Ertas for PDF Parsing and Transformation
Train a small model in Ertas Studio that turns text you have already extracted from PDFs into structured output such as JSON fields, normalised records or short summaries. Export it and run it on your own servers, alongside the PDF extraction tool you already use.
The Challenge
Enterprise document archives contain diverse PDF types: scanned, native, mixed layouts, multi-column pages, tables and technical drawings. OCR engines and PDF parsers can pull the text out, but what comes back is raw text that still has to become the fields, records or summaries a downstream system expects. Rule-based scripts for that step break on every new layout, and sending every page to a large frontier model gets expensive at archive scale. Service providers handling client documents need a repeatable way to do that transformation.
The Solution
Ertas covers the step after extraction. Run your PDFs through the extraction tool you already use, then train a model in Ertas Studio that turns the extracted text into the output you need: a JSON object with named fields, a normalised record, a classification or a short summary. Build the training set in Data Craft, with extracted text as the USER turn and the correct structured output as the ASSISTANT turn, or upload pairs you already have as JSONL. Domain experts review and approve the examples there before training.
Studio fine-tunes an open model on Ertas's cloud GPUs through a visual canvas, and several runs can train in parallel so you can compare model sizes on held-out documents. The finished model is exported from Hub as GGUF for llama.cpp or Ollama, or as full safetensors for vLLM, and runs on your own servers next to your extraction step. Documents processed after deployment stay on that infrastructure, and the cost per page is your own compute. Training data is uploaded to Ertas, so build the training set from pages you are free to share, with personal details removed. Ertas does not extract text from PDF files itself, so a scanned archive still needs an OCR or PDF parsing tool in front of the model.
Key Features
Structured Output From Extracted Text
Train a model to map raw extracted text to the fields, labels or format your downstream system expects. The output schema is whatever your training examples show.
Training Pairs in Data Craft
Build examples as USER turns of extracted text and ASSISTANT turns of the correct output. Edit and version them, and track which ones a reviewer has approved before training.
Compare Runs Before You Commit
Run several LoRA fine-tuning runs in parallel on different base models or settings, and check each against held-out documents the training set never saw.
Runs Next to Your Extraction Step
Export the model as GGUF or full safetensors and run it on your own servers with llama.cpp, Ollama or vLLM. Documents processed at inference time stay in your infrastructure, with no per-page API bill.
Example Workflow
An AI consultancy receives a large archive of construction PDFs (tender documents, specifications and bills of quantities) from a client who needs cost line items pulled into its estimating system. The team runs the archive through an open-source PDF extraction tool they already use, which produces raw text and tables for each page. They take a few hundred representative pages and build training examples in Data Craft: the extracted text as the USER turn and a JSON list of line items (description, quantity, unit and rate) as the ASSISTANT turn. A quantity surveyor on the client's side reviews and approves the examples. In Ertas Studio they train three runs in parallel on Qwen 2.5 1.5B, Qwen 2.5 3B and Llama 3.2 3B, check each against a held-out set of pages, and export the smallest model that meets the estimators' accuracy bar as GGUF. It runs with llama.cpp on the consultancy's own server, straight after the extraction step. When the next construction client arrives, the team reuses the same dataset structure and adds examples from that client's documents.
Related Resources
Ship AI that runs on your users' devices.
Start free with 150 credits after card verification. Paid plans from $20/mo USD.