Construction PDF Binder Extraction

Four text extraction methods were tested head-to-head on real construction PDFs. Vision wins on every page type for structure and layout. See references/extraction-findings.md for the full comparison data.

Vision is the primary extraction method. It handles structure, spatial understanding, watermark transparency, drawing interpretation, and rasterized content reading. No other method comes close on construction PDFs.

Tesseract supplements vision for numeric accuracy. Testing on dense cover sheets revealed that vision at 1568px resolution can hallucinate specific numeric values — "65.0 sq ft" becomes "856", "475 sq ft" gets missed entirely. Tesseract's character-level OCR reliably captures exact digits even on dense pages. The hybrid approach: run both, give subagents both outputs, and cross-reference numbers. On drawing-heavy pages where Tesseract produces garbage, subagents are instructed to ignore it.

Do NOT use pdftotext or pdfplumber for construction PDFs. pdftotext produces garbage on drawings and empty output on rasterized pages. pdfplumber produces reversed text (*TEEHS REVOC*) — completely unusable.

At ~1,500 tokens per page PNG, a full 30-page binder costs ~45K tokens for complete vision extraction — trivial at production pricing.

Extraction Process

Step 1: Prepare Output Directory

Create the output directory structure:

{output-dir}/
├── pages-png/           # One PNG per PDF page (resized to 1568px max)
├── pages-text/          # Tesseract OCR text per page (numeric cross-ref)
├── pages-vision/        # Per-page outputs from vision subagents:
│   ├── page-NN.md       #   Structured markdown (detailed content)
│   └── page-NN.json     #   Manifest fragment (routing entry)
└── binder-manifest.json # Assembled manifest (routing artifact)

Step 2: Extract Page PNGs + Tesseract Text

Run scripts/extract-pages.sh to split the PDF into page PNGs, resize them for API consumption, and run Tesseract OCR:

scripts/extract-pages.sh INPUT.pdf OUTPUT_DIR

The script does three things:

1. Split PDF — Uses pdftoppm at 200 DPI. Each page becomes pages-png/page-01.png, page-02.png, etc.
2. Resize PNGs to 1568px max — Claude's API internally resizes images to 1568px on the longest side. Construction PDFs at 200 DPI on D-size sheets produce 7200x4800 PNGs — resizing before upload saves bandwidth and avoids the 32MB API payload limit without losing any information the model would actually see. Uses ImageMagick (Linux) or sips (macOS).
3. Tesseract OCR — Runs tesseract on each resized PNG to produce pages-text/page-01.txt, etc. These raw text dumps supplement vision extraction by providing reliable numeric values for cross-reference. On drawing-heavy pages, Tesseract output will be garbage — subagents are instructed to recognize and ignore it.

Note: CAD-generated construction PDFs commonly produce Poppler warnings like "Syntax Error: insufficient arguments for Marked Content". These are harmless — the PNGs render correctly. The script suppresses these via 2>/dev/null.

If pdftoppm is not available, fall back to ImageMagick:

magick -density 200 input.pdf -quality 90 output-dir/pages-png/page-%02d.png

Then manually resize and run Tesseract on the resulting PNGs.

Step 3: Vision Extract Every Page (Rolling Window)

Vision extraction is the most time-intensive step. Use a rolling window of parallel subagents — one page per subagent, max 3 in flight at any time. The full prompt template is in prompts/vision-extract-page.md — read it and use it as the prompt for each subagent.

Why One Page Per Subagent

Each subagent conversation accumulates every image it reads into the message history. With multi-page batches, by the time the subagent processes page 3, all 3 PNGs are in context for every API call. This causes:

• API image limits: Claude's API enforces a 2000px-per-image cap when

  20 images are in the conversation. Construction PNGs at 200 DPI are typically 7200x4800 — well over this limit.

• Token waste: Each additional image in context costs ~1,500 tokens per API round-trip, even when only analyzing the current page.
• Quality degradation: More images in context = more noise for the model.

One page per subagent means exactly one image in context. No multi-image limits, cleaner extraction, and the Tesseract text file provides numeric cross-reference without adding image tokens.

Resource Constraints

Maximum 3 concurrent subagents. One page per subagent.

This is a hard constraint for deployment to Vercel sandboxes (4 GB RAM total). The orchestrator + 3 subagents = 4 processes, each getting ~1 GB RAM. Do not exceed 3 concurrent subagents under any circumstances.

Rolling Window Orchestration

Instead of fixed rounds (launch 3, wait for all 3, launch next 3), use a rolling window: launch 3 subagents, and as each one completes, immediately launch the next. This keeps 3 subagents in flight at all times until all pages are processed.

Task tool parameters (per subagent):
  name:            "vision-page-NN"
  subagent_type:   "general-purpose"
  mode:            "bypassPermissions"
  run_in_background: true
  prompt:          (read from prompts/vision-extract-page.md,
                    replace {{PAGE_PNG}}, {{TEXT_FILE}}, {{OUTPUT_MD}},
                    {{OUTPUT_JSON}}, and {{SKILL_DIR}} with actual paths)

1. Count the total page PNGs in pages-png/
2. Launch subagents for pages 1, 2, and 3 (3 in parallel)
3. As each subagent completes, immediately launch the next page
4. Continue until all pages are queued
5. Wait for the final subagents to complete
6. Verify all pages-vision/page-NN.md AND page-NN.json files exist

Throughput

Each subagent takes ~3-4 minutes (read references, read PNG, write .md, write .json). With 3 concurrent, a 26-page binder completes in ~30 minutes.

Output Format

Each subagent writes two files per page to pages-vision/:

1. page-NN.md — Structured markdown with full extracted content:

   • Title block identification (sheet number, title, firm)
   • All text content (tables, notes, schedules, specifications)
   • Spatial zone mapping for every content element
   • Drawing descriptions for non-text content
   • Confidence annotations for watermark-obscured or low-resolution content

2. page-NN.json — Manifest fragment for routing:

   • Page metadata (sheet_id, category, subcategory)
   • key_content array with specific values (guided by extraction priorities)
   • topics keyword tags for corrections letter matching
   • drawing_zones spatial map
   • "NOT SHOWN: [item]" entries for expected-but-absent content
   • Cover sheet fragment includes _project metadata

The extraction priorities reference (references/adu-extraction-priorities.md) guides subagents on what to capture with specificity and what to flag as absent for each content type. This produces manifest entries targeted for corrections letter routing without any decision-making about compliance.

See prompts/vision-extract-page.md for the full prompt template including both output formats.

Step 4: Assemble the Manifest

The manifest is what makes everything else useful. It enables an agent to route to the correct page(s) without loading all pages into context.

Since each vision subagent already wrote a JSON manifest fragment per page (in Step 3), assembly is deterministic — no LLM needed.

Run the assembly script:

python3 scripts/assemble-manifest.py {output}/pages-vision {output}/binder-manifest.json

The script:

1. Reads all page-NN.json fragments from pages-vision/
2. Extracts _project metadata from the cover sheet fragment
3. Combines into { "project": {...}, "pages": [...] }
4. Validates required fields and page numbering
5. Writes binder-manifest.json

Exit codes: 0 = clean, 1 = assembled with issues, 2 = fatal error.

The orchestrator MUST always review the assembled manifest (see Step 4a). The assembly script is deterministic but not smart — it can concatenate JSON but it cannot catch semantic issues like a wrong category, a vague key_content entry, or a missing _project field that should have been extracted. A quick orchestrator read-through catches things the script never could.

The assembled manifest follows the schema in references/manifest-schema.md. Each page entry captures:

1. Sheet ID and title — from the title block
2. Category — general, architectural, structural, energy, code_compliance, mechanical, plumbing, electrical
3. What's on the page — key content items with exact values, specific enough to match correction letter items
4. What's NOT on the page — "NOT SHOWN: [item]" entries for expected- but-absent content (guided by extraction priorities)
5. Topics — keyword tags for routing
6. Drawing zones — spatial map of where things are on the page

Step 4a: Orchestrator Review (ALWAYS — Not Optional)

After the assembly script runs, the orchestrator must read binder-manifest.json and review it. This takes seconds and catches things the script cannot.

Standard Review Checklist

1. Read the full binder-manifest.json
2. Verify project metadata is populated (address, type, owner, sqft)
3. Verify page count matches PNG count
4. Spot-check sheet_id values look reasonable
5. Check that key_content arrays have specific values, not vague entries
6. If the script reported issues (exit code 1), fix them
7. Fix any JSON errors, missing fields, or wrong categories

Cross-Page Consistency Check (Critical)

Vision models can hallucinate individual digits — a "3" read as "2", a "5" as "6". When this happens on the cover sheet, the wrong value cascades into project metadata and poisons everything downstream.

Every page's JSON fragment includes a title_block_address field — the address as read independently from that page's title block. The orchestrator must use these to verify project-level values:

1. Collect all title_block_address values from every page entry
2. Majority vote on the address: the value that appears on the most pages is the correct address. If the project.address differs from the majority, fix it.
3. Apply the same logic to other repeated values: project type, designer firm, and structural engineer firm appear on multiple title blocks. When there's a conflict, the majority wins.
4. Log any corrections: when the orchestrator overrides a value, note what was changed and why (e.g., "Fixed address from 1222 to 1232 — cover sheet hallucination, 14/15 pages read 1232").

This check exists because in testing, the vision model misread "1232" as "1222" on one page, and that single error propagated through the entire manifest. With 15 pages each independently reading the title block, a single-page hallucination is trivially detectable.

This review is cheap (one file read + a few string comparisons) and prevents the scenario where subagents did great work but a single hallucination or script assembly glitch ruins the output.

Reading Title Blocks

Construction plan title blocks follow consistent conventions:

• Location: Bottom-right corner or right edge of each sheet
• Contains: Sheet number (e.g., "A2", "S1"), sheet title, designer/engineer name, project info, revision dates
• Sheet numbering convention:
  • CS = Cover Sheet
  • A prefix = Architectural (site plans, floor plans, elevations)
  • S prefix = Structural (foundation, framing, details)
  • SN prefix = Structural Notes
  • T prefix = Title 24 / Energy
  • AIA prefix = CalGreen/code checklists
  • M prefix = Mechanical
  • P prefix = Plumbing
  • E prefix = Electrical

Drawing Zone Mapping

To enable precise references like "Sheet S2, detail 8, mid-left quadrant":

• Divide each page into a grid (top/middle/bottom x left/center/right)
• For detail sheets with numbered detail bubbles, map bubble numbers to zones
• For plans, note which drawing is in which half (e.g., "left-half: foundation plan, right-half: framing plan")

Step 5: Validate Outputs

After extraction, verify:

• PNG count matches PDF page count
• Vision markdown files exist for every page
• Manifest JSON is valid and has entries for every page
• Every sheet_id in the manifest matches what's visible in the PNG title block

Using Extraction Results

For Corrections Response (Flow 2)

When interpreting a corrections letter against extracted plans:

1. Parse each correction item for keywords
2. Match keywords against manifest topics and key_content arrays
3. Load only the matched page PNGs into context (vision) for verification
4. Use the pages-vision/ markdown for quick text searches
5. Reference corrections by sheet ID and drawing zone: "See Sheet S2 (page 11), Shearwall Schedule in the mid-left quadrant"

For Permit Checklist (Flow 1)

When generating a permit checklist from extracted plans:

1. Load the cover sheet manifest entry for project overview
2. Walk each category (architectural, structural, energy) loading relevant pages
3. Use vision markdown files for data extraction, PNGs for verification
4. Cross-reference against ADU regulatory skill requirements

Typical Sheet Types in ADU Binders

For reference, a typical California ADU plan binder contains:

Orchestration Summary

The full extraction workflow. Hard limit: max 3 concurrent subagents, 1 page per subagent (4 GB RAM deployment environment).

Example for a 15-page binder:

Step 1: mkdir -p {output}/pages-png {output}/pages-text {output}/pages-vision

Step 2: bash scripts/extract-pages.sh INPUT.pdf {output}
        → Split PDF into PNGs (200 DPI)
        → Resize PNGs to 1568px max (API internal limit)
        → Run Tesseract OCR → pages-text/page-01.txt through page-15.txt
        → produces pages-png/page-01.png through page-15.png

Step 3: Rolling window of vision subagents (prompts/vision-extract-page.md)
        → Launch page-01, page-02, page-03 in parallel (3 in flight)
        → page-01 completes → launch page-04 (still 3 in flight)
        → page-03 completes → launch page-05
        → ... continue until all 15 pages queued ...
        → Wait for final subagents to complete
        → Verify: pages-vision/page-NN.md AND page-NN.json exist for all 15

Step 4: python3 scripts/assemble-manifest.py {output}/pages-vision {output}/binder-manifest.json
        → Reads all page-NN.json fragments
        → Assembles binder-manifest.json (deterministic, no LLM)

Step 4a: Orchestrator reads binder-manifest.json (ALWAYS, not optional)
        → Cross-page consistency check: majority-vote address + repeated values
        → Verifies project metadata, page count, key_content quality
        → Fixes any assembly issues, hallucinations, or missing fields

Step 5: Validate all outputs

Steps 1-2 are sequential (bash). Step 3 uses a rolling window — as each subagent finishes, the next page launches immediately. Each subagent reads one PNG, writes one .md and one .json. Step 4 is a fast Python script (no LLM call). Step 5 is orchestrator validation.

Why one page per subagent? Each subagent's conversation accumulates every image it reads. With 3 pages per subagent, the 3rd page's API calls include all 3 PNGs in context — wasting tokens, risking API image limits (2000px cap for >20 images in conversation), and degrading quality. One-per-subagent keeps exactly one image in context at all times.

Why inline fragments? Each vision subagent already has the page image in context and has done the deep analysis. Writing a manifest entry at that point is nearly free — just reformatting what it already knows into JSON. This is faster and more accurate than a separate manifest subagent re-reading all the markdown files.

Resources

scripts/

• extract-pages.sh — Split PDF into per-page PNGs (200 DPI), resize to 1568px, run Tesseract OCR for hybrid text cross-reference
• assemble-manifest.py — Assemble page JSON fragments into binder-manifest.json

prompts/

• vision-extract-page.md — Subagent prompt template for single-page vision extraction (produces both markdown and JSON manifest fragment for one page)
• vision-extract-batch.md — Legacy batch prompt (retained for reference; the single-page approach in vision-extract-page.md supersedes this)
• build-manifest.md — Legacy manifest subagent prompt (retained for reference; the inline fragment approach supersedes this)

references/

• manifest-schema.md — JSON schema and field descriptions for binder-manifest.json
• adu-extraction-priorities.md — Domain-aware extraction guide: what to capture, what to flag as absent, and corrections letter terminology by content type
• extraction-findings.md — Lessons learned from testing on real construction PDFs