count-engineering-drawing-symbols

TakeoffLens Drawing Audit

Use the takeoff-lens MCP tools. Treat every automatic detection as a candidate until it is visually confirmed.

Workflow

1. Call get_discipline_catalog to distinguish active from planned disciplines. Then call get_symbol_rules with active system_id=POWER, LIGHTING, FIRE_ALARM, DATA_VOICE, or CCTV_SECURITY to lock the v0.2.x scope. Never imply that a planned discipline is already supported.
2. Prefer DWG/DXF block, attribute, layer, and XREF extraction when CAD is available. For this v0.2.x MCP, process vector/hybrid PDF directly.
3. On a clean vector sheet with meaningful equipment layers, call analyze_vector_layers first. Group by rotation-normalized dimensions, then apply only a user- or legend-confirmed project mapping. Save that mapping and reuse signature_mapping_path; never infer equipment class from a frequent shape alone.
4. Call prepare_sheet_audit. It builds one native shared sheet context, profiles the page, renders an overview, detects every symbol with a ready template, and creates one candidate contact sheet. Use the returned manifest instead of repeating separate inspection calls.
5. Read references/symbol-rules.md before classification. Every class without a bundled starter requires either an unambiguous confirmed layer signature or a clean project-specific template from the current legend; build only templates listed in template_required, pass its exact symbol_id to build_symbol_template, then rerun prepare_sheet_audit.
6. Keep MCP responses at the default response_detail=compact. Read the detailed JSON artifact only when coordinates, diagnostics, or ambiguity review are needed. Use response_detail=full for debugging, not routine turns.
7. Use detect_symbol_candidates separately only for tuning or rerunning one symbol. Keep a high-recall shortlist.
8. Inspect every crop and markup. Also inspect filtered_candidates.json when Candidate Filtering v3 suppresses matches. Classify indoor and explicit WP variants from the crop and project legend; do not assume extracted PDF text contains a visible WP glyph. Independently sweep all wall faces, wall corners, both sides of doors, rotated symbols, and paired power/data symbols; do not treat a low candidate count as evidence of zero.
9. Reject candidates in legends, notes, details, title blocks, and unrelated systems.
10. Add visually verified misses as manual_points; do not silently change the final number.
11. Call confirm_symbol_count with every candidate classified in accepted_ids, rejected_ids, or uncertain_ids; pass the matching detection_manifest_json, include floor_or_region, and set wall_door_sweep_completed only after the independent sweep. A final result requires verified PDF, template, and candidate hashes.
12. If a symbol, legend mapping, boundary, or candidate decision is unknown or not sufficiently supported, keep it uncertain, record its page and coordinates, create a crop or markup, and ask the user a concise verification question. Group related uncertain points in one question when practical.
13. Continue reviewing independent clear items while waiting for clarification, but do not guess, silently exclude the item, or finalize any affected subtotal.
14. After the user answers, record the decision and rerun confirmation. Treat the result as final only when review_complete is true, clarification_required is false, and provenance_verified is true. Report count by page/floor and include final markup and CSV/JSON paths.
15. For detector tuning, call evaluate_detection_accuracy only with ground truth whose review is complete and whose PDF-point coordinates were visually verified. Compare shortlist recall with candidate-pool recall to identify filtering and shortlist-limit misses.