EvidencePipelineKit

Apple-platform document intelligence that treats OCR and RAG as evidence pipelines, not as permission to turn uncertain text into confident facts.

The package demonstrates a confidentiality-safe subset of patterns used in production-style document workflows:

• PDFKit embedded-text fast path before OCR.
• Apple Vision accurate OCR fallback with configurable English/Russian recognition.
• Spatial line reconstruction from normalized Vision bounding boxes.
• Hard page, file-size and raster limits.
• Confidence and provenance retained as first-class data.
• Deterministic retrieval for downstream RAG context.
• Synthesis blocked when eligible sources contradict each other.
• Explicit reviewRequired outcome when evidence is missing or too weak.

It contains no customer documents, private endpoints, employer code, credentials, trained models or proprietary extraction rules.

Build and test

xcrun swift-format lint --recursive --configuration .swift-format Sources Tests Package.swift
xcrun swift test --parallel
xcrun swift build -c release
xcrun swift run evidence-pipeline-demo

Requirements: Xcode 16.4+ or a compatible Swift 6 toolchain on macOS 14+. The library targets iOS 17+ and macOS 14+.

Architecture

local PDF
  -> file/page safety gates
  -> PDFKit embedded-text fast path
     OR PDF rasterization -> Apple Vision OCR
  -> normalized lines + confidence + source provenance
  -> deterministic fact extraction owned by the host app
  -> GroundedContextBuilder
       -> confidence gate
       -> contradiction gate
       -> deterministic query/fact ranking
       -> cited context OR reviewRequired OR blockedByContradiction
  -> optional downstream LLM under host-app policy

The package intentionally stops before an LLM call. A model may summarize the supplied context; it must not invent absent evidence or resolve contradictory official facts by itself.

Core types

AppleDocumentReader

Reads local PDFs with a fast, accurate text-layer path and an Apple Vision fallback. It preserves the chosen path per page and reports mean OCR confidence rather than flattening every result into a plain string.

OCROrdering

Groups Vision observations into rows using normalized geometry, then orders row fragments left to right. The function is deterministic and unit-tested without requiring camera or document fixtures.

GroundedContextBuilder

Builds a small, cited context for downstream RAG. Low-confidence facts are rejected, conflicting eligible values block synthesis, and deterministic tie-breaking keeps tests and audit logs stable.

let result = GroundedContextBuilder(minimumConfidence: 0.8).build(
    query: "What is the permitted use?",
    facts: extractedFacts
)

switch result.verdict {
case .grounded:
    sendToModel(result.context, citations: result.citations)
case .reviewRequired, .blockedByContradiction:
    routeToHumanReview(result.rejectedReasons)
}

What this sample proves

• Native Apple document APIs rather than a web-service wrapper.
• Swift 6 value semantics and explicit Sendable boundaries.
• Honest uncertainty and failure handling.
• Separation of OCR, deterministic extraction, retrieval and generative synthesis.
• Testable RAG guardrails that can fail closed.

What it does not claim

• OCR is not guaranteed correct because a confidence value is high.
• Token overlap is not a replacement for semantic retrieval at large scale; it is a deterministic reviewable baseline.
• RAG does not make an LLM authoritative.
• The sample is not legal, cadastral, medical or financial advice.
• Host applications still own sandboxing, source authentication, PII policy, retention, schema validation and human review.

See ARCHITECTURE.md, THREAT-MODEL.md and VERIFICATION.md.

License

MIT.