PoliRural+ — Bridging Rural and Urban Europe · knowledge baseInteractive book · llms.txt · For AI agents

From data to decisions, with provenance

Follow a potato-suitability question through JackDaw, and see what the answer must carry before a planner can use it.

What must a GeoAI answer carry before a planner can use it in a Regional Action Plan?

Visible references to the datasets behind it, with their source, update status, quality status and known limitations, so the reader can tell official statistics from pilot data, external web information, expert interpretation or experimental analysis (D4.2 §4.4).

Open in the interactive book · missions, mini-games and memory cards there

One question travelling through JackDaw, step by step A user draws an area and asks a question; a router picks up to five of JackDaw's eighteen tools, GeoRAG finds matching datasets in the Raven catalogue, the tools run on the drawn area, and the LLM loops until no more tool calls are needed. The answer is streamed back with tables and references to the datasets used, and a follow-up question continues the same thread. Sample prompt shipped with JackDaw: “Is this area suitable for a specific crop?” 1 Ask about an area user draws a rectangle on themap or searches a place name;the area travels as WKT withevery chat message 2 Route the request one LLM call classifies theintent and picks up to fiveof the 18 tools; a matchingskill playbook adds guidance 3 Find the right data GeoRAG matches datasets inthe Raven catalogue by meaning;provenance should favourcurrent, well-described data 4 Query the data tools run on the same area:climate indicators, land cover,LPIS crop suitability, POIs;flagged steps wait for approval 5 Reason and loop results return to the LLM,which may call further toolsuntil it judges that nothingmore is needed more tool calls 6 Answer with sources streamed text, tables, charts,references to datasets anddocuments; next step: clearerprovenance in every answer follow-up keeps the thread Every figure in the answer comes from a tool call on the drawn area, so it can be traced to a catalogue dataset and its provenance record. Sources: D4.2 §3.3.1–3.3.2, §4.4 · D4.1 §6.3
A user draws an area and asks a question; a router picks up to five of JackDaw's eighteen tools, GeoRAG finds matching datasets in the Raven catalogue, the tools run on the drawn area, and the LLM loops until no more tool calls are needed. The answer is streamed back with tables and references to the datasets used, and a follow-up question continues the same thread.

From data to decisions: one question and its evidence trail

main tellingdepth: standard3 min markdown

The weak point in GeoAI is rarely the generated text. It is whether a reader can tell where each number came from. The test case below is D7.2’s: a user asks whether a particular plot is suitable for potatoes.

One question, step by step

  1. Area. The user draws a rectangle around the plot; JackDaw attaches it to the message as WKT.
  2. Routing. One model call classifies the request and picks relevant tools, for example land-cover statistics (OLU_EU), elevation from a digital elevation model, and agroclimatic indicators such as frost occurrence and growing degree units from ERA5-Land.
  3. Approval. If a step is flagged as significant, an approval card shows its parameters and waits for the user to confirm.
  4. Tools run. Each tool computes its result inside its own service, for the drawn area only.
  5. Answer. The model explains the combined result in text; the underlying figures appear as tables and charts below it.

What makes the answer usable

D4.2 is explicit that a Regional Action Plan cannot rest on fluent text alone. Users need to know whether an output rests on official statistics, partner-provided pilot data, external web information, expert interpretation or experimental analysis. For the platform, D4.2 defines four levels of provenance:

LevelRecords
SourceOriginal creator, provider, ownership
IngestionWhen and how the resource entered the platform
ProcessingChanges after upload: reprojection, cleaning, enrichment
UsageHow services and pilot applications used it, where appropriate

Every dataset should also carry a quality status: validated, partially validated, experimental, external source, historical or superseded. D4.2 asks the system to prefer resources that are relevant, current, sufficiently described and valid for the requested territory and period; an experimental or outdated dataset should change how an answer is presented. Metadata should distinguish “unknown”, “not applicable” and “not provided”. Access is not reuse: a user may open a dataset whose licence still forbids redistribution, and derived maps inherit the conditions of their inputs.

Today and next

The Raven ingestion workflow already captures the basic provenance package (title, description, tags, temporal coverage, column descriptions, attached methodology). D4.2 also names the gap: keeping evidence trails transparent across several tool calls is hard, and better exposure of source provenance is a next step. GeoRAG and knowledge-graph work is expected to add source-to-answer traces toward the final platform. Until then, treat a JackDaw answer as screening evidence: keep its tables, name its datasets, and validate with people who know the plot before it enters a Regional Action Plan.

Where is that number from? (a comic)

another tellingexample world: czech-bavariadepth: intro1 min markdown

The scene is invented; the rules it follows are not. A Klatovsko council asks the question used as the worked example in open geodata for Klatovsko: how much farmland lies next to a Natura 2000 site?

JackDaw is drawn as a jackdaw because the PoliRural+ tools carry corvid names (JackDaw, Raven, Vulture, Magpie). The map in panel 3 is a sketch, not real parcel data.

Odkud je to číslo? (komiks)

another tellinglanguage: csexample world: czech-bavariadepth: intro1 min markdown

Scéna je vymyšlená, pravidla, podle kterých se odvíjí, ne. Zastupitelstvo na Klatovsku se ptá na totéž, co modelový příklad v bloku Co o Klatovsku řeknou otevřená data: kolik zemědělské půdy leží u lokality Natura 2000?

JackDaw je nakreslený jako kavka, protože nástroje PoliRural+ nesou jména krkavcovitých ptáků (JackDaw, Raven, Vulture, Magpie). Mapa v okénku 3 je náčrt, ne skutečná data o pozemcích.

Zdroje:

Check yourself: A JackDaw answer for the Action Plan

A regional planner drafting the agri-food part of a Regional Action Plan draws a rectangle around a group of plots in JackDaw and asks whether they are suitable for potatoes. The answer says “largely suitable” and shows tables of land-cover shares, elevation and frost and growing-degree indicators. The draft goes to stakeholders on Friday.

A) Paste the answer into the draft. It comes from the project’s own GeoAI tool and reads well.

B) Keep the tables and the datasets behind them, mark the result as screening evidence, and check it with an adviser and local data before the meeting.

C) Ask a general-purpose chatbot instead. It is faster and needs no map.

D) Leave AI out and commission a GIS specialist analysis. Bring it into the plan once it is ready.

More than one answer can be defended. Choose one, then read where it works and where it breaks.

Feedback on each option
  • A) This is where it breaks. D4.2 calls JackDaw a Beta whose answers are sensitive to metadata completeness, and says that keeping evidence trails transparent across several tool calls is still hard. A Regional Action Plan needs readers to see whether a claim rests on official statistics, pilot data or experimental analysis; pasted prose hides that.
  • B) The sound default. The tables under the answer name the tools and datasets (land cover, elevation, climate indicators), so you can record them with their dates and limits and label the result as screening evidence. Checking with an adviser and local data is what D4.2's provenance rules and D7.3's call to involve AKIS advisers ask for.
  • C) Faster, but weaker on exactly this question. D7.2's comparison found that a pure language model gives broad potato advice without the plot's climate, terrain or land cover, and D4.1 notes that LLMs struggle with lesser-known places. You would gain fluency and lose the place.
  • D) Defensible when the decision is binding or costly, such as land purchase or investment: a specialist analysis with documented methods is the right instrument there. For a first screening of options in a draft plan, it skips evidence you already have and puts off stakeholder discussion.

Play

Related

Sources