Ertas for Document Classification

The Challenge

Organizations process thousands of documents daily, including contracts, invoices, correspondence, reports, applications, and compliance filings, and routing each document to the right team or workflow depends on accurate classification. Manual classification is slow, inconsistent, and scales poorly. When a single misrouted document can delay a legal filing or lose a time-sensitive business opportunity, the cost of errors is significant.

Generic AI models struggle with document classification in specialized domains because they lack context about an organization's specific document taxonomy. A general model might distinguish between an invoice and a contract, but it cannot reliably differentiate between a master services agreement and a statement of work, or between a regulatory filing and an internal compliance memo. These fine-grained distinctions require domain knowledge that can only come from training on the organization's actual document corpus, which is exactly the kind of task fine-tuning is designed to solve.

The Solution

Ertas Studio lets organizations fine-tune classification models on their own document taxonomy using real examples from their archives. Teams upload labelled document samples as JSONL, where each entry pairs document text with its correct category, and train an open model with LoRA through Studio's visual canvas. The model learns the vocabulary, structure and patterns that separate one document type from another in that organization's taxonomy. Several runs can train in parallel, so the team can compare base models and settings against a held-out set before choosing one.

The trained model is listed in Hub and downloads as GGUF for llama.cpp or Ollama, or as full safetensors for vLLM. Run it on your own servers and the documents being classified stay inside your network. Training happens on Ertas's cloud GPUs, so the labelled examples are uploaded to Ertas: remove or mask personal and confidential details first, and delete the dataset when you no longer need it. When the taxonomy changes, with categories added, split or merged, the team updates the examples, trains again and swaps in the new model. Teams that already send documents to a frontier model API can instead use the adaptive model endpoint, a sales-led service where a custom model trained on their own requests takes over a growing share of the work once it passes their evals.

Key Features

Example Workflow

A large insurance company receives 10,000+ documents daily across email, fax and web portal channels. The documents include new claims, policy amendments, medical records, adjuster reports and legal correspondence, each needing to reach a different department. The team exports 50,000 labelled examples from its archive, masks policyholder names and identifiers, and uploads the JSONL dataset to Ertas Studio. They fine-tune Qwen 2.5 7B and Llama 3.1 8B with LoRA in parallel against their 28-category taxonomy, then compare both runs and a generic model on a held-out test set the models never saw. The better model is exported from Hub and runs on the company's own servers behind its document intake system, returning a category for each incoming document. High-stakes categories, and any output that does not match a known category, go to a person for review, and those reviewed documents become training examples for the next run.