A field by its label
Name an index field after the label printed on the form, such as "Invoice Number", and the model finds the label and reads the value next to it.
Local LLM OCR, on your own hardware
SimpleIndex can hand the fields OCR can't read to a language model running on your own PC or server, through Ollama, LM Studio or vLLM. There's no API key and no per-page bill, and no page leaves your network. Local OCR still reads every page first, so the model sees only what's left.
01
What a model on your own server reads
The same work a cloud model does, with nothing sent out. Reading the page image needs a vision model; the job wizard knows which ones can.
Tables and line items
A table field comes back as rows, such as an invoice's line items, written out as CSV beside the other values.
Ticked boxes and filled circles come back as values.
Ask the model what kind of document a page is, and file or route it by the answer.
Draw a zone around one handwritten box and send just that zone. How well it reads depends on the model and on how neat the writing is, so test it on your own forms.
JSON or XML
Turn on raw model output and the answer comes back in the format your prompt asks for.
Setting it up takes four steps
Run a model server
Install Ollama, LM Studio or vLLM on the same PC or on a server your network reaches. Any server that speaks the OpenAI API works.
Load a vision model
The job wizard suggests Qwen3-VL, Gemma 3, Llama 3.2 Vision and DeepSeek-OCR for Ollama. Any other model name can be typed in.
Pick Custom in the job
In the LLM settings, choose the Custom provider and your model. Left blank, the address points at Ollama on the same computer, and no API key is needed.
Choose which pages it reads
Set a registration field: if local OCR leaves it blank, the page isn't sent to the model, so pages the job isn't meant for never use its time.
03
How it fits with local OCR
Tesseract, SimpleOCR or ABBYY FineReader reads every page first and makes it searchable, and zones, patterns and barcodes fill most fields at no cost. The model is asked only for the fields still empty, and a field it already filled is not asked for again.
Hybrid, the default, sends the model the page image together with the OCR text and where each word sits. Text only sends just the OCR text: lighter on a small server, and it works with models that can't read images. Images only skips local OCR, so the PDF gets no searchable text.
Most documents need no model at all: OCR vs. LLM: you don't need AI for that.
04
What running it yourself means
There's no API key to manage and no provider bill: the cost is the hardware the model runs on. It has been tested on a server with 8 GB of memory, and it runs with no internet connection at all, including on air-gapped networks.
A small model on modest hardware is slower than a cloud model and less sure on hard pages, so send it only the fields that need it, and try it on your own documents first.
Before you set this up
What is local LLM OCR?
Using a language model on your own computer or server to read fields from scanned documents, instead of sending pages to a cloud AI service. In SimpleIndex it works alongside regular OCR, not instead of it.
What hardware does it need?
It has been tested on a server with 8 GB of memory. A small vision model, such as Qwen3-VL 4B or Gemma 3 4B, is the place to start; a larger model or a graphics card makes it faster.
Does it work with no internet connection?
Yes. It runs on air-gapped networks with no internet connection at all.
Which model servers work?
Ollama, LM Studio, vLLM and llama.cpp's server, or anything else that speaks the OpenAI API. DeepSeek-OCR also runs locally through Ollama.
How is this different from AWS Textract or a cloud LLM?
Same job setting, different engine. AWS Textract bills per page and cloud LLM providers bill by tokens, and both need an internet connection. A model on your own server has no per-page cost and no outside service to reach.
Which edition do I need?
Professional. A model on your own server runs under the Cloud OCR license, the same as a cloud model, and Professional includes it; a Standard license that already carries the Cloud OCR add-on can use it too. The trial includes it.
Can the model make my PDFs searchable?
No. Searchable text comes from local OCR, which records where each word sits. The model supplies field values.
Does SimpleIndex need the cloud at all?
No. SimpleOCR, Tesseract and ABBYY FineReader all run on your own computer, and a model on your own server keeps the AI part local too. Pages go out only to an engine you choose, such as AWS Textract or a cloud LLM. See how local-first compares on cost.
Try it on your own server
The free trial runs the full software, local models included, and our team in Knoxville can help you set up the first job.
