Floor-Plan PDF Extractor

A small, runnable pipeline that reads an architectural floor-plan PDF and extracts the parts a downstream system needs — title-block metadata, per-room stamps (name / number / floor area), and the fixtures table — into structured JSON + Excel, with a validation layer that flags low-confidence values for manual review instead of letting them enter the pipeline as fact.

Honest scope, up front. The demo runs on a synthetic sample plan I generate (src/make_sample.py), so it has something to run on without any client's confidential drawings. It shows the method — layout-aware parsing, a defined output schema, and confidence flagging — on the native-text (vector PDF) path. Real plans vary (different CAD exports, layer names, stamp formats, scanned pages); the pipeline is built to be pointed at yours, and the parsing rules are configured per document family. See "On real plans".

Run it

python -m venv .venv && . .venv/bin/activate      # (Windows: .venv\Scripts\activate)
pip install -r requirements.txt
python src/run.py

Output (console + output/output.json + output/output.xlsx):

Rooms (5):
  nr  name            area m²  review
  01  Wohnzimmer         24.5     ok
  02  Küche              12.8     ok
  03  Bad                 6.2     ok
  04  Schlafzimmer       18.0     ok
  05  Abstellraum           —   ⚠ area_unparsable    ← routed to manual review
Document needs_review: True

The last room's area is deliberately broken (BF: -- m²) to show the validation layer doing its job: a value it cannot trust is flagged, not guessed.

How it maps to a real extraction job

Job requirement	Where it lives
Native-text and scanned (OCR) ingestion	vector path in extract.py; OCR path described below
Text + table extraction, German field names	extract.py (_title_block, _rooms, _fixtures)
Layout-aware parsing (values ↔ correct room)	line clustering + column split, so side-by-side stamps don't merge
Structured output (JSON schema + Excel)	run.py → output.json, output.xlsx
Validation layer flagging low-confidence	per-value flags + document needs_review
CLI to run the pipeline	python src/run.py

Low-confidence & scanned plans (the strategy)

The core discipline: a value the pipeline is not sure about must be flagged for review, never silently emitted. Concretely:

• Native-text plans: values come straight from the PDF text layer (high confidence). Range/format checks (e.g. a floor area outside 1–200 m², an unparsable stamp) still route to review.
• Scanned plans: the same parsers run on OCR output, and each value carries the OCR confidence. Below a threshold → needs_review. Two levers raise accuracy before flagging: (1) fuse OCR with any residual vector text when a plan is hybrid; (2) cross-check a value against its neighbours (a room area that contradicts the fixtures table, a dimension that breaks the plan scale). What can't be made confident is handed to a human with the crop that produced it — cheap to verify, never wrong-by-guess.

This is the same accuracy-by-fusion + confidence-gating approach behind a table recognition system I took from 55% to 99.5% on hard documents.

On real plans

Point the pipeline at your documents and configure per family:

• stamp patterns — _rooms matches Raum Nr <n> + BF: <area> m²; real offices use different stamp formats, added as patterns / a small abbreviation map.
• layers — when a CAD export has named layers, they disambiguate rooms; when it doesn't, geometry + stamp proximity carry it.
• schema — output.json is one shape; align it to your JSON schema and the Excel columns follow.

Files

• src/make_sample.py — generates the synthetic sample plan
• src/extract.py — layout-aware extraction + confidence flagging
• src/run.py — one-command: PDF → JSON + Excel + review summary

License

MIT — see LICENSE.