Batch Book Translation Workflow

Process books through the complete pipeline: Crop → OCR → Translate

Roadmap Reference

See .claude/ROADMAP.md for the translation priority list.

Priority 1 = UNTRANSLATED - These are highest priority for processing:

• Kircher encyclopedias (Oedipus, Musurgia, Ars Magna Lucis)
• Fludd: Utriusque Cosmi Historia
• Theatrum Chemicum, Musaeum Hermeticum
• Cardano: De Subtilitate
• Della Porta: Magia Naturalis
• Lomazzo, Poliziano, Landino

bash
1# Get roadmap with priorities
2curl -s "https://sourcelibrary.org/api/books/roadmap" | jq '.books[] | select(.priority == 1) | {title, notes}'

Roadmap source: src/app/api/books/roadmap/route.ts

Overview

This workflow handles the full processing pipeline for historical book scans:

1. Generate Cropped Images - For split two-page spreads, extract individual pages
2. OCR - Extract text from page images using Gemini vision
3. Translate - Translate OCR'd text with prior page context for continuity

API Endpoints

Processing Options

Option 1: Lambda Workers via Job System (Primary Path)

The primary processing path uses AWS Lambda workers via SQS queues. Each page is processed independently with automatic job tracking.

bash
1# Queue OCR for a book's pages
2curl -s -X POST "https://sourcelibrary.org/api/jobs/queue-books" \
3  -H "Content-Type: application/json" \
4  -d '{"bookIds": ["BOOK_ID"], "action": "ocr"}'
5
6# Queue translation
7curl -s -X POST "https://sourcelibrary.org/api/jobs/queue-books" \
8  -H "Content-Type: application/json" \
9  -d '{"bookIds": ["BOOK_ID"], "action": "translation"}'
10
11# Queue image extraction
12curl -s -X POST "https://sourcelibrary.org/api/jobs/queue-books" \
13  -H "Content-Type: application/json" \
14  -d '{"bookIds": ["BOOK_ID"], "action": "image_extraction"}'

IMPORTANT: Always use gemini-3-flash-preview for all OCR and translation tasks. Do NOT use gemini-2.5-flash.

Option 2: Gemini Batch API (50% Cheaper, Automated Pipeline)

The post-import-pipeline cron uses Gemini Batch API for automated processing of newly imported books. Results arrive in ~24 hours at 50% cost.

OCR Output Format

OCR uses Markdown output with semantic tags:

Markdown Formatting

• # ## ### for headings (bigger text = bigger heading)
• **bold**, *italic* for emphasis
• ->centered text<- for centered lines (NOT for headings)
• > blockquotes for quotes/prayers
• --- for dividers
• Tables only for actual tabular data

Metadata Tags (hidden from readers)

Inline Annotations (visible to readers)

Critical OCR Rules

1. Preserve original spelling, capitalization, punctuation
2. Page numbers/headers/signatures go in metadata tags only
3. IGNORE partial text at edges (from facing page in spread)
4. Describe images/diagrams with <image-desc>, never tables
5. End with <vocab>key terms, names, concepts</vocab>

Step 1: Analyze Book Status

First, check what work is needed for a book:

bash
1# Get book and analyze page status
2curl -s "https://sourcelibrary.org/api/books/BOOK_ID" > /tmp/book.json
3
4# Count pages by status (IMPORTANT: check length > 0, not just existence - empty strings are truthy!)
5jq '{
6  title: .title,
7  total_pages: (.pages | length),
8  split_pages: [.pages[] | select(.crop)] | length,
9  needs_crop: [.pages[] | select(.crop) | select(.cropped_photo | not)] | length,
10  has_ocr: [.pages[] | select((.ocr.data // "") | length > 0)] | length,
11  needs_ocr: [.pages[] | select((.ocr.data // "") | length == 0)] | length,
12  has_translation: [.pages[] | select((.translation.data // "") | length > 0)] | length,
13  needs_translation: [.pages[] | select((.ocr.data // "") | length > 0) | select((.translation.data // "") | length == 0)] | length
14}' /tmp/book.json

Detecting Bad OCR

Pages that were OCR'd before cropped images were generated have incorrect OCR (contains both pages of the spread). Detect these:

bash
1# Find pages with crop data + OCR but missing cropped_photo at OCR time
2# These often contain "two-page" or "spread" in the OCR text
3jq '[.pages[] | select(.crop) | select(.ocr.data) |
4  select(.ocr.data | test("two-page|spread"; "i"))] | length' /tmp/book.json

