Appendix B — Quick-Reference Cheatsheet

The book in one page. Print this, tape it to a wall, refer back when the marketing language starts to drift.

B.1 The five attributes

A system is agentic OCR if and only if it is:

  1. Goal-driven — operates against a structured target (schema, extraction goal), not against the document.
  2. Multi-pass — iterates with at least one self-review pass, not single-shot.
  3. Self-correcting — has working detection, repair, and escalation subsystems (all three).
  4. Tool-using — routinely invokes specialized tools (layout detector, table extractor, schema validator, KB lookup), not just function-calling-capable.
  5. Grounded — every output field traces to a source region (page, bounding box).

Three of five is “VLM with retries.” Four of five is borderline. Five of five is the bar.

See Chapter 3 for the full definition; Chapter 13 for the anti-patterns that come from missing each attribute.

B.2 The three architecture patterns

Pattern Description Best for Failure mode
A — Single-pass VLM One model call; structured output Low-stakes, low-variability workloads; commodity OCR Tables, long documents, ungrounded output
B — Agentic loop with reflection Plan → extract → critique → re-prompt → validate The 2026 default; messy enterprise docs at moderate stakes Silent hallucinations; retry thrash
C — Hybrid CV+VLM Layout-first; route regions to specialized extractors High-stakes, adversarial tables, regulated workloads Engineering complexity; operational overhead

See Chapter 4 for the full treatment.

B.3 The decision tree in one diagram

START
  │
  ├─ Volume < 10k pages/mo? ───→ Hyperscaler API (Pattern A)
  │
  ├─ Volume > 1M pages/mo + engineering capacity? ───→ Self-hosted open-source
  │
  └─ Otherwise:
       ├─ High stakes (compliance, healthcare, finance) ───→ Pattern C + HITL
       ├─ Wide variability, moderate stakes ───→ Pattern B (AI-native specialist)
       └─ Narrow + stable workload ───→ Classic IDP or template-based

See Chapter 9 for the full framework, including the “should I build this at all?” check in Section 9.5.

B.4 The cost rule of thumb

All-in cost-per-correct-extraction for true agentic OCR in 2026: $0.05 – $0.30 per page.

This includes parsing API spend, HITL amortized over typical 5% escalation, and eval/observability/vendor management.

Defensibility thresholds: - Downstream per-page value below 1¢: agentic OCR has already eaten your budget. Don’t build. - Downstream per-page value 1–5¢: marginal at best; wait for prices to drop. - Downstream per-page value 5–25¢: economically defensible if the workload justifies the architecture. - Downstream per-page value above 25¢: comfortably above the threshold (healthcare prior-auth, KYC, contract review, clinical-note coding).

The objective function to minimize: cost-per-correct-extraction-with-provenance, not cost-per-page.

See Chapter 9 and Chapter 11.

B.5 Benchmarks at a glance

Benchmark Year Best use Caveat
ParseBench 2026 Commercial vendor comparison (LlamaParse Agentic 84.9%) Vendor-published
olmOCR-Bench 2025 Open-source model comparison (Chandra 83.1) English-only
OmniDocBench v1.7 2026 Cross-validation (use hard subset) Saturated at the top
RD-TableBench 2025 Table extraction (Reducto ~0.90) Vendor-published
DABstep 2025 Reality check (~14.55% ceiling) Small N
SCORE-Bench 2026 Generative-parser eval New methodology
OfficeQA 2026 Shift-left thesis evidence (16% gain) Databricks-maintained

The most important rule: benchmark rank ≠ production rank. Your golden set wins.

See Chapter 10.

B.6 Anti-pattern smell tests

A one-line check for each of the five dominant anti-patterns from Chapter 13:

  1. Wrapper test: Does the system satisfy all five attributes? (If not, it’s wrapper-washing.)
  2. Schema-flexibility test: What does the system do with an unexpected field? (If it drops it silently, schema rigidity.)
  3. HITL test: What happens to a confident-but-wrong output? (If it ships, ignoring HITL.)
  4. Calibration test: Plot confidence vs. accuracy on your golden set — is the curve smooth? (If not, you’re over-relying on confidence.)
  5. Decomposition test: When a downstream answer is wrong, can you tell whether parsing or reasoning was at fault? (If not, you’re conflating them.)

Print these. Use them before signing.

B.7 Six eval properties of a healthy 2026 deployment

From Chapter 11:

  • Golden set: 200–500 docs from production, refreshed quarterly.
  • Metrics: field-level F1, schema-validity, grounding accuracy (per-field, not aggregate).
  • Regression cadence: weekly subset, monthly full, quarterly refresh.
  • HITL closed-loop: corrections flow into golden set, prompt tuning, vendor selection.
  • Cost-per-correct-extraction: tracked as the primary economic metric.
  • OOD slice: held out and measured separately.

B.8 RAG + agentic OCR

The three concrete ways agentic OCR helps RAG (Chapter 12):

  • Structural metadata → semantic chunking (chunks = sections, not fixed windows).
  • Table awareness → tables as atomic chunks with cell-relationship metadata preserved.
  • Grounding → real citations, audit infrastructure.

The diagnostic before upgrading the parser: hand-trace 50 wrong-answer queries. If parsing failures are under 30% of total, parsing is not your bottleneck. Don’t spend $500k upgrading the wrong layer.

B.9 The three named takes worth memorizing

  1. “If your agent can read perfectly but reason imperfectly, you have a model problem. If your agent reasons brilliantly on the wrong inputs, you have a document problem — and 2026 says the document problem is bigger.”

  2. “‘Agentic’ is the word that lets people charge 10× for a retry loop. The five-attribute test exists so you don’t get fooled.”

  3. “Most teams pick a vendor before they pick a metric. This is backwards. Choose the metric that determines your downstream cost, then choose the vendor that minimizes that metric.”

B.10 What 2027 commoditizes (and doesn’t)

Will commoditize: parsing quality, single-pass VLM extraction, basic table extraction, multilingual coverage.

Will not commoditize (yet): agentic orchestration platforms, compliance and grounding infrastructure, observability for document agents, vertical specialization.

Build your vendor relationships and your engineering investment around the categories that will not commoditize.

B.11 The thesis, restated

Reading is the ceiling on agent quality, and reading is mostly solved. The next ceiling is trust.

That is the entire book.