Chandra OCR Model: Complex Tables, Forms, and Handwriting

What Happened

Datalab released Chandra, an OCR model that preserves full document layout while parsing complex tables, forms, and handwritten text. The model is distributed via the company's GitHub repository under an open license, positioning it as a self-hostable alternative to closed document-intelligence APIs. Chandra is positioned specifically for document-heavy RAG pipelines and structured data-extraction workflows, where downstream retrieval quality depends on retaining table structure, reading order, and spatial relationships between elements.

Why It Matters

Most retrieval failures in enterprise document AI trace back to the ingestion layer, not the model layer. When OCR collapses a multi-column form into a single text stream or discards table cell boundaries, every downstream chunk, embedding, and retrieval result inherits that corruption. Layout-preserving parsing converts documents into structured representations that chunkers and extractors can operate on without reconstructing spatial semantics from scratch. An open model in this category reduces dependence on per-page closed APIs, which carry cost, rate limits, data-residency constraints, and vendor lock-in. For regulated industries — legal, insurance, healthcare, financial services — self-hosting the ingestion layer is often a compliance requirement, not a preference.

Technical Details

Chandra outputs structured representations that retain table grids, form field positions, and reading order rather than flattening pages to plain text. It handles handwriting alongside printed content in the same pass, which matters for scanned forms and annotated contracts that mixed pipelines typically route to separate models. The repository provides weights and inference code for local deployment, making it compatible with air-gapped and on-premise environments where external API calls are disallowed. As with any OCR model, accuracy degrades on low-resolution scans, heavy skew, unusual scripts, and dense multi-language documents; operators should benchmark against their own corpora rather than assume parity across document classes. Throughput and memory footprint depend on the deployment hardware and batching strategy, which the repository documentation covers.

Operational Impact

Teams currently paying per-page fees to closed OCR APIs can now run ingestion in-house, converting a variable per-document cost into fixed compute cost — a material shift at volume. Document-heavy RAG pipelines gain a cleaner input stage: table-aware chunking becomes viable, reducing the retrieval noise that comes from mangled tables and merged columns. Data-extraction workflows for invoices, claims, and contracts can move from brittle regex-over-flat-text to structure-aware parsing, cutting downstream validation and correction work. The practical constraint shifts from API quotas and per-page budgets to GPU provisioning, model serving, and evaluation infrastructure. Operators who already run local inference stacks can integrate Chandra without new vendor relationships; those without will need to stand up serving capacity before the cost advantage materializes.