Skip to content

Technical Document Module

This page translates the Forward research into a demo-ready Rapid Craft module for the meeting on Thursday, May 14, 2026.

The right story for Forward

Forward does not need a generic "chat with documents" pitch.

The pattern across the current research is much more specific:

  • battery-enclosure programs generate repeated engineering knowledge,
  • material-card and validation evidence are hard to track and reuse,
  • regulations like GB 38031-2025 increase documentation burden,
  • repetitive CAE glue work is a real cost center,
  • old project files contain reusable know-how but are hard to search.

So the demo should feel like an engineering knowledge system with source traceability, not a consumer-style PDF chatbot.

What to demo

The cleanest demo story is:

  1. Upload a mixed set of sources: engineering PDF, PowerPoint, design note, and validation report.
  2. Ingest them into a technical wiki that preserves source, page, section, and document type.
  3. Ask focused engineering questions such as: Where do we discuss material-card calibration? Which pages mention bottom impact, rocker integration, or thermal propagation? Show me the regulation-related requirements and the tests they affect.
  4. Show generated wiki pages or topic pages such as: Material cards Battery enclosure validation GB 38031 readiness Crash and load-path questions
  5. Show that the same ingestion layer can later feed workflow automation: report generation, requirements traceability, test-correlation notes, and CAE setup reuse.

The fastest believable Rapid Craft architecture is four layers.

1. Intake

Accept pdf, pptx, docx, md, txt, and exported images.

Store one document record with:

  • original filename
  • source type
  • checksum
  • project / module
  • upload time
  • version

2. Page intelligence

Convert each document into document objects, not just one flat text blob.

The object types should be:

  • section
  • page
  • table
  • figure
  • caption
  • symbol-region
  • slide

This matters because Forward-style documents are full of tables, pictures, load-case diagrams, symbols, and mixed text density.

For scanned or drawing-heavy inputs, reuse the existing DraftLint posture:

  • local document preprocessing
  • OCR runtime selection
  • optional layout detection
  • optional symbol detection

The draft note at C:\Users\adeel\OneDrive\100_Knowledge\203_TextCAD\01_Product_Project_Management\00_Project_Management_n_skills\01_tracks\cad-intelligence\handover\260411_draftlint-repo-assessment-and-doc-map.md already points to OCR/runtime-split pieces worth extracting.

3. Knowledge normalization

Normalize extracted content into source-traceable chunks and facts.

For the demo, the minimum useful record is:

  • chunk text
  • source file
  • page or line range
  • tags
  • related figures/tables
  • confidence / extraction mode

Then add lightweight derived structure:

  • topic tags like material-cards, regulation, validation, crash, CFD, meshing
  • glossary terms
  • entity links
  • suggested wiki pages

4. Retrieval and synthesis

Expose the normalized objects through:

  1. technical wiki pages,
  2. source-linked search,
  3. promptable Q&A with citations,
  4. downstream workflow helpers such as report generation or checklist assembly.

What to build now versus later

For Thursday, keep the demo scope tight.

Build now

  • PDF plus Markdown ingestion
  • chunking with page/section traceability
  • source-linked keyword or hybrid search
  • one generated technical wiki page per topic
  • explicit flags for needs OCR, table-like, or layout review

Fake honestly

  • deep table understanding
  • figure understanding
  • symbol semantics
  • PPTX fidelity
  • embeddings quality

If asked, say these are the next layer once the document-object model is in place.

Why this fits Rapid Craft

Rapid Craft already has nearby ingredients:

  1. a collaboration shell with comments, revisions, and shareable context,
  2. CAD and review workflows built around traceable evidence,
  3. prior OCR and document-processing experiments from DraftLint,
  4. planning notes that already mention eventual semantic search.

That means the quickest path is not a separate product. It is a new module that reuses the same ideas:

  • uploads and versions from collaboration,
  • evidence and traceability from CAD intelligence,
  • OCR/layout parts from DraftLint,
  • wiki and search as the first pilot-facing surface.

Forward-specific use cases to mention

The most credible first use cases are:

  1. Technical wiki for battery-enclosure research, standards, and validation methods.
  2. Search across old reports for material card, bottom impact, thermal propagation, sealing, or venting.
  3. Auto-assemble a regulation evidence dossier for GB 38031-2025 or R100.
  4. Capture reusable notes for crash pre-processing, contact-definition strategy, and meshing setup.
  5. Link test data, design assumptions, and CAE reports into one traceable view.

In this folder

The lightweight prototype for this idea lives in:

  • prototype/forward_ingest_demo.py
  • prototype/output/document_manifest.json
  • prototype/output/chunks.json
  • prototype/output/query_results.json
  • Technical ingest demo