Windows setup guide / Updated
Local AI setup guide Windows + NVIDIA RTX

Extract PDF and image data. On your RTX GPU.

Turn PDFs and images into editable fields and tables with the RTX GPU in your PC. Connect Ollama to Dynamite Docs, check the source, and export the data you reviewed.

Set up local extraction Open workspace

Included on Starter, Pro, Scale. Requires the Windows companion and a compatible Ollama vision model.

By Dynamite Docs Editorial Team
Super fast extractions with your RTX GPU on your Windows PC with Dynamite Docs
  1. 01

    Add a document

    Images & PDFs

  2. 02

    Extract locally

    Your RTX + a vision model

  3. 03

    Review & export

    Editable fields & tables

NVIDIA GeForce RTXWindowsNVIDIA NIM logoNIMOllama Companion

The short version

Install Ollama and a vision model. Start the Windows companion. In AI Settings → Local, select that model. Add a PDF or image, run local extraction, check the source, and export.

01 / Local extraction guide

Start with the PC you already have.

You need a local model that can read images, a bridge to the workspace, and a plan with local processing.

Windows PC + NVIDIA RTX

Use a supported RTX GPU and an up-to-date NVIDIA driver. Available GPU memory determines which models fit.

Ollama + a vision model

Install Ollama separately. Choose an image-capable model for photos and scanned PDFs, not a text-only model.

Dynamite Docs companion

The Windows companion connects the web workspace to Ollama on the same PC. Keep both apps running.

Starter, Pro, Scale access

Local Ollama is included on these plans. Free does not include local companion inference. Sign-in and account services need internet access.

Check Ollama’s GPU support and Dynamite Docs plan access before installing a large model.

02 / Local extraction guide

From installation to local model ready.

Follow these steps on the Windows PC that will run the extraction. You only need PowerShell for the Ollama commands.

  1. 1

    Install Ollama on your Windows PC

    Download Ollama for Windows and start it. After installation, open a new PowerShell window so the ollama command is available. Keep Ollama running in the background.

    Get Ollama for Windows
    PowerShell · confirm installation
    ollama --version

    A version number appears. If the command is not found, reopen PowerShell after installing Ollama.

  2. 2

    Download a model that can read images

    Start with a local vision model such as gemma3:4b. This is an example to try on your documents, not an accuracy guarantee. The smaller Gemma 3 1B and 270M variants are text-only, so do not use them for scans or image extraction.

    Check Gemma 3 model details
    PowerShell · download and list models
    ollama pull gemma3:4b
ollama list

    gemma3:4b appears in the installed model list. Leave enough disk space and memory for the model and document images.

  3. 3

    Install the Dynamite Docs Windows companion

    Install DynamiteDocs by SevenOneLabs from Microsoft Store. Launch the companion and keep it running alongside Ollama. You do not need Node.js, Python or the source code.

    Get the Windows companion

    The companion is running on the same PC as Ollama. Closing its window leaves it in the system tray; Quit stops it.

  4. 4

    Choose the local model in AI Settings using Dynamite Docs

    Sign in to Dynamite Docs on an eligible plan. Open AI Settings, select Local, and find Your computer (private). If it says Local model offline, choose Check Again. Under Default local model, select the vision model you downloaded.

    Open your workspace

    The panel shows Local model ready. Keep the companion and Ollama open while you extract.

03 / Local extraction guide

Make sure your RTX is doing the work.

Ollama chooses the hardware for the model. Check the loaded model while extraction is running, using another PowerShell window.

PowerShell · while a model is running
ollama ps

Read the PROCESSOR column

100% GPU

The model is loaded entirely in GPU memory.

CPU/GPU

The model is split between system and GPU memory. A smaller model may fit fully on your RTX.

100% CPU

The model is in system memory. Check driver support, model size and available GPU memory.

An empty list means no model is loaded at that moment. See Ollama’s explanation of GPU memory use.

04 / Local extraction guide

Try one document. Check every row.

Start with a short PDF or a clear invoice image. In your workspace, import the file, choose Local (Ollama) for extraction, and run the model you selected in AI Settings.

