[LionelZoubritzky-IGN]

Description

Reject any query that requires a too-big CQL filter ("too big" is at 2 MiB, checked experimentally).

Related issues (if applicable)

Close #151, close #215.

Motivation

Too big filters lead to timeout or to silently wrong answers (see #215).

Implementation

Check the size of the intersect_feature_filter and reject it if too large with a clear error message. Also guard against too large CQL filters in general (which could result from a too-large where clause as well).

The filter never silently switches to the -CARTO-PE version because the id can change: departement.72 does not mean the same thing in ADMINEXPRESS-COG.LATEST and in ADMINEXPRESS-COG-CARTO-PE.LATEST. However, the error message gives a hint to the LLM that it should use the -CARTO-PE version.

Testing

npm run build && npm run verify: 611+4+35 tests green and "how many buildings in Vendée" correctly returns around 840000.

Checklist

• The PR is focused and of a reasonable size.
• The commit history is clean.
• Relevant documentation has been updated.
• Relevant tests have been added or updated.

[mborne]

Tested with ignfab/geocontext-test test_count_batiment_vendee.py (claude-haiku-4-5, local server on this branch):

• The oversized ADMINEXPRESS-COG.LATEST:departement geometry is now rejected as expected (3803688 octets de filtre encodé (limite : 2097152)).
• But with the original question ("Combien y a-t-il de bâtiments dans le département de la Vendée ?"), the model follows the "ou utiliser un filtre bbox" hint: it fetches the department bbox, counts with bbox_filter and answers 1 111 470 buildings (neighbouring departments included) without any warning.
• When asking for a precise count (ignfab/geocontext-test@a8d1b8d), the model still tries the bbox first, then switches to ADMINEXPRESS-COG-CARTO-PE.LATEST:departement and answers 854 362 (2/2 runs).

Suggestion: drop the bbox alternative from the error message (or state that it overcounts and must not be used for counting within an administrative unit), and maybe hint how to find the feature in the CARTO-PE layer (e.g. where code_insee = '85') since ids may differ.

🤖 Generated with Claude Code