tune-multi-modal-docproc — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited tune-multi-modal-docproc (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
| Page Content | Use OCR | Use Vision | Use Both | Cost Ratio |
|---|---|---|---|---|
| Pure text | ✅ | ❌ | ❌ | 1x (cheapest) |
| Simple table | ✅ | ❌ | ❌ | 1x |
| Complex table | ✅ | ✅ (verify) | ✅ | 5x |
| Bar/line chart | ❌ | ✅ | ❌ | 10x |
| Photo/image | ❌ | ✅ | ❌ | 10x |
| Stamp/seal | ❌ | ✅ | ❌ | 10x |
| Handwriting | ✅ (try) | ✅ (fallback) | ✅ | 5x |
| Mixed (text+visual) | ✅ (text) | ✅ (visual) | ✅ | 5x |
Cost impact: Routing all pages through GPT-4o vision costs ~10x more than OCR-only. Intelligent routing saves 60-80% on text-heavy documents.
| Resolution | Quality | Speed | VRAM | Best For |
|---|---|---|---|---|
| 150 DPI | Low | Fast | Small | Quick preview / classification only |
| 300 DPI | Good | Medium | Medium | Standard OCR + vision |
| 600 DPI | High | Slow | Large | Fine print, stamps, signatures |
| 1024px max | Standard | Fast | Medium | GPT-4o vision (API limit) |
| 2048px max | High | Medium | Large | GPT-4o vision (max quality) |
Rule: OCR at 300 DPI, vision at 1024-2048px. Never send full-resolution scans to vision (wastes tokens and time).
| Visual Element | Prompt Strategy | Output Format |
|---|---|---|
| Bar chart | "Extract data points as {label: value}" | JSON array |
| Line chart | "Extract x,y coordinates for each series" | JSON array of arrays |
| Pie chart | "Extract segment labels and percentages" | JSON object |
| Stamp/seal | "Read text, identify type, note position" | JSON object |
| Signature | "Describe position, note if signed" | JSON object |
| Logo | "Identify company, describe visual" | JSON object |
| Photo | "Describe content, extract visible text" | JSON object |
Prompt optimization tips:
| Strategy | Pages/min | Concurrency | Best For |
|---|---|---|---|
| Sequential | 4-6 | 1 | Testing, small batches |
| Parallel OCR | 20-30 | 5 | Text-heavy documents |
| Parallel vision | 8-12 | 3 | Visual-heavy documents |
| Mixed parallel | 15-20 | OCR:5, Vision:3 | Production mixed documents |
| Queue-based | 50+ | Auto-scale | High-volume enterprise |
Bottleneck: GPT-4o vision API has lower throughput than OCR. Parallelize OCR aggressively, rate-limit vision calls.
Cost breakdown (10-page mixed document):
| Component | Pages | Cost |
|---|---|---|
| Document Intelligence (OCR) | 7 text pages | $0.07 |
| GPT-4o vision | 3 visual pages | $0.15 |
| Storage | Upload + cache | $0.001 |
| Cosmos DB | Result storage | $0.001 |
| Total | 10 pages | $0.22 |
Without intelligent routing (all pages through vision): $0.50 — 56% savings from routing.
Optimization levers:
After tuning, compare:
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.