Invoice → reviewed fields and rowsIllustrative example

Invoice number
INV-2048
Document total
1,432.00
Line items
2 rows
Illustrative extracted invoice rows to check against the source
DescriptionQtyUnit priceAmount
Design services11,200.001,200.00
Print materials2116.00232.00

Before you export

Check the invoice number, decimal places, row count and total beside the original. Correct missed or misplaced values in the fields and table editor.

After your review

Export the checked result as Excel, CSV or JSON. Keep document fields separate from repeating line items when the output calls for it.

Extract my first document Download a sample invoice

The downloadable sample is separate from the illustrative invoice above. Check your model’s actual results against the sample source.

05 / Local extraction guide

Pick for your document, then your GPU.

A model that loads successfully still needs to return complete fields and rows. Compare results on the same document before using it for recurring work.

Starting example

Gemma 3 · 4B vision

The gemma3:4b model accepts text and images. The 1B and 270M versions do not accept images. Try the 4B version on a short, legible document before considering a larger vision model.

Model availability, memory use and output quality vary. Download size is not a VRAM requirement, and a larger model is not a guarantee of a better extraction.

  1. 1.Confirm image support for the exact model tag you install.
  2. 2.Check small text, multi-line descriptions and table boundaries.
  3. 3.Compare missing rows and correction effort, not just how fast the first answer arrives.
How to evaluate extraction quality

06 / Local extraction guide

Local inference, with a clear boundary.

The model runs on your PC. Your account and web workspace still use hosted Dynamite Docs services.

On your PC

For the local inference route, document bytes go from the browser to the loopback companion, then to Ollama. That model request does not call a hosted AI provider.

Still hosted

Sign-in, plan checks, token issuance and workspace services require internet access. Cloud-backed file storage is separate from the local model choice.

Choose browser-only file storage when source files should stay on your computer, and check your storage selection before importing. Local model inference does not make the entire application offline or self-hosted.

Review storage and processing controls

07 / Local extraction guide

Get unstuck without starting over.

AI Settings says Local model offline

Start the Windows companion and Ollama on this PC, then select Check Again in AI Settings → Local. Confirm that you are signed in on an eligible plan. If the session has expired, sign in again before retrying.

The model list is empty

Run ollama list in PowerShell. If your vision model is missing, finish downloading it with ollama pull gemma3:4b. Then return to AI Settings and choose Check Again. Installing Ollama alone does not install a model.

Extraction runs on the CPU or feels slow

Check ollama ps while the model is running. CPU/GPU means the model is split across system memory and GPU memory. Close other GPU-heavy apps, use a smaller vision model, and check your NVIDIA driver against Ollama’s hardware support documentation.

A scan produces missing fields or broken rows

Check that the selected model supports images. Use a straight, legible scan and start with one page. Review column boundaries, decimal places and repeated headers. Try another compatible vision model if the same document keeps returning incomplete results.

A large document times out

Start with a shorter document or a smaller vision model. The companion has a default 120-second extraction timeout. Memory, image batches and document complexity affect what your PC can finish within that time.

More extraction troubleshooting

08 / Local extraction guide

A few things you might be wondering.

Can I use my own NVIDIA RTX GPU for PDF and image extraction?
Yes. Ollama runs the model on your Windows PC and handles GPU acceleration on supported hardware. The Dynamite Docs companion connects that local model to the extraction and source review workflow. Use ollama ps during inference to check whether the model is using GPU memory.
How much VRAM do I need?
There is no single VRAM requirement for every model and document. Model size, quantization, context length and document images all affect memory use. Start with a smaller vision model, check Ollama’s hardware guidance, and use ollama ps to see whether the loaded model fits on the GPU or is split across CPU and GPU.
Which plans include local Ollama?
Starter, Pro, Scale include the local companion. Free does not. This entitlement is separate from BYOK.
Can I use Ollama for offline document extraction?
Internet access is required for the web workspace, sign-in, and account services. Ollama model inference runs locally on your PC. Download Ollama and your chosen model separately.
How do I sign in for local extraction?
Sign in to Dynamite Docs in your browser. On an eligible plan, open AI Settings and connect the local companion to select an Ollama model.
What is the best Ollama model for document extraction?
Choose a vision model that fits your hardware and returns the required structured response. Compare candidate models on the same scans, small text and multi-page tables. Measure missing fields, row completeness and correction time rather than choosing from a single OCR example.
How do I run Ollama OCR locally on Windows?
Install Ollama and a compatible vision model, then launch the Windows companion. On Starter, Pro, Scale, sign in to the web workspace, connect the companion in AI Settings and select the local model. Import a PDF or image, run extraction and check the result beside the source.
Can Ollama convert a PDF or image to Markdown?
Ollama PDF to Markdown conversion generally requires rendering scanned pages as images for a suitable vision model. Image to Markdown conversion also depends on model output. The Dynamite Docs companion returns structured extraction data. For selectable PDF text, use the separate local PDF to Markdown converter, which runs in the browser without Ollama.
How should I prepare batch PDF extraction or batch image OCR with Ollama?
Test representative files first, including the weakest scan and longest table. Keep a consistent schema, record the expected pages and rows, and review failed or incomplete results separately. The companion processes scanned PDF pages in image batches within its request limit. That page batching is separate from stored multi-file jobs; check the processing route before assuming a job uses local inference.
Where can I find an Ollama OCR Python tutorial or API integration example?
The official Ollama vision API guide (opens in a new tab) includes Python and REST examples for sending images to a vision model. A PDF workflow also needs text extraction or page rendering and output validation. Dynamite Docs connects through the Windows companion without a Python script. Its authenticated local routes are separate from the hosted REST API.
Can I use DeepSeek OCR with Ollama on Windows?
The DeepSeek OCR model page (opens in a new tab) documents image OCR and Markdown prompts. Follow the Ollama Windows setup guide (opens in a new tab) to run a local model. Test its response against the companion extraction contract before choosing it for Dynamite Docs. Text or Markdown output alone does not establish compatibility with the structured fields and rows expected by the companion.
Can I run document extraction locally with Ollama without exposing data to external APIs?
The model-inference request stays on the local loopback path between the browser, companion and Ollama. Account, token and workspace services remain hosted, so review the full storage and retention path before treating the workflow as self-hosted.

Sources & next steps

Ollama documentation checked 9 November 2026. Product steps reflect the Dynamite Docs Windows companion and current local-processing entitlements.

Review PDF fields and tablesChoose a private processing route

Topic Cluster · In This Series

In This Series: AI Document Extraction Guides

Explore all 6 connected workflows in this series to master table extraction, invoice processing, data privacy boundaries, and local Ollama inference.

View guides hub
Core Workflow

Finance Document Extraction

Turn piles of mixed business paperwork into verified line items and tables with source-linked human check.

Explore guide
Multi-Page PDFs

PDF Data Extraction to Spreadsheets

Extract PDF tables and line items beside original pages, resolving wrapped rows and page-break continuations.

Explore guide
Tables & Forms

Complex PDF Forms & Nested Grids

Parse borderless tables, dense fillable forms, and multi-tier subtotal grids into normalized structured data.

Explore guide
Invoices to XLSX

Invoices to Audit-Ready Excel

Separate invoice headers from repeating item lines into a clean, audited Excel workbook with math checks.

Explore guide
Data Security

Private Extraction & Zero Retention

Choose your processing path: zero-retention cloud models, client-isolated BYOK keys, or local loopback.

Explore guide
Current guideOn-Device AI

Local RTX GPU Inference with Ollama

Run vision models on your Windows PC via the companion app with zero external API calls for document bytes.

You are reading this workflow

Looking for other workflows? AI document data extraction and OCR guides covers all supported formats and procedures.

Browse all 6 workflows in this cluster
Dynamite DocsNVIDIA GeForce RTX

Your next document

Give your Nvidia RTX card something useful to read.

Install the companion, select your local vision model, and review your first extracted table.

Get the Windows companion

Can we save optional cookies?

Essential cookies keep you signed in. With your permission, optional analytics cookies and masked session replays help us understand usage and diagnose errors. Privacy policy. Change your choice anytime in Cookie settings.