The unified interface for every reading.

Only --backend changes between these runs. Same structure back every time: document → pages → blocks, coordinates and confidence on a [0,1] scale, originals kept in bbox_native. Where providers disagree, compare shows you exactly where.

{
  "schema_version": "0.3",
  "status": { "state": "succeeded" },
  "backend": { "id": "pymupdf", "type": "oss_library",
                "output_paradigm": ["block_tree"] },
  "document": {
    "page_count": 2,
    "pages": [{ "page_number": 1, "width": 612.0, "height": 792.0,
                "unit": "pdf_point",
      "blocks": [{ "type": "text", "native_type": "text",
                   "text": "FTCSTMT042 042 20140531 UDSPDFSTMT …",
                   "bbox": { "x": 0.05882, "y": 0.00387, "w": 0.41671,
                             "h": 0.01010, "page": 1,
                     "bbox_native": { "coords": [36.0, 3.064, 291.024, 11.064],
                                      "origin": "top_left", "unit": "pdf_point" } },
                   "reading_order": 0 }] }]
  },
  "warnings": [{ "code": "confidence_unavailable",
                 "field": "block_confidence",
                 "message": "PyMuPDF is a deterministic parser;
                              per-element confidence does not exist" }]
}

{
  "schema_version": "0.3",
  "status": { "state": "succeeded" },
  "backend": { "id": "tesseract", "type": "oss_library",
                "output_paradigm": ["element_list"] },
  "document": {
    "pages": [{ "page_number": 1, "width": 1275.0, "height": 1650.0,
                "unit": "pixel",
      "blocks": [{ "type": "text", "native_type": "line",
                   "text": "Statement Period Date: 5/1/2014 - 5/31/2014",
                   "confidence": 0.9,        // tesseract reports 0–100
                   "text_type": "printed",
                   "bbox": { "x": 0.67922, "y": 0.01697, "w": 0.26431,
                             "h": 0.00667, "page": 1,
                     "bbox_native": { "coords": [866.0, 28.0, 1203.0, 39.0],
                                      "origin": "top_left", "unit": "pixel",
                                      "dpi": 150.0 } },
                   "reading_order": 0 }] }]
  },
  "warnings": null
}

{
  "schema_version": "0.3",
  "status": { "state": "succeeded" },
  "backend": { "id": "aws-textract", "type": "hosted_api",
                "operation": "AnalyzeDocument",
                "output_paradigm": ["block_graph"] },
  "document": {
    "pages": [{ "page_number": 1,
      "blocks": [{ "type": "title",
                   "text": "Loan Application",
                   "confidence": 0.994,
                   "bbox": { "x": 0.1, "y": 0.05, "w": 0.5,
                             "h": 0.04, "page": 1 } }] }]
  },
  "typed_fields": { "Total": { "value": "$4,400.00",
                                "confidence": 0.965 } }
}

The local tabs are real openreading parse output; the aws-textract tab is a captured response converted the same way. And note warnings on pymupdf: it cannot produce confidence, so the field is left out and a warning says why. No value is ever made up.

Parse gets any document into one schema. Compare tells you which provider to trust. Strategy encodes that decision so it keeps happening without you. And each feature works for three audiences: a human at a terminal, a program, or an agent.

Every cell speaks the same JSON Schemas. The product grows by deepening the cells it has, not by adding boxes. A new provider deepens parse. A new verdict deepens compare.

A strategy is your answer to "what happens when it fails?", written down and versioned with your code. It is testable offline. If you have no openreading.yaml, nothing changes.

version: 1

strategies:
  cheap_first:             # local first; escalate only when the result looks bad
    try: [pymupdf, reducto]
    escalate_when: looks_bad

  hosted_duel:             # run both, keep the better result
    parallel:
      - reducto
      - azure-document-intelligence
    pick: best

On a digital PDF, cheap_first accepts pymupdf and reducto never runs. On a scan, the gate fires, reducto runs, and the response says so with a quality_escalated warning. explain shows why a run went the way it did; replay re-runs it offline; calibrate tunes thresholds from your own documents.

Deterministic by default. LLM decision points stay off unless an operator turns them on, and even then they only pick between options you wrote.

Most document tools assume a human reads the output. openreading assumes the caller might be an LLM, so the contract is built for that.

One briefing file. Point your assistant at docs/llms.md and it can drive the whole product. Providers, flags, exact JSON shapes and recipes are all in there.

Outputs you can branch on. Status is a typed state; a partial result says "partial", never a bare success; warnings carry codes; a compare verdict is EQUIVALENT or DIVERGENT with a score. And every strategy run leaves a typed, replayable trace, so an agent can show its work afterwards.

The callable surface is coming soon. That means openreading mcp and triage. What you get today is the briefing, the typed contracts and the traces.

The reading itself is identical in all four. Same connectors, same strategies, same output. What changes is who operates it and whose provider accounts get called. The open-core version leaves out only what someone else running it for you would provide: user accounts, provisioned keys, uptime and administration.

$ pip install 'openreading[pymupdf,tesseract]'

$ openreading backends                 # who is ready, and which env vars are missing
$ openreading parse loan.pdf --backend pymupdf          # no key, no network
$ openreading parse invoices/ --backend pymupdf         # a directory → one batch JSON
$ openreading compare loan.pdf --backends pymupdf,tesseract --format diffs

import openreading

doc = openreading.run("loan.pdf", backend="pymupdf")
doc.keys()
# schema_version, status, backend, document, usage,
#   warnings, backend_raw, channel_provenance

① CONTENT   ✔ EQUIVALENT  content shared by all: 1.00 · 0 real misses
② TABLES    table counts differ: {'pymupdf': 1, 'tesseract': 0}

Same JSON Schemas on every surface: the command line, Python, and an HTTP API you run yourself (openreading serve → /v1/parse, /v1/batch, /v1/compare, /v1/jobs). A web UI comes with it, at openreading serve-ui. All of that is in the package. Hosting it for you is the separate thing.

Pre-1.0, not yet on PyPI. See ways to run it for what is available today and what is not. No customer logos because there are no customers yet.