Why specialist document AI matters for historical collections

Documents can now be uploaded to frontier AI models such as Gemini, Claude, GPT, Qwen, Llama, Mistral, and other multimodal systems, and they can return surprisingly detailed results. These models are extremely powerful. We use them ourselves.

But historical collections are different from ordinary modern documents.

They are often difficult because of their language, layout, handwriting, abbreviations, damaged paper, faded ink, marginal notes, complex tables, obsolete terminology, multilingual content, and inconsistent structure. A model may produce an impressive-looking result while missing important details, misunderstanding the source, applying the wrong structure or "corrections", or failing silently on the parts of the document that matter most.

Successful document AI is therefore not just a question of choosing a powerful model. It depends on knowing the source material, selecting the right technology, preparing the images correctly, combining OCR, HTR, layout analysis, visual language models, validation tools, and structured extraction pipelines, and designing the workflow around customer data requirements - precisely.

This is what Osiris does.

We combine historical source knowledge with machine learning development and research-oriented production. We help clients choose the right models, adapt them where needed, and build the data pipeline required to produce reliable outputs at scale.

The AI model landscape is also changing rapidly. Hugging Face now hosts millions of models, with specialist communities forming around particular domains, languages, and problem types. The challenge is no longer simply finding a model; it is knowing which model, workflow, and validation process is right for a specific collection. Osiris helps clients navigate this landscape, select appropriate models, fine-tune them where required, and build a model and workflow that just works.

What Osiris provides

We work with complex archival and historical material, including handwritten records, early printed texts, tabular sources, administrative documents, registers, newspapers, maps, forms, and multilingual collections.

We assess which models and tools are appropriate for each collection, including OCR, HTR, layout analysis, multimodal AI, and specialist open-source or commercial models. Where required, we train or fine-tune models for the script, layout, language, and structure of the source.

Off-the-shelf tools are useful, but many collections require a bespoke or adapted workflow. We build the model and pipeline needed for your documents, your accuracy requirements, and your final data use.

We manage the full process from image preparation and layout analysis through to OCR/HTR, structured data extraction, validation, reporting, and export.

We do not simply upload documents to a model and accept the answer. We combine models with preprocessing, segmentation, batching, extraction logic, validation, review tools, and export formats designed for production use.

We identify where the workflow performs well, where errors are likely, and what level of review or correction is needed for the intended use. This helps isolate strengths and weaknesses before and during large-scale processing.

Where required, we can process collections from start to finish on local or private servers, avoiding unnecessary data transfer to third-party AI platforms and reducing the risk of data leakage.

We deliver data in formats suited to search, research, publication, discovery systems, databases, maps, catalogues, or customer ingestion pipelines.

Our multi-stage pipeline for document understanding.

Osiris is not a single LLM system. It is a multi-stage pipeline combining image preparation, custom HTR/OCR, segmentation, structured extraction, validation, and feedback loops to process complex documents reliably at scale.

Each stage is engineered to address specific failure modes — preserving structure, reducing hallucinations, and verifying outputs against the source. This allows us to deliver accurate, auditable results on material where generic AI tools break down.

Image prep

Sampling, de-warp, de-skew, crop, and enhancement for cleaner inputs.

HTR / OCR fine-tuning

Assisted training around your collection and document patterns.

Pre-segmentation

Split pages into useful zones, records, and fields before extraction.

Structured extraction

Targeted prompts, smaller inputs, deterministic checks, and review-ready outputs.

Structured output

Corrected text, structured data, and error metrics. We can also provide readable formats like PDF.

Accuracy you can measure — and trace.

General-purpose LLMs can transcribe reasonably well — but they also invent plausible, correct-looking text. Osiris measures and controls this risk explicitly.

Osiris gets the best from fast-evolving AI through research and development to provide trustworthy and useful information from complex records.

We regularly benchmark our models against competing systems (including industry leaders). The examples below report overall accuracy (F1) against ground truth and break errors down by type — isolating hallucination risk, numeric failures, and OCR noise.

Overall Accuracy (F1) — 18th C English handwriting
Osiris   ████████████████████   90%
ChatGPT  █████████████████      79%

Error Types (Extra Tokens vs GT)
Plausible errors (easy to miss)
Osiris   ██████                 3.97%
ChatGPT  ██████████████████████ 14.67%

Numeric errors (high impact)
Osiris   █████████████████      4.30%
ChatGPT  ██████████████████████  5.67%

Minor OCR errors
Osiris   ██████████████████████  1.66%
ChatGPT  ██████████████████       1.33%

Other major errors
Osiris   ██████████████████████  0.66%
ChatGPT                           0.00%

Overall Accuracy (F1)
Osiris   ████████████████████   92%
ABBYY    ████████████████       71%
ChatGPT  ███████████████████    85%

Error Types (Extra Tokens vs GT)
Plausible errors (easy to miss)
Osiris   ███████████████        4.66%
ABBYY    █████████████████      5.27%
ChatGPT  ██████████████████████  6.78%

Numeric errors (high impact)
Osiris   ███████                0.62%
ABBYY    ██████████████████████  1.86%
ChatGPT  ███████                0.58%

Minor OCR errors
Osiris   █                      0.52%
ABBYY    ██████████████████████ 14.68%
ChatGPT  ██                     1.05%

Other major errors
Osiris   ████                   1.35%
ABBYY    ██████████████████████  6.83%
ChatGPT  ██                     0.58%