search happened to surface for each one.
Numbers and prose in one answer
An index score sits next to the paragraph that explains the policy behind it. The structured side keeps the figure exact, the retrieval side keeps the context intact, and neither has to be flattened into the other.
Gaps become visible and reportable
Because the base knows what it should hold, it can say what it does not. That converts an uncomfortable silence into a list, and a list is something a research budget can be pointed at.
Provenance survives aggregation
The failure mode of any comparison table is that the sources vanish into it. Here each cell keeps its link back to the document it came from, so a reviewer can audit one number without unpicking the whole table.
One base, many outputs
A chatbot for interactive work, country briefings for circulation, comparison tables for a committee. These stop being separate production jobs and become views of the same underlying base.

IN THE FIELD

Two country-level knowledge bases, in build now

With the Copenhagen Institute for Futures Studies, we are building the knowledge infrastructure behind two international health policy programmes. Both follow the same pattern: a country-level base combining policy documents with structured indices, an interrogation layer over the top, and prototypes reviewed by an expert board before they reach a policy audience. Technical governance runs jointly with CIFS and the University of Southern Denmark.

RSV infant immunisation

A country-level knowledge base for ten countries, built to help each of them make evidence-based decisions on RSV immunisation for infants: Spain, Italy, Germany, France, Portugal, Belgium, Ireland, the Netherlands, Chile and the United States.

Decision makers query the base directly and compare across countries, and best practice from the countries that moved first is mapped so it can travel to the ones that have not. Prototypes are tested with a sounding board of internal and external experts before going anywhere near a policy audience.

The Respiratory Health Initiative

A knowledge base supporting countries in delivering their chronic respiratory disease commitments under the 2030 political declaration. It integrates the Severe Asthma Index and the COPD Index with policy documents and with new targeted research written specifically to fill the gaps the base exposes.

Coverage runs to countries with populations above forty million plus priority countries, phased across 2026. First France, Germany, Italy, Poland, Spain and the United Kingdom, then Canada, China, India, Japan, South Korea, South Africa, Egypt and Algeria. Latin America is included both for population and for the value of learning across a shared language: Brazil, Mexico, Colombia and Argentina, with Chile, Peru and Costa Rica.

Alongside the interrogation layer, the base produces short country briefings carrying the essential indicators for every country in the initiative, and prototypes are reviewed by an advisory board of pulmonologists.

THE METHOD

How a knowledge base gets built

The order matters more than any single step in it. The most common way these projects fail is to gather everything first and work out afterwards what it was supposed to answer.

1. Settle the questions first
The entity model is derived from the questions the base has to answer, not the other way round. Get this wrong and you build something that holds everything and answers nothing.
2. Assemble the sources
Policy documents, indices, published research and grey literature, each carrying its source class, publisher, date and geography, because a retrieval query is often only useful once it can be restricted to national sources from the last three years.
3. Model what the things are
Countries, policies, indicators, interventions and periods, and the decisions that come with them: what counts as the same thing under two names, what relationship is worth recording, what granularity a comparison actually needs.
4. Index for how people ask
Passages made searchable by meaning and by exact term together. Policy work is full of exact terms, and a search that only understands meaning will quietly miss the one document that names the scheme.
5. Join the two layers
A question is resolved against the graph to decide what to look at, then answered from the retrieved passages. This is where most of the engineering goes, and it is what separates a working base from two systems in a trench coat.
6. Test with the people who will use it
Prototypes in front of an expert board, early, before anything reaches the audience it is meant to inform. Domain experts find the failures that no benchmark catches.
7. Keep it current
Sources are refreshed on a schedule, and the language model is treated as a replaceable part. Better ones keep arriving, and a base built properly outlives every one of them.

OWNERSHIP AND GOVERNANCE

It stays yours

A knowledge base is an asset that takes real work to build, and the organisation that paid for it should not discover later that it cannot leave with it. We contract on that basis.

You own the material and everything derived from it. The documents, the metadata, the embeddings, the vectors and the outputs remain yours, whatever format they end up in.
Exportable on request. A complete export in a commonly used, machine-readable and non-proprietary format, with help migrating it somewhere else. That commitment is written into our agreements rather than offered as goodwill.
Continuous access. You keep access to your own data throughout, including if the relationship ends.
Model independence by design. The chat model is a component, not the foundation. It can be swapped for a better one without rebuilding the base underneath.
Deployment where the data has to live. European hosting, or inside your own infrastructure where the material cannot leave it.

Bring us your hardest question

The useful first conversation is not about the technology. It is about the question your organisation keeps having to answer by hand, and whether the material to answer it already exists somewhere in your files. Usually it does, in six formats and four languages.