The wet-area engine already treats kitchen, baths, WC, laundry, wash and pantry
as wet (keyword list), but the arch room-extraction prompt only said 'list every
room' - so the AI sometimes skipped small wet rooms (kitchen/laundry), under-
counting the ground-floor wet area. The prompt now explicitly requires every wet
room (kitchen, all baths, WC, powder, laundry/wash, pantry) with its dimensions.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
fix(qto): line-weight filter in the vector wall reader (root over-count fix)
extract_arch_vectors treated EVERY straight edge as a candidate wall face, so
furniture, fixtures, hatching, site-plan and dimension lines were counted as
walls — inflating internal wall length ~10× (a floor plan read 1330 m instead
of ~120 m). Keep only the heavier wall-weight lines (with a safe fallback to all
lines when there's no clear weight distinction). On real drawings this brought
per-floor readings from 1330 m → 120 m (correct). The floor-area clamp remains
as a second safety net.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New pdf_engine/vector_intelligence.py — every helper is additive and falls back
to the existing behaviour, so a messy drawing can never make output worse:
1. Self-calibrating scale from the drawing's own dimension chains — used ONLY
when the sheet has no "1:N" label (never overrides a real label), replacing
the blind default of 1:100. Wired into detect_page_scale (both vector modules).
2. Real closed-polygon area (shoelace) instead of a bounding box → measure_
architectural_bounds now returns gf_area when a convincing footprint is found.
3. Line-weight-filtered building bounds (drops dimension/grid lines that inflate
a raw bbox); used for length/width/perimeter with the bbox as fallback.
4. Vector geometry handed to the vision model as grounding context (appended to
the deterministic injection in extract_page) so it reasons over real numbers.
5. Reconciliation of vector vs bbox into a _confidence flag on the result.
6. Scale-bar reader scaffold (kept inert/None until DP…
Update UI, add feedback system, extract name from email
perf: eliminate Step 3 hangs by shrinking max PNG DPI to 50 for all blueprints, preventing massive network timeouts when sending images to Gemini. Increased concurrent extraction workers to 8 to slash processing time.
fix: entirely bypass massive Numpy/PIL conversions in Step 3 AI Extraction by feeding raw cached images directly to Gemini. Fixes OOM memory thrashing freezing at 5%
Fix extraction speed and rotation: unconditionally use KeyManager to rotate through all available API keys instead of locking onto a single env variable