Turn your documents into a searchable AI workspace. In this project, you will set up a browser-based assistant on the Acer Veriton GN100, add your own reference files, and ask questions with supporting source passages.

Think of a student comparing research papers, a developer searching technical manuals, or a hobbyist organizing project notes. Instead of opening every file separately, you can ask a focused question and inspect the evidence behind the answer.

What you will build

The assistant uses retrieval-augmented generation (RAG): software extracts text from your files, builds a searchable index, and retrieves passages relevant to a question. A language model then uses those passages as context. Adding a document does not retrain the model.

Acer lists the VGN100-UD13 with the NVIDIA GB10 Grace Blackwell Superchip, 128GB LPDDR5X memory, and DGX OS. This guide starts with a smaller model to keep the first experiment manageable. It does not represent the largest model the workstation can run, and a basic document assistant is also possible on less powerful hardware.

Before you begin

• Complete the GN100’s first-run setup and any manufacturer-recommended updates.
• Use the GN100 locally with a display, keyboard, mouse, and browser. All terminal commands below run on the GN100.
• Have internet access for downloading software and model files.
• Check that Docker and the NVIDIA container runtime are configured for GPU access.
• Reserve free storage for the container, models, documents, and index. As a planning allowance, keep at least 40GB free for this starter setup; actual usage varies.
• Start with English plain-text files or PDFs with selectable text. Scanned PDFs need an OCR workflow that is outside this tutorial.

No paid cloud AI API is required for this configuration. Setup time depends heavily on download speed and whether your system prerequisites are already working.

Step 1 — Check the workstation

Open Terminal on the GN100. Run each command separately:