Construction Plan Intelligence

Construction Plan Intelligence Skill

Sharpen ForemanOS's existing extraction pipeline for construction plan PDFs. The infrastructure exists — this skill addresses 10 specific gaps in extraction quality, cross-referencing, takeoff accuracy, and RAG retrieval.

Codebase Map

Assets

Symbol Library Usage

The symbol library (construction_symbols_library.json) serves three roles:

1. Vision prompt injection (Gap 1): During two-pass extraction, after the lightweight classification determines discipline, load only the relevant division's symbols into the main extraction prompt. E.g., mechanical plan gets Division 23 symbols (17 entries) + Common Hatch Patterns, not all 230. Each entry has a vision_hints field written specifically for guiding the vision model on what to look for visually.

2. Symbol learner baseline (Gap 10): Load as initial vocabulary in symbol-learner.ts before runtime pattern matching begins. Symbols matched against this library receive +25 confidence boost. Cross-validate against SheetLegend.legendEntries for per-project ground truth.

3. Assembly resolution (Gap 5): The CSI division structure maps directly to the MasterFormat keynote resolution chain. Symbol IDs use D{division}-{category}-{seq} format enabling programmatic lookup by division number.

Division-to-discipline mapping for prompt injection:

Processing Flow (Current)

Upload → S3 → processDocument() → classifyDocument()
  → PDF? → processDocumentBatch() [per-page]:
      convertPdfToImages() → getVisionPrompt() [GENERIC] → vision API → parseJSON
      → formatVisionData() → store DocumentChunk
  → Post-processing:
      runIntelligenceExtraction(phases: ['A','B','C'])
        Phase A: title blocks, scales, legends, drawing classification
        Phase B: dimensions, annotations, callouts, rooms
        Phase C: spatial correlation, MEP tracing, symbol learning
  → extractRoomData(), extractScheduleData(), syncFeatures()

10 Gaps — Implementation Guide

Gap 1: Discipline-Specific Vision Prompts

Problem: getVisionPrompt() (line 251, document-processor-batch.ts) sends identical instructions for every page type. Structural foundation plans get asked for rooms and doors.

Fix: Two-pass extraction. Before the main vision call, run a lightweight classification call (~$0.0002/page) to determine discipline and drawing type, then select the appropriate prompt template.

Integration point: Inside the per-page loop in processDocumentBatch(), before the main analyzeWithSmartRouting() call.

Reference: Read references/discipline_prompts.md for all prompt templates. Load relevant symbol subsets from assets/construction_symbols_library.json using the division-to-discipline mapping in the Assets section above.

Gap 2: Cross-Reference Resolution

Problem: DocumentChunk.crossReferences (Json?) is populated but never resolved. Detail bubble "3/A-501" on Sheet A-201 is stored as text, not linked to Sheet A-501.

Fix: Post-processing step after Phase B (after line 414 in intelligence-orchestrator.ts) that builds a project-level adjacency map. New DrawingCrossReference Prisma model.

RAG enhancement: In retrieveRelevantDocuments(), after initial scoring, check high-scoring chunks' cross-references and pull in referenced chunks with +40 score boost.

Reference: Read references/cross_reference_patterns.md for pattern formats and resolution logic. Read references/schema_migrations.md for the new Prisma model.

Gap 3: Material Takeoff Formulas

Problem: takeoff-memory-service.ts reads/formats DB data but has no calculation logic connecting extracted dimensions to material quantities.

Fix: Add formula engine that takes extracted spatial data (dimensions, rooms, counts from DocumentChunk) and produces TakeoffLineItem records with show-your-work math.

Integration point: New function called after Phase B extraction completes, or on-demand when takeoff context is requested.

Reference: Read references/takeoff_formulas.md for all formulas, waste factors, and sanity checks.

Gap 4: RAG Sheet-Type Routing

Problem: detectQueryIntent() (line 487, rag.ts) has 6 broad categories. "Ceiling height in Room 201" and "foundation depth at grid C-3" both get the same plans_drawings intent with no sheet-type discrimination.

Fix: Add sub-classification within plans_drawings that maps query content to specific sheet types, then apply discipline multiplier in calculateRelevanceScore().

Integration point: Extend detectQueryIntent() to return sub-intents. Add sheet-type boost logic after applyCategoryBoost() at line 439.

Reference: Read references/query_routing.md for the complete mapping table.

Gap 5: Assembly Identification

Problem: System extracts individual elements but doesn't understand that wall type "WT-1" is an assembly of studs + gypsum + insulation + cladding.

Fix: Post-processing that recognizes assembly identifiers (wall types, ceiling types, floor assemblies) and follows references to schedules/legends for component lists. Enables accurate multi-material takeoffs from a single wall dimension.

Reference: Read references/assembly_patterns.md for assembly types and component resolution logic.

Gap 6: Multi-Page Drawing Continuity

Problem: Large floor plans span multiple sheets (A-101, A-102) with match lines. System processes each independently — room counts and dimensions may be incomplete.

Fix: Post-processing step after all sheets are extracted that detects continuation indicators and groups spatially adjacent sheets. New SheetContinuity model. RAG enhancement: when any sheet in a continuation group scores high, pull all group members.

Reference: Read references/continuity_patterns.md for detection logic. Read references/schema_migrations.md for the new Prisma model.

Gap 7: Revision Delta Tracking

Problem: DocumentChunk.revision field exists, vision prompt mentions revision clouds, but no comparison logic exists between document versions.

Fix: When a new plan set is uploaded, compare against previous version's chunks by sheet number. Flag chunks where content differs significantly. Store revision metadata for RAG boosting when users ask "what changed."

Reference: Read references/revision_tracking.md for detection and comparison logic.

Gap 8: Multi-Scale Context Association

Problem: DocumentChunk has hasMultipleScales and scaleData but the vision prompt doesn't ask the model to associate specific scales with specific extracted elements.

Fix: Enhance the discipline-specific prompts (Gap 1) to request per-element scale association. Add validation in Phase A that checks dimension values against their associated scale for plausibility.

Reference: Addressed within references/discipline_prompts.md (scale extraction instructions) and references/validation_rules.md (scale plausibility checks).

Gap 9: Specification-Drawing Linkage

Problem: No explicit linking between spec documents and drawing keynotes. RAG gives a generic 1.2x related-category boost between plans_drawings and specifications.

Fix: Build CSI MasterFormat mapping so keynote "09 21 16" extracted from a drawing automatically links to the corresponding spec section. Enhance RAG to follow these links during retrieval.

Reference: Read references/assembly_patterns.md (keynote-to-CSI section) for the MasterFormat mapping.

Gap 10: Symbol-Legend Validation

Problem: symbol-learner.ts runs regex patterns against chunk text but never validates matches against SheetLegend.legendEntries — the architect's ground truth.

Fix: After learnFromDocument() runs, cross-check extracted symbols against SheetLegend entries for the same sheet. Confirmed matches get confidence boost. Contradictions get flagged. Unmatched legend entries become new learned symbols.

Integration point: Modify learnFromDocument() (line 182, symbol-learner.ts) to query SheetLegend for the same documentId after initial regex extraction.

Reference: Load baseline vocabulary from assets/construction_symbols_library.json before runtime matching. Logic described inline above — the fix is a 30 line addition to the existing function plus a library loader (50 lines).

Implementation Priority

Execute in this order — each gap builds on previous ones:

1. Gap 1 (Discipline prompts) — Foundation for everything else. Better extraction data improves all downstream processing.
2. Gap 8 (Multi-scale context) — Included in Gap 1's prompt redesign.
3. Gap 10 (Symbol-legend validation) — Quick win, ~30 lines.
4. Gap 2 (Cross-reference resolution) — Requires schema migration.
5. Gap 6 (Multi-page continuity) — Requires schema migration, pairs with Gap 2.
6. Gap 4 (RAG sheet-type routing) — Benefits from Gap 1's better classification data.
7. Gap 7 (Revision tracking) — Independent, can be done anytime.
8. Gap 5 (Assembly identification) — Depends on Gaps 2 and 4.
9. Gap 9 (Spec-drawing linkage) — Depends on Gap 5's assembly work.
10. Gap 3 (Takeoff formulas) — Last because it benefits from all other improvements.

Schema Changes Required

Read references/schema_migrations.md before making any Prisma schema changes. Two new models needed: DrawingCrossReference and SheetContinuity.