Civil Dwg Processing

Civil DWG Processing

Overview

Process civil engineering DWG/DXF files through a standardized 7-stage pipeline: file reception, deterministic cleaning, feature classification, data extraction, validation, provenance tagging, and integration handoff. This ensures geometry-derived data is accurate, structured, and safe for downstream AI processing.

Announce at start: "I'm using the civil-dwg-processing skill to extract structured data from CAD drawing files."

When to Use This Skill

Trigger Conditions:

• Incoming civil engineering drawings (DWG/DXF) require data extraction
• DWG geometry needs analysis or conversion to structured data (CSV/JSON)
• Drawing-to-document consistency checking is required
• Quantity takeoff from CAD drawings is needed
• Clash detection between disciplines requires model extraction

Prerequisites:

• DWG/DXF file available with revision metadata
• Extraction schema defined for the drawing type (site plan, drainage layout, road alignment, structural GA, etc.)
• CE-DWG-001 agent or equivalent CAD processing capability available
• Target data schema ready to receive extracted features

Step-by-Step Procedure

Step 1: File Reception & Pre-Check

Verify input integrity before processing:

1. Confirm DWG/DXF file is readable (not corrupted)
2. Verify revision number matches expected version
3. Check file size is within processing limits (<500MB recommended)
4. Identify drawing type from title block or filename convention
5. Confirm required extraction schema exists for this drawing type

Pre-check checklist:

• File opens in CAD viewer without errors
• Drawing units confirmed (metric/imperial)
• Coordinate system identified
• Title block present with revision info
• Relevant layers identified
• No unresolved Xrefs

If any check fails: Flag for human CAD review, do NOT proceed to extraction.

Step 2: Deterministic DWG Cleaning (7-Stage Pipeline)

Run these cleaning stages in order. Each stage must complete before the next:

Stage 1 — AUDIT: Run CAD audit/repair to fix database errors and corrupt entities.

Stage 2 — PURGE (Pass 1): Remove unused layers, blocks, text styles, linetypes, regapps.

Stage 3 — OVERKILL: Merge overlapping/duplicate lines, arcs, polylines; delete zero-length or tiny stray objects.

Stage 4 — PURGE (Pass 2): Re-purge after overkill to catch newly orphaned entries.

Stage 5 — EXTENTS FIX: Zoom to extents, identify and remove far-away ghost objects that corrupt bounding boxes.

Stage 6 — XREF CLEAN: Bind or detach unnecessary Xrefs; resolve missing references or flag for human review.

Stage 7 — LAYER STANDARDIZATION: Map incoming layers to project standard naming (e.g., "WALLS" → "A-WALL", "DRAIN" → "C-DRAIN").

Post-cleaning verification:

• File size reduced by ≥20% (typical)
• No audit errors remaining
• Bounding box matches expected drawing extents
• Layer mapping completed successfully

Step 3: Feature Classification & Extraction

Classify drawing elements into the extraction schema:

Classification rules:

• Only extract from layers matching the schema's target layers
• Ignore presentation-only content (hatches, image references, viewport boundaries)
• Flag elements with low classification confidence (<80%) for human review

Step 4: Data Structuring (CSV/JSON Output)

Transform extracted features into structured data:

{
  "extraction_metadata": {
    "source_file": "original_filename.dwg",
    "revision": "C",
    "extraction_date": "2026-03-31",
    "drawing_type": "site_drainage_plan",
    "coordinate_system": "Local Grid"
  },
  "features": {
    "drainage_pipes": [
      {
        "tag": "P-001",
        "start_node": {"x": 100.0, "y": 200.0},
        "end_node": {"x": 150.0, "y": 225.0},
        "diameter_mm": 300,
        "gradient": "1:100",
        "invert_level_start": 52.350,
        "invert_level_end": 52.250,
        "provenance": "layer=C-DRAIN-PIPE, entity=LWPOLYLINE, id=4521"
      }
    ],
    "manholes": [
      {
        "tag": "MH-01",
        "location": {"x": 100.0, "y": 200.0},
        "cover_level": 53.500,
        "invert_level": 52.350,
        "provenance": "layer=C-DRAIN-STRUCT, entity=BLOCK, name=MH_STD_1200"
      }
    ]
  },
  "validation": {
    "total_features": 245,
    "flagged_for_review": 3,
    "confidence_score": 0.96
  }
}

Step 5: Validation & Range Checking

Run validation rules against extracted data:

If any validation fails: Do NOT pass to LLM layer. Flag specific failures for human CAD review.

Step 6: Provenance Tagging

Attach provenance metadata to every extracted feature:

• Source drawing filename and revision
• Original entity type (LWPOLYLINE, BLOCK, TEXT, etc.)
• Original entity ID in DWG
• Layer name before standardization
• Extraction timestamp

This enables click-through from extracted data back to the source drawing element.

Step 7: Integration Handoff

Pass validated, structured data to downstream consumers:

• Quantity takeoff agents → BOQ generation
• Design consistency checkers → Drawing-to-spec comparison
• BIM coordinators → Clash detection
• Documentation agents → Drawing schedules and references

Handoff payload: Structured JSON with metadata, features, validation results, and provenance tags.

Success Criteria

• All 7 cleaning stages completed without errors
• Extraction schema matched drawing type successfully
• All features classified with ≥80% confidence (or flagged)
• Validation rules passed (or failures flagged and isolated)
• Provenance tags attached to all extracted features
• Structured JSON output generated and delivered to downstream agents
• LLM layer received only validated, structured data — never raw geometry

Common Pitfalls

1. Skipping DWG Cleaning — Never feed raw/un-cleaned DWG files into extraction. Corrupted entities and junk layers cause extraction errors.
2. Assuming Geometry = Truth — DWG geometry must be validated before use. Drawing errors exist; do not let LLM "invent" values to fill gaps.
3. Ignoring Provenance — Without provenance tags, reviewers cannot trace extracted values back to source elements. Always attach source metadata.
4. Over-Extraction — Only extract what the schema defines. Do not extract everything in the file; this creates noise and false positives.
5. Missing Xref Dependencies — If the drawing references external files that are unresolved, extraction may be incomplete. Verify Xrefs before proceeding.

Cross-References

Related Skills

• civil-quantity-takeoff — Use extracted data for BOQ generation
• civil-site-assessment — DWG processing for survey data
• civil-design-development — Extraction feeds design calculations
• pre-task-assessment-readiness — Assess readiness before starting extraction

Related Agents

• CE-DWG-001 (deep-agents) — Primary DWG processing agent with KIMI Vision Analysis
• CE-LIB-001 (OpenClaw) — Knowledge management and provenance tracking
• CE-QA-001 (OpenClaw) — Quality validation of extraction results

Example Usage

Scenario: Extract drainage network from site drainage plan (Revision C)

1. Reception: File DWG-0043-REV-C.dwg received, 45MB, drainage layout
2. Cleaning: 7-stage pipeline completed, file reduced to 32MB, 0 audit errors
3. Extraction: 12 drainage pipes, 8 manholes, 4 catch basins, 3 outfalls
4. Validation: All topology checks passed, 1 pipe gradient at 0.08% flagged (marginally below 0.1% minimum)
5. Provenance: All 27 features tagged with source layer, entity ID, and extraction timestamp
6. Handoff: Structured JSON passed to quantity takeoff agent for pipe length schedules

Performance Metrics

Target Performance:

• Processing time: <30 seconds for files up to 100MB
• Extraction accuracy: >95% feature recognition
• Validation pass rate: >98% of features within expected ranges
• False positive rate: <5% of extracted features flagged incorrectly