pyIfcLayer

Python scripts for IFC post-processing: layer assignment, cleanup, format conversion, and self-learning LLM classification.

Quick Start

# Windows (PowerShell / cmd)
.\.venv\Scripts\python.exe pyifclayer_llm.py           # Full run: layers + LLM learning + writes _clean.ifc
.\.venv\Scripts\python.exe pyifclayer_llm.py --dry-run # Test run: LLM classifies but writes nothing

# Or from venv
.venv\Scripts\python.exe pyifclayer_llm.py

Requirements: Python 3.10+, ifcopenshell, requests, python-dotenv (install via pip install -r requirements.txt or pip install ifcopenshell requests python-dotenv)

Environment: Create .env with LLM_API_KEY=sk-or-... (OpenRouter key for Nemotron 3 Nano)

---

Scripts Overview

Script	Purpose	Usage
pyifclayer_llm.py	Main pipeline: loads layers_rule.csv + layers_rule_learned.csv, assigns layers via rules, sends unassigned elements to LLM (Nemotron 3 Nano via OpenRouter), writes learned patterns to layers_rule_learned.csv, merges learned into base on real run.	python pyifclayer_llm.py [--dry-run]
pyifclayer.py	Core layer assignment only (no LLM). Reads layers_rule.csv, assigns IfcPresentationLayerAssignment by element type/name patterns. Writes <file>_clean.ifc.	python pyifclayer.py (processes all .ifc in cwd)
llm_classify_prototype.py	DRY-RUN prototype: loads rules, finds unassigned elements, calls LLM, prints results. Writes nothing. Good for testing prompt quality.	python llm_classify_prototype.py
pyArchIFC2NavisIFC.py	ArchiCAD → Navisworks IFC: strips storeys/layers/materials, groups elements into IfcElementAssembly by "Naviswork - уровни" pset, applies colors/coords from coords.txt/system.txt.	python pyArchIFC2NavisIFC.py
geonics.py	NanoCAD GeoniCS IFC → target: converts IfcSpatialZone/IfcValve/IfcPipeSegment to proxies, creates building/storey, migrates to IFC2X3.	python geonics.py
ifc_delim.py	Splits IFC by IfcPresentationLayerAssignment into separate files (one per layer). Draft (test.ifc).	python ifc_delim.py
stories.py	Collapses all IfcBuildingStorey into single base storey (level 0).	python stories.py
allplan_ifc.py	Text-based IFC fix for Allplan: IFCQUANTITY* → IFCPROPERTYSINGLEVALUE, IFCELEMENTQUANTITY → IFCPROPERTYSET. Works on raw strings.	python allplan_ifc.py
renga_ifc_compat.py	Converts IfcSpace zones from Tessellation to SweptSolid (slab extrusion) for Renga compatibility.	python renga_ifc_compat.py input.ifc output.ifc
batchifctonwd	Batch IFC→NWD via Navisworks FileToolsTaskRunner.exe. Paths hardcoded to C:\IFCTEMP.	python batchifctonwd

---

Key Files

File	Description
layers_rule.csv	Base layer rules (UTF-8/CP1251, ; delimiter). Columns: Layer Name;IfcType;Name;.... IfcType=All → fixed list; empty → all IfcProduct by name; specific type → exact match.
layers_rule_learned.csv	Learned patterns from LLM (appended to base = lower priority). Auto-managed.
.env	LLM_API_KEY=sk-or-... (OpenRouter). Not committed.

---

Layer Rule Format (layers_rule.csv)

Layer Name;IfcType;Name;Category;Workset;FamilyName
2_0_Колонны.КЖ;IfcColumn;;;;
2_0_Стены наружные.КЖ;IfcWall;;;;
5_1_Оборудование.ТХ;;.*приборы.*,.*рудование.*;;;
я82_Скрытый.КЖ;;.*скрытый.*;;;

• Col 1: Layer name (must match IFC layer name)
• Col 2 (IfcType): All = fixed list (IfcColumn,IfcSlab,IfcDiscreteAccessory,IfcBeam,IfcMember,IfcGrid); empty = search by name among ALL IfcProduct; specific type (e.g. IfcWall, ifcstair) = exact type match
• Col 3 (Name): Case-insensitive regex patterns for element.Name
• Col 4 (Category): Case-insensitive regex patterns for Category Pset property
• Col 5 (Workset): Case-insensitive regex patterns for Workset Pset property
• Col 6 (FamilyName): Case-insensitive regex patterns for FamilyName Pset property
• Matching logic: A rule matches if ANY of its non-empty pattern sets matches (Name OR Category OR Workset OR FamilyName)
• Priority = row order (first match wins). Learned rules appended at end = lowest priority

Note: Rows with empty IfcType AND all pattern columns empty are stored as "no-rule layers" for LLM validation only. They do NOT act as catch-all rules that would consume all subsequent rows.

---

LLM Self-Learning Loop (pyifclayer_llm.py)

layers_rule.csv + layers_rule_learned.csv
         │
         ▼
   set_layer()  ── typed rules ──▶ assigned
         │
         ▼
   set_layer()  ── allprod rules ──▶ assigned
         │
         ▼
   Unassigned elements ──▶ LLM resolver (Nemotron 3 Nano)
         │
         ├─▶ LLM picks layer FROM allowed list (validated client-side)
         └─▶ Client builds SAFE regex from element names (LCP + re.escape)
              │
              ▼
         layers_rule_learned.csv  (append only, no duplicates)
         │
         ▼ (on REAL run only, not --dry-run)
   merge_learned_into_base() → append to layers_rule.csv, clear learned

Safety: LLM only chooses layer name. Regex pattern built by client (prevents ReDoS/injection). Layer name validated against CSV list.

Rate limits: Free OpenRouter = 1000 req/day. CHUNK=10, sleep(6) between batches (~10 rpm). 429 → exponential backoff ×5.

---

Typical Workflow

1. Prepare: Place .ifc in folder, ensure layers_rule.csv exists, .env has LLM_API_KEY
2. Test: python pyifclayer_llm.py --dry-run — see LLM classification distribution, no files written
3. Run: python pyifclayer_llm.py — writes <file>_clean.ifc, updates layers_rule_learned.csv, merges learned into base
4. Iterate: Next run uses merged rules → fewer LLM calls

---

Notes

• All scripts work in-place on *.ifc in current directory, output <name>_clean.ifc
• pyifclayer_llm.py processes first .ifc found (not all)
• Encoding auto-detected: UTF-8 → CP1251 → CP1252 → Latin1
• Python: use python (3.11) for most, but py -3.10 with PYTHONPATH= for ifcopenshell/numpy compatibility
• LLM:free OpenRouter model: nvidia/nemotron-3-nano-30b-a3b:free (1000 req/day)