2027 SOTA Document Intelligence Operating Model
Date: 2026-06-16 Status: Implementation target Owner: pdf-reader-mcp Related: docs/adr/0001-2027-sota-document-intelligence-boundary.md
Goal
Make @sylphx/pdf-reader-mcp a public, commercial-grade quality, state-of-the-art document-intelligence MCP package for agents, while preserving local-first privacy and a clean boundary from hosted Platform services.
Non-Goals
- Do not add hosted auth, billing, storage, tenancy, or data retention inside the package.
- Do not add direct provider secrets or Sylphx AI Gateway credentials to the package.
- Do not optimize for one agent client at the expense of MCP contract compatibility.
- Do not claim release readiness without docs, tests, benchmarks, and changelog evidence.
Capability Map
| Area | SOTA requirement |
|---|---|
| MCP contract | Stable tool schemas, versioned options, compatibility fixtures, generated/API docs. |
| Text extraction | Page-aware and full-document extraction with deterministic ordering and clear whitespace policy. |
| Layout/provenance | Page, bounding region, image, table, and source metadata sufficient for citations and verification. |
| Visual/OCR adapters | Optional provider interface for OCR and region analysis with local-only default behavior. |
| Batch processing | Multiple sources, partial failure isolation, concurrency limits, cancellation-safe behavior. |
| Performance | Reproducible benchmarks by document class, page count, image density, and operation type. |
| Security/privacy | Path handling, URL handling, file size limits, remote-processing opt-in, no secret logging. |
| Release quality | Typecheck, tests, benchmark gate, docs build, changelog, npm package smoke, compatibility smoke. |
Portfolio Integration Rules
- Gateway may register this package as an external MCP tool source, but Gateway owns any model-facing manifest, native late-loading projection, tenant policy, hosted execution, cache diagnostics, billing, and usage surface around that registration.
- Product apps own product permissions, credential handles, audit, and durable outcomes created after using these tools.
- A hosted multi-tenant document-intelligence product is a separate service, not a mutation of this local MCP package.
Provider Boundary
Provider adapters may exist for OCR, vision, or layout analysis, but they must:
- be optional dependencies or optional runtime adapters;
- accept credential handles from the caller environment, never hardcoded keys;
- declare whether documents leave the local machine;
- return typed, provenance-preserving results;
- expose cost/latency/error metadata where available;
- fail per page/source when possible.
Commercial Paths
The local package remains free/open-source quality software. Commercial value can be created through:
- a hosted document-intelligence service in a separate repo/service with Platform Auth/Billing/Storage/Durable Work controls;
- enterprise support and compatibility certification;
- benchmark-backed performance claims;
- private provider adapters or compliance packs;
- integration bundles for agent platforms and CI/document workflows.
Acceptance Criteria
- Every public tool option is covered by schema tests and docs.
- Extraction results include source and page provenance.
- Visual/OCR provider support is adapter-based and local-only remains the default.
- Benchmarks are reproducible and versioned with hardware/runtime context.
- Release gate runs typecheck, tests, package build, benchmark release gate, docs build, and npm smoke where credentials permit.
- A failed page/source does not hide successful extraction from other pages or sources unless the caller requested fail-fast behavior.
- The README, docs site, changelog, and
progress.mdagree before release.
Implementation Slices
- Add schema-level contract fixtures for existing tools.
- Add provenance normalization for pages, images, and future regions.
- Add provider adapter interface for visual/OCR analysis without bundled provider secrets.
- Add benchmark release gate that separates small, scanned, image-heavy, table-heavy, and large-document classes.
- Add docs explaining privacy modes, local-only behavior, remote-provider opt-in, and commercial hosted boundary.
- Add release checklist that blocks publish when docs, benchmarks, changelog, and package smoke are out of sync.