eng-drawing-extractor

name

eng-drawing-extractor

description

Extract structured engineering drawing data from documents — PDFs, images, markdown reports, specs, datasheets — and produce EDIS JSON (Engineering Drawing Interchange Schema) for the eng-schematic-renderer skill. Trigger on "extract drawing data", "pull dimensions from", "parse engineering specs", "get geometry from this document", "extract parts and dimensions", "prepare data for a schematic", "read this drawing", "extract BOM", "pull tolerances", or any request to convert engineering documents into structured data. Also trigger when a user uploads a technical PDF or engineering report and asks to "visualize", "draw", "make a schematic", or "render the geometry" — run this extractor first, then hand off to eng-schematic-renderer. Even casual "what does this part look like" or "draw this pipe" on an engineering document should trigger this skill as step one.

Engineering Drawing Extractor

Extract engineering content from documents into the EDIS (Engineering Drawing Interchange Schema) format — a structured JSON that captures parts, geometry, dimensions, materials, tolerances, and assembly relationships.

Before You Begin

1. Read the EDIS schema — references/edis-schema.md defines every field, type, and convention. Read it before extracting anything. The schema is the contract with downstream consumers (especially eng-schematic-renderer). Do not invent fields.

2. Read the extraction heuristics — references/extraction-rules.md contains domain-specific rules for identifying parts, dimensions, and views from different document types.

---

Core Workflow

Phase 1 — Classify the Source Document

Before extracting, determine what kind of document you're working with. This drives the extraction strategy.

Document Type	Primary Extraction Method	Expected Yield
Engineering report (Markdown, DOCX)	Parse tables, equations, and prose for dimensions, materials, tolerances	Parts + dimensions + materials; views must be inferred
Technical drawing (PDF with vector graphics)	Extract dimension callouts, title block, BOM, section labels	Full: parts + views + dimensions + annotations
Scanned drawing (PDF image / photo)	OCR + spatial analysis of text positions relative to geometry	Partial: depends on scan quality
Specification / datasheet	Parse tables and structured text for properties and limits	Parts + materials + tolerances; geometry may be minimal
3D model export (STEP, IGES metadata)	Extract feature tree, dimensions from metadata	Parts + geometry summary; no 2D views

If the document type is ambiguous, state your classification and confidence before proceeding.

Phase 2 — Extract Parts

Identify every distinct physical component mentioned in the document. For each part:

1. Name it — use the document's own terminology. Do not rename.
2. Assign a part_id — short, snake_case, unique. E.g., steel_host_pipe, hdpe_liner, end_ring.
3. Extract material — look for:
   • Explicit material callouts ("API 5L X52", "PE100")
   • Material property tables (E, ν, α, σ_y)
   • Process-implied materials ("swaged HDPE")
4. Extract geometry summary — OD, ID, WT, length, shape type.
5. Log the source — where in the document each value came from.

Rules:

• If a value is computed (e.g., ID = OD − 2×WT), mark the source as "computed: OD - 2×WT = <value>" and trace the inputs.
• If a value is assumed (e.g., WT unchanged after swaging), mark it as "assumed: <reason>" and set confidence to low or medium.
• If a value is missing, omit the field — do not guess.

Phase 3 — Extract or Infer Views

Engineering documents may contain explicit views (section drawings, detail callouts) or may require you to infer appropriate views from the described geometry.

Explicit views: Extract them directly. Look for:

• Section labels ("Section A-A", "Detail B")
• Scale callouts
• Projection indicators (first-angle, third-angle)
• Coordinate system references

Inferred views (from reports/specs): Generate logical views based on the part geometry:

Geometry Type	Default Views to Generate
Concentric cylinders (pipe-in-pipe)	Cross-section (radial) + longitudinal section
Plate / flat part	Plan view + edge view
Ring / annular part	Cross-section + plan view
Complex assembly	Exploded view + assembled cross-section

For each view:

1. Define the coordinate system — origin, axes, and what they mean physically.
2. Generate geometry entities — translate dimensions into drawable primitives (lines, arcs, circles, rectangles).
3. Assign roles — outline, centerline, hidden line, hatch.
4. Position entities — use extracted dimensions to compute coordinates. All coordinates should be in the document's unit system.

Phase 4 — Extract Dimensions and Tolerances

For every dimensional value found in the document:

1. Classify it — linear, diameter, radius, angular, etc.
2. Record the numeric value and unit.
3. Extract any tolerance — bilateral (±), unilateral, limit, fit class (H7/p6), GD&T frame.
4. Link it to geometry — which entity or entities does this dimension span?
5. Preserve the label — record the dimension as written on the drawing (e.g., "∅155.575", "6.35 ±0.05").

Phase 5 — Extract Annotations, BOM, and Notes

• Annotations: GD&T frames, weld symbols, surface finish callouts, balloon numbers, flags.
• BOM: Item numbers, part references, quantities, material specs.
• Notes: General notes, process notes, inspection requirements, safety callouts.

Classify each note by category (general, process, inspection, material, tolerance, assembly, safety).

Phase 6 — Assess Completeness

Before producing the final EDIS JSON:

1. Count what you extracted vs. what you expected to find.
2. List known gaps — dimensions you couldn't find, views you couldn't infer, parts mentioned but not dimensioned.
3. Rate overall confidence — high / medium / low.
4. Log unknowns — anything detected but not classifiable.

---

Coordinate Generation Rules

When generating entity coordinates for inferred views (Phase 3), follow these conventions:

Cross-Section View (Radial Cut)

For concentric cylinders, draw from outside in:

Origin: geometric center of outermost cylinder
X-axis: horizontal (left = negative)
Y-axis: vertical (up = positive)

Outer cylinder: circle at origin, radius = OD_outer/2
Inner surface of outer cylinder: circle at origin, radius = ID_outer/2
Outer surface of inner cylinder: circle at origin, radius = OD_inner/2
Inner surface of inner cylinder: circle at origin, radius = ID_inner/2

Apply hatching between inner and outer surfaces of each part (different hatch patterns per material).

Longitudinal Section View

For pipe-in-pipe assemblies:

Origin: left end of assembly, at centerline
X-axis: axial direction (right = positive)
Y-axis: radial direction (up = positive from centerline)

Draw upper half only (symmetric about centerline), then mirror.
Centerline: horizontal dashed line at y=0.

Each wall appears as a rectangle:

• Outer pipe upper wall: rectangle from (0, ID_host/2) to (L, OD_host/2)
• Liner upper wall: rectangle from (0, ID_liner/2) to (L, OD_liner/2)

Dimension Placement

• Diameters: Place outside the largest circle, with leader lines.
• Wall thicknesses: Dimension between adjacent surfaces.
• Radial interference (δ): Show as a detail with exaggerated gap.
• Axial dimensions: Place above or below the longitudinal view.

---

Output

Produce a single EDIS JSON file. Save it as:

/home/claude/edis_output.json

Then either:

• Present it to the user for review, or
• Hand it off to eng-schematic-renderer if the user asked for a visual output.

Handoff Protocol

When handing off to the renderer, provide:

1. The EDIS JSON file path.
2. A brief summary: how many parts, how many views, confidence level.
3. Any renderer hints:
   • Preferred view to render first
   • Whether to show dimensions
   • Whether to show hatching
   • Scale preference

---

Domain-Specific Extraction Rules

Pipe-in-Pipe / Lined Pipe Systems

These are the most common use case (interference fits, swaged liners, CRA-lined pipe). Special rules:

• Interference: Compute δ, ΔD, r from extracted diameters. Show these as derived dimensions in the EDIS.
• Pre/post process states: If the document describes a part in multiple states (pre-swage vs. post-reversion), create separate geometry summaries or dimension sets, labeled by state.
• Contact interface: The interface between liner and host is a critical feature. Mark it explicitly as an annotation.

Pressure Vessels / Cylindrical Shells

• Extract design pressure, MAWP, hydrostatic test pressure.
• Identify longitudinal and circumferential stress directions.
• Look for corrosion allowance dimensions.

Structural Steel

• Extract section designations (W12×26, HSS 6×4×3/8).
• Connection details: bolt patterns, weld callouts, plate sizes.
• Identify load paths from member annotations.

Flanges and Fittings

• Extract flange class, face type, bolt circle diameter.
• Gasket dimensions and material.
• Bore/hub dimensions.

---

Error Handling

Situation	Action
Document is unreadable (corrupt PDF, blank image)	Report failure. Do not fabricate data.
Dimensions conflict (OD in text ≠ OD in table)	Extract both, flag conflict in extraction_notes.
Units ambiguous (bare "6.625" — inches or mm?)	Infer from context (API 5L → inches). State assumption.
Part referenced but never dimensioned	Create part entry with geometry_summary: null. List in missing.
Document contains multiple independent assemblies	Create separate EDIS instances or clearly partition views by assembly.

---

Reminder

The EDIS schema in references/edis-schema.md is the authoritative contract. Every field you produce must conform to it. Every value must carry a source annotation. Never fabricate dimensions — if you cannot extract a value, omit it and log the gap.