The PoliRural+ Advisor and RAG services

The Advisor and Magpie: answers from the project's own evidence

main tellingdepth: standard3 min markdown

The PoliRural+ Advisor answers from the project’s own documents rather than the open web. Magpie does a similar job for geospatial metadata catalogues.

The Advisor

The Advisor is a custom GPT built with ChatGPT’s retrieval-augmented generation (RAG) technology. It gives pilot stakeholders access to the project’s Knowledge Space and References database, and it supports RAP preparation, policy briefs and reviews of pilot evidence (D4.1 §6.2; D4.2 §3.7). Anyone with the link can use it. Its contents stay in the project’s Team workspace and are not used by ChatGPT to train its wider models (D4.1 §6.2).

D4.2 distinguishes two modes:

• textual RAG over reports, policy documents and pilot material;
• GeoRAG, which adds metadata records, map layers, indicators and analytical services, because many regional questions depend on place and not on prose alone.

The goal is an answer whose citations, excerpts and reference trails the user can check, not just text that sounds right. A planned next step links the RAG pipelines to a Knowledge Graph, for entity linking and clearer evidence trails (D4.2 §3.7).

What testers found

In February 2025, partners and pilot users completed 8 structured cognitive walkthroughs (D2.3 Annex B). Opening the Advisor caused no problems (average difficulty 1.0). The dialogue was rated useful (2.9 overall on a 1–4 scale). One tester got good advice on funding instruments for Community-Led Local Development. Another got a wrong reference to the Horizon 2020 programme; it was corrected in later answers. Free-account daily limits cut some sessions short. One pilot warned that advice about the RAP’s future “can be tricky. We need to think for ourselves.”

D2.3 itself shows the Advisor in use. Its literature review on rural-urban linkages was written with the Advisor and the 700+ sources in the project’s References Database (D2.3 §3).

Magpie

Magpie embeds ISO 19115/19119 metadata, the standards behind INSPIRE and OGC catalogues, in a Qdrant vector database. It answers plain-language queries with records listing layer names and WMS/WFS endpoints, and can build a map from them (D4.1 §6.5).

In the Beta, Raven’s data-selection workflow has taken over Magpie’s core function, and its results appear mostly through JackDaw (D4.2 §3.4). Its limit is the metadata: no search makes up for incomplete records.

Next: why one broad question rarely gets a good RAG answer.

How-to: Using the Advisor for RAP work · Finding data with Magpie and Raven.

How to use the PoliRural+ Advisor for RAP work

another tellingdepth: intro..standard2 min markdown

When to use it: you are drafting or reviewing a RAP, a policy brief or a pilot evidence summary and want answers from the project’s own documents rather than the open web.

What you need

• The Advisor link. It is a custom GPT in ChatGPT, “publicly available to any user with the link” (D4.1 §6.2); D7.4 describes access “through a dedicated link”. The pack does not give the address, so get it from your pilot or the project.
• No upload is needed: its contents stay in the project’s Team workspace and are not used by ChatGPT to train its wider models (D4.1 §6.2).

Steps

1. State the task and the place. Name your region, the RAP theme and the kind of output you want (SWOT, measure list, funding options).
2. Keep the question narrow. A RAG engine returns a small top-ranked set of chunks, often 10 to 20; a broad question pulls up general statements (D4.2 §3.6).
3. Ask for the sources. The service is designed to return citations, excerpts and reference trails (D4.2 §3.7). Open them.
4. Continue the dialogue. Clarify, challenge and follow up; one question and one answer does not use the tool well (D4.2 §3.6).
5. Edit and question the draft. D7.4 advises users to edit responses and question outputs rather than accept them.
6. Move repeatable work to an agent. If the same assessment runs for every measure, the Boxes & Arrows Agent Builder in JackDaw takes steps and quality checks.

What you get

Answer drafts grounded in the project corpus, with citations you can check and reuse (D4.2 §3.7). In Annex O of D2.2, a pilot stakeholder asked “How can I develop an excellent RAP for my region?” and got an eight-part structure, from regional context to monitoring, with pointers to D2.1, D5.1 and D2.2.

How the consortium used it

• May 2025: a summary of every pilot’s RAP, its gap and a possible radical innovation, plus a SWOT per pilot. For County Monaghan it named the gap as rural mobility and remote-work infrastructure (D2.3 Annex A).
• August 2025: the Advisor analysed project outputs with the AI Act and EU guidelines on generative AI to draft the Participatory AI Principles; the first is that AI responses must be understood and questioned, not blindly accepted (D2.3 Annex C).
• November 2025: an EU glossary for RAP and grant alignment, built by scanning all public project outputs (D2.3 Annex C).

Watch out

• The private RAG is only as good as the completeness and preparation of the corpus; GeoRAG still needs better links between texts, metadata and spatial services (D4.2 §3.7).
• A Knowledge Graph link for clearer evidence trails is planned, not built.

Boxes & Arrows: build an agent instead of asking one question

another tellingdepth: standard2 min markdown

In the Beta, many users treated the JackDaw chat like a search box: one question, one final answer (D4.2 §3.6). Used that way, a conversational agent gives far less than it can.

Why one broad question falls short

A RAG engine compares the question, or an expanded version of it, with indexed document chunks and returns a small top-ranked set, often 10 to 20 chunks. That is a relevance match against a small subset of the documents, not an analysis of the whole corpus. A specific, well-scoped question can be answered well. A broad question without clear limits tends to pull up general statements with little substance.

So the quality of the answer depends heavily on how the question is scoped. Some tasks, such as research or assessment work with known steps, prior knowledge and quality criteria, need a structured conversation rather than a single prompt.

What the builder does

The Boxes & Arrows Agent Builder, a module of JackDaw, lets users draw an agent as a graph:

• each box is a step with its own instructions, background data or knowledge resources, and a defined output;
• arrows can branch, loop and merge;
• quality checks can sit between steps.

This lets complex research and assessment tasks be structured and partly automated. It is designed so that non-technical users can configure their own agents, which opens agentic AI to a wider group of users.

Status

The builder is a new module in the Beta. D4.2 does not report pilot results for it. More support for model builders is one of D4.2’s priorities before the final release: demos, workflow templates and configuration examples. The candidate demos include rural attractiveness analysis, stakeholder mapping and RAP support (D4.2 §11.1).

Check yourself: A broad question, a generic answer

Your pilot team in Vidzeme, where youth retention is a priority for the rest of the project, asks the PoliRural+ Advisor: “What should our Regional Action Plan do about youth outmigration?” The answer reads well but is generic. It names no specific measures, funds or evidence for your region. What do you do next?

A) Rephrase the same question until a better answer comes back

B) Narrow it down: name the territory, the group and the measure, ask for sources, check them and keep the dialogue going

C) Build a Boxes & Arrows agent with steps for context, options, funding and a quality check

D) Stop using AI; the RAP team reads the documents and decides

Choose one, then read why it works or where it breaks.