Step 2: Generate Cropped Images

For books with split two-page spreads, generate individual page images:

bash
1# Get page IDs needing crops
2CROP_IDS=$(jq '[.pages[] | select(.crop) | select(.cropped_photo | not) | .id]' /tmp/book.json)
3
4# Create crop job
5curl -s -X POST "https://sourcelibrary.org/api/jobs" \
6  -H "Content-Type: application/json" \
7  -d "{
8    \"type\": \"generate_cropped_images\",
9    \"book_id\": \"BOOK_ID\",
10    \"book_title\": \"BOOK_TITLE\",
11    \"page_ids\": $CROP_IDS
12  }"

Process the job:

bash
1# Trigger processing (40 pages per request, auto-continues)
2curl -s -X POST "https://sourcelibrary.org/api/jobs/JOB_ID/process"

Step 3: OCR Pages

Option A: Using Job System (for large batches)

bash
1# Get page IDs needing OCR (check for empty strings, not just null)
2OCR_IDS=$(jq '[.pages[] | select((.ocr.data // "") | length == 0) | .id]' /tmp/book.json)
3
4# Create OCR job
5curl -s -X POST "https://sourcelibrary.org/api/jobs" \
6  -H "Content-Type: application/json" \
7  -d "{
8    \"type\": \"batch_ocr\",
9    \"book_id\": \"BOOK_ID\",
10    \"book_title\": \"BOOK_TITLE\",
11    \"model\": \"gemini-3-flash-preview\",
12    \"language\": \"Latin\",
13    \"page_ids\": $OCR_IDS
14  }"

Option B: Using Lambda Workers with Page IDs

bash
1# OCR specific pages (including overwrite)
2curl -s -X POST "https://sourcelibrary.org/api/jobs/queue-books" \
3  -H "Content-Type: application/json" \
4  -d '{
5    "bookIds": ["BOOK_ID"],
6    "action": "ocr",
7    "pageIds": ["PAGE_ID_1", "PAGE_ID_2"],
8    "overwrite": true
9  }'

Lambda workers automatically use cropped_photo when available.

Step 4: Translate Pages

Option A: Using Job System

bash
1# Get page IDs needing translation (must have OCR content, check for empty strings)
2TRANS_IDS=$(jq '[.pages[] | select((.ocr.data // "") | length > 0) | select((.translation.data // "") | length == 0) | .id]' /tmp/book.json)
3
4# Create translation job
5curl -s -X POST "https://sourcelibrary.org/api/jobs" \
6  -H "Content-Type: application/json" \
7  -d "{
8    \"type\": \"batch_translate\",
9    \"book_id\": \"BOOK_ID\",
10    \"book_title\": \"BOOK_TITLE\",
11    \"model\": \"gemini-3-flash-preview\",
12    \"language\": \"Latin\",
13    \"page_ids\": $TRANS_IDS
14  }"

Option B: Using Lambda Workers (Recommended)

Lambda FIFO queue automatically provides previous page context for translation continuity:

bash
1# Queue translation for pages that have OCR but no translation
2curl -s -X POST "https://sourcelibrary.org/api/jobs/queue-books" \
3  -H "Content-Type: application/json" \
4  -d '{"bookIds": ["BOOK_ID"], "action": "translation"}'

The translation Lambda worker processes pages sequentially via FIFO queue and fetches the previous page's translation for context.

Complete Book Processing Script

Process a single book through the full pipeline using Lambda workers:

bash
1#!/bin/bash
2BOOK_ID="YOUR_BOOK_ID"
3BASE_URL="https://sourcelibrary.org"
4
5# 1. Fetch book data
6echo "Fetching book..."
7BOOK=$(curl -s "$BASE_URL/api/books/$BOOK_ID")
8TITLE=$(echo "$BOOK" | jq -r '.title[0:40]')
9echo "Processing: $TITLE"
10
11# 2. Queue OCR (Lambda workers handle all pages automatically)
12NEEDS_OCR=$(echo "$BOOK" | jq '[.pages[] | select((.ocr.data // "") | length == 0)] | length')
13if [ "$NEEDS_OCR" != "0" ]; then
14  echo "Queueing OCR for $NEEDS_OCR pages..."
15  curl -s -X POST "$BASE_URL/api/jobs/queue-books" \
16    -H "Content-Type: application/json" \
17    -d "{\"bookIds\": [\"$BOOK_ID\"], \"action\": \"ocr\"}"
18  echo "OCR job queued!"
19fi
20
21# 3. Queue translation (after OCR completes — check /jobs page)
22NEEDS_TRANS=$(echo "$BOOK" | jq '[.pages[] | select((.ocr.data // "") | length > 0) | select((.translation.data // "") | length == 0)] | length')
23if [ "$NEEDS_TRANS" != "0" ]; then
24  echo "Queueing translation for $NEEDS_TRANS pages..."
25  curl -s -X POST "$BASE_URL/api/jobs/queue-books" \
26    -H "Content-Type: application/json" \
27    -d "{\"bookIds\": [\"$BOOK_ID\"], \"action\": \"translation\"}"
28  echo "Translation job queued!"
29fi
30
31echo "Jobs queued! Monitor progress at $BASE_URL/jobs"

Fixing Bad OCR

When pages were OCR'd before cropped images existed, they contain text from both pages. Fix with:

bash
1# 1. Generate cropped images first (Step 2 above)
2
3# 2. Find pages with bad OCR
4BAD_OCR_IDS=$(jq '[.pages[] | select(.crop) | select(.ocr.data) |
5  select(.ocr.data | test("two-page|spread"; "i")) | .id]' /tmp/book.json)
6
7# 3. Re-OCR with overwrite via Lambda workers
8curl -s -X POST "https://sourcelibrary.org/api/jobs/queue-books" \
9  -H "Content-Type: application/json" \
10  -d "{\"bookIds\": [\"BOOK_ID\"], \"action\": \"ocr\", \"pageIds\": $BAD_OCR_IDS, \"overwrite\": true}"

Processing All Books

Use the Lambda worker job system for bulk processing:

bash
1#!/bin/bash
2BASE_URL="https://sourcelibrary.org"
3
4# Get all book IDs
5BOOK_IDS=$(curl -s "$BASE_URL/api/books" | jq -r '[.[].id]')
6
7# Queue OCR for all books (Lambda workers handle parallelism and rate limiting)
8curl -s -X POST "$BASE_URL/api/jobs/queue-books" \
9  -H "Content-Type: application/json" \
10  -d "{\"bookIds\": $BOOK_IDS, \"action\": \"ocr\"}"
11
12# After OCR completes, queue translation
13curl -s -X POST "$BASE_URL/api/jobs/queue-books" \
14  -H "Content-Type: application/json" \
15  -d "{\"bookIds\": $BOOK_IDS, \"action\": \"translation\"}"

Monitor progress at https://sourcelibrary.org/jobs

Monitoring Progress

Check overall library status:

bash
1curl -s "https://sourcelibrary.org/api/books" | jq '[.[] | {
2  title: .title[0:30],
3  pages: .pages_count,
4  ocr: .ocr_count,
5  translated: .translation_count
6}] | sort_by(-.pages)'

Troubleshooting

Empty Strings vs Null (CRITICAL)

In jq, empty strings "" are truthy! This means:

• select(.ocr.data) matches pages with "" (WRONG)
• select(.ocr.data | not) does NOT match pages with "" (WRONG)
• Use select((.ocr.data // "") | length == 0) to find missing/empty OCR
• Use select((.ocr.data // "") | length > 0) to find pages WITH OCR content

Rate Limits (429 errors)

Gemini API Tiers

Optimal Sleep Times by Tier

Note: Use ~50% of max rate to leave headroom for bursts.

API Key Rotation

The system supports multiple API keys for higher throughput:

• Set GEMINI_API_KEY (primary)
• Set GEMINI_API_KEY_2, GEMINI_API_KEY_3, ... up to GEMINI_API_KEY_10
• Keys rotate automatically with 60s cooldown after rate limit

With N keys at Tier 1, you get N × 300 RPM = N × 150 safe req/min

Function Timeouts

• Jobs have maxDuration=300s for Vercel Pro
• If hitting timeouts, reduce CROP_CHUNK_SIZE in job processing

Missing Cropped Photos

• Check if crop job completed successfully
• Verify page has crop data with xStart and xEnd
• Re-run crop generation for specific pages

Bad OCR Detection

Look for these patterns in OCR text indicating wrong image was used:

• "two-page spread"
• "left page" / "right page" descriptions
• Duplicate text blocks
• References to facing pages