Agentic map builder

Evidence producers and evidence users often need to know what evidence already exists on a topic and just as importantly, where the gaps are. Traditionally, building an evidence gap map means manually searching databases, screening hundreds or thousands of titles and abstracts and sorting studies into a taxonomy by hand. This is slow, labour-intensive and hard to keep up to date as new research emerges. 

The agentic map builder is an agentic AI tool that speeds up this process while keeping a human expert in control of the key decisions. It works in stages:

1. The user enters a research question in plain language. 
2. The research question is broken down into its underlying areas of interest. From this, it generates of an open-ended set of candidate search strings. 
3. The user reviews these and keeps as many, or as few, as they judge relevant. 
4. Each retained search string is handed to its own subagent, which independently runs the search, pages through the results and decides for itself when it has seen enough to stop. 
5. A screening agent proposes inclusion/exclusion criteria for the user to approve, then screens the retrieved titles and abstracts. 
6. A classifier agent suggests taxonomy axes from the scoped repository’s taxonomy and sorts the included studies into a structured map. 

The result is a rapidly generated, transparent evidence gap map showing where research is concentrated and where it is not. It is built for researchers and evidence teams who want an up-to-date picture of a field without months of manual work while still reviewing and approving every key decision along the way. 

Example of how the agentic map builder is used

A team member wants to map the evidence on “the impacts of air quality on respiratory disease”, a question that is broad enough to generate an evidence gap map because it spans multiple pollutants, populations and health outcomes. They enter their question into the tool.

The tool maps the query against the DESTINY climate and health taxonomy and generates a set of candidate search strings, for example:

(“air quality” OR “air pollution” OR “particulate matter” OR “PM2.5” OR “PM10” OR “NO2” OR “ozone”) AND (“respiratory disease” OR “respiratory diseases” OR asthma OR COPD OR bronchitis OR “lung disease” OR “respiratory health”)

The team member reviews these search strings and keeps the ones that best match their research question.

Each search string is assigned to an agent which runs the search, pages through the results and stops once it has judged that it has covered enough literature.

A screening agent then proposes eligibility criteria to determine which records should be included which the team member approves before screening runs. For example:

Inclusion: Evidence must report on air quality measures or exposures (e.g., levels of particulate matter, ozone, nitrogen dioxide, sulphur dioxide, or other relevant pollutants).

Inclusion: Evidence must examine effects or associations with respiratory diseases (including but not limited to asthma, COPD, lung cancer, respiratory infections, or other respiratory conditions).

The output is an evidence gap map of, for example, air quality indicator versus respiratory disease type showing where research is concentrated and where there are gaps.

Caveats and future development

The performance of the agentic map builder has not yet been formally evaluated. The tool is in active development and its accuracy, coverage and reliability have not yet been benchmarked. Outputs should currently be treated as a starting point for expert review, not a final or authoritative evidence base.

Planned areas of future development

Citation snowballing

A feature that traverses the citation network of screened studies i.e., following references and citations, to surface additional relevant records that may be missed by keyword searching alone.

Expanding coverage beyond the repository

Exploring ways to supplement results with extra-repository references (for example, citation chasing, other databases, grey literature).

Supporting iterative, sub-question search

Considering whether the tool can move beyond a single search pass, for example, generating and running follow-up sub-questions once an initial result set has been found.

Improving state and replicability

Considering how a completed agentic search could be captured and shared e.g., a link to a specific search’s results so that others can view, reproduce, or build on the same run.

Moving towards a “living” map

Considering whether the tool could refresh its results automatically (for example, daily) as new evidence is published, rather than requiring a fresh manual run each time.