LlamaIndex's OpenDocRouter Puts 10 Document Parsers Behind One API

LlamaIndex has launched OpenDocRouter, a hosted service that places multiple document-parsing models behind one HTTP endpoint. Developers submit a document, page range, and model identifier, then receive Markdown through a consistent response format. The service aims to reduce the integration work required to test and switch optical character recognition and document-understanding models.

Document parsers vary in prompts, deployment requirements, rate limits, output formats, and support for tables or page layouts. OpenDocRouter packages those differences into versioned “parsing recipes” that pair each model with its prompt and post-processing rules. Applications can change backends while keeping the surrounding pipeline intact.

Ten backends, one request shape

LlamaIndex’s launch catalog lists ten models, divided between frontier APIs and open-source parsers:

A basic request includes a bearer token, model ID, document URL, and optional page range:

curl https://www.opendocrouter.ai/v1/parse \
  -H "Authorization: Bearer $OPEN_DOC_ROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemini-3.8-flash-low",
    "document": {
      "url": "https://arxiv.org/pdf/1706.03762"
    },
    "pages": "1-3"
  }'

Changing backends requires replacing the model value while preserving the request and Markdown response structure. LlamaIndex publishes a TypeScript SDK under the @llamaindex/opendocrouter namespace and says a Python SDK is also available.

Page failures stay local

OpenDocRouter schedules each page as an independent model call when capacity becomes available. Every page receives its own Markdown output, status, and charge, allowing clients to retry failed pages or skip pages already processed. A failure near the end of a large PDF therefore requires a targeted retry instead of another full-document run.

Setting layout: true adds grounded bounding boxes and a shared layout-class vocabulary across the catalog, including models that lack native layout output. That normalized structure can support spatial chunking, highlighted citations, document viewers, and extraction pipelines that depend on coordinates.

Model choice changes the bill

LlamaIndex evaluates the catalog with ParseBench, its internal benchmark suite. The published results compare quality and cost across tables, charts, formatting, faithfulness, and grounding. These category-level scores can help teams narrow the field before testing models against their own invoices, reports, forms, or research papers.

ParseBench remains a vendor-operated benchmark, so production evaluations should include representative documents, expected Markdown structures, latency measurements, and field-level accuracy checks. A shared schema simplifies integration, while model behavior can still vary on merged cells, handwriting, rotated scans, footnotes, and reading order.

The wide price range supports per-document routing. A team could send clean, repetitive invoices to a lower-cost open-source parser and reserve a more expensive frontier model for dense tables or visually complex reports.

The router follows a fast-moving field

LlamaIndex already operates LlamaParse, maintains the open-source LiteParse engine, and has worked with Kaggle on a document OCR leaderboard. OpenDocRouter extends that document tooling into a model-aggregation service, using the company’s evaluation and parsing layers to expose third-party backends through one contract.

Rapid releases from frontier labs and open-source projects have made parser selection a recurring engineering task. Teams building retrieval-augmented generation systems, search indexes, or extraction workflows often maintain several adapters to compare new models. A router moves that variation into the service layer and reduces the code required for evaluations, fallbacks, and cost-based routing.

Where it fits, and what to verify

• Model testing: One request format supports controlled comparisons across frontier and open-source backends.
• Cost routing: Applications can select a parser according to document type, expected complexity, or budget.
• Targeted recovery: Page-level status and billing limit retries to failed work.
• Layout-aware output: Normalized bounding boxes give downstream systems a common coordinate format.
• Managed open models: Teams can use MinerU or PaddleOCR-VL without operating GPU infrastructure.

Adopting the router adds another network dependency and gives LlamaIndex access to the documents being processed. Security reviews should cover retention periods, encryption, data residency, provider subprocessors, rate limits, service-level commitments, recipe versioning, and deletion controls. Teams with strict compliance requirements should also confirm whether individual backends send document content to external model providers.

Page-by-page inference can affect documents whose meaning crosses page boundaries, including continued tables, repeated headers, and footnotes. Evaluations should measure how reliably downstream processing reconstructs those relationships and whether recipe updates change established outputs.

The service fits teams that regularly compare parsers or maintain several OCR integrations. Organizations that need custom prompts, direct control over model infrastructure, or tightly constrained data handling may prefer their existing integrations despite the added maintenance.