Use Cases + Implementation

How do you reduce the cost of OCR workflows with model routing?

OCR workflows should pay models for judgement, not for reading. Run OCR with your own engine, use confidence scores to split pages into easy and hard, route text cleanup and field extraction to cheap tiers with JSON schemas, and escalate only low-confidence pages. Per-page routing and checkpointing keep retries cheap and prevent one bad page from repricing the document.

Key facts

Router modelplugsky-fusion escalates per page request across tiers (live)
PipelineOCR or parse first, then classify, clean and extract per page
ConfidenceUse OCR confidence scores to decide which pages escalate
JSON modeLive for per-page field extraction schemas
PricingFlat monthly self-serve plans with no per-token charges on self-serve
Free tierplugsky-micro and plugsky-lite on the free plan, no card required
Usage controlsScoped keys and usage analytics per OCR pipeline
RoadmapThe images endpoint is coming soon; keep OCR in your pipeline today

TL;DR

  • OCR outside the model; models handle cleanup and extraction.
  • Route per page using OCR confidence, not per document.
  • Cheap tiers clean and extract; low-confidence pages escalate.
  • Checkpoint pages so retries never restart the whole file.
  • Measure cost per processed page and per exception.

How it works, step by step

  1. Run OCR with a deterministic engine and capture per-word or per-page confidence.
  2. Normalise OCR output cheaply: fix common confusions, merge broken lines, strip artifacts.
  3. Classify each page layout with a cheap tier and a small schema.
  4. Extract fields per page with JSON mode, validating against document-type schemas.
  5. Escalate low-confidence pages and failed validations to stronger tiers.
  6. Checkpoint per page and cache parsed text by file hash so reruns only touch changed pages.
  7. Track cost per processed page, exception rate and downstream correction rate.
1Run OCR with adeterministicengine and capture2Normalise OCRoutput cheaply: fixcommon confusions,3Classify each pagelayout with a cheaptier and a small4Extract fields perpage with JSONmode, validating5Escalatelow-confidencepages and failed6Checkpoint per pageand cache parsedtext by file hash

Try it yourself

Open the tokens-to-pages converter →

Cheap extraction, selective reasoning

The reading step belongs to OCR, which is deterministic, fast and cheap at volume. The model's role starts after text exists: cleaning artifacts, classifying layout, extracting fields and resolving exceptions. Teams that send page images to a language model pay premium prices for work an OCR engine does better and cheaper — and the images endpoint is coming soon rather than live today.

Confidence scores are the routing signal. Pages with high OCR confidence are usually safe for cheap-tier cleanup and extraction. Pages with low confidence, unusual layouts or multilingual mixes escalate. That single split typically removes most model spend from a digitisation pipeline.

Page-level pipelines and checkpoints

Documents are uneven, so process them unevenly. A per-page queue lets you route each unit independently and retry failures without reprocessing siblings. Checkpoint state after every page and cache OCR text by file hash so a re-run of a corrected document touches only changed pages.

  • Classify page layout before extraction so each type gets an appropriate schema.
  • Validate extracted fields deterministically: totals, dates, identifiers and totals that must reconcile.
  • Escalate once, then queue for human review — never loop on the same failing configuration.
  • Keep confidence and model tier per page for audit and tuning.

Throughput and cost accounting

At scale, throughput matters as much as unit cost: a pipeline that processes pages faster with cheap tiers shortens backlogs and reduces storage of intermediate files. Measure pages per hour alongside cost per page, and batch or stream work to match your infrastructure.

Report cost per processed page including exceptions and corrections, because correction labour usually dominates when extraction fails. Start on the free plan with plugsky-micro and plugsky-lite for cleanup and extraction, then use the 14-day full-access trial for stronger models on low-confidence pages; the live pricing page lists plans, and batch endpoints for large runs are coming soon.

Honest comparison

OCR workflow choiceRouted OCR pipelineVision model per pageOne cheap model per document
Text readingDeterministic OCR engineModel reads every imageDepends on OCR feed
Easy pagesCheap tier cleanup and extractionExpensive per pageCheap and adequate
Hard pagesEscalated by confidenceUniform costFails silently
RetriesCheckpointed per pageWhole-document restartWhole-document restart
EconomicsCost per processed pageCost per pageHidden correction cost

Frequently asked questions

Should OCR run inside the model?

No. Deterministic OCR engines are cheaper and better at reading, and the images endpoint is coming soon rather than available today. Use models after text extraction for cleaning, classification and field extraction.

How do confidence scores help routing?

They identify pages likely to fail cheap-tier extraction. Escalating low-confidence pages selectively is far cheaper than sending every page to a strong model.

Why process per page?

Documents vary internally, and per-page routing makes retries cheap and isolates failures. One difficult page should not reprice an entire document.

How do I avoid re-paying on reruns?

Checkpoint after each page and cache OCR text by file hash. Corrected or updated documents then only reprocess the pages that changed.

What about multilingual OCR?

Detect language per page, apply the right OCR model, and route extraction per language tier. Mixed documents benefit from per-page rather than per-file decisions.

Are files or images endpoints available?

Not yet — files and images endpoints are coming soon. OCR, parsing and storage stay in your pipeline today.

What should I measure?

Cost per processed page including exceptions, exception rate by document type, throughput in pages per hour, and downstream correction rate.

Can I pilot cheaply?

Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial covers stronger models for low-confidence pages.