fix: prevent truncated Ollama answers

2026-07-15 07:39:03 +02:00
parent 5be8f628ef
commit 56f823f02a
13 changed files with 89 additions and 4 deletions
@@ -8389,7 +8389,7 @@ Validation evidence:
 - Focused temporal, source, Ollama, navigation and Unraid regression tests
  passed, including direct coverage for multi-metric history and DockerMan host
  mapping.
- Full readiness passed 615 backend tests, backend compilation, 91 documented
+- Full readiness passed 617 backend tests, backend compilation, 91 documented
  API routes, one Alembic head, frontend TypeScript typecheck/build and all
  shell syntax gates.
 - Tower deployment passed the all-in-one container health check, live PostGIS
@@ -8403,6 +8403,15 @@ Validation evidence:
  message keys now use a session-local monotonic id generator; no persisted or
  security-sensitive identity depends on it. Typecheck, production build and
  full readiness reran successfully.
+- The regional operator then reused the five governed forest snapshots and
+  imported 15 new authoritative snapshots: water, built functions and
+  transport for 2013, 2016, 2019, 2022 and 2025. A live exact-Area Mol water
+  comparison returned 606.16 ha in 2013 and 641.75 ha in 2025 across all five
+  observations; the API retained the 10 m raster and object-lineage warnings.
+- Browser QA also reproduced an Ollama `done_reason=length`: the previous
+  implicit 4096-token context left only 86 tokens after a 4010-token grounded
+  prompt. GeoIntel now requests a configurable 16,384-token context and rejects
+  any future length-truncated response instead of presenting a partial answer.

 Known limitations:
 - Water volume remains unavailable until a governed depth/bathymetry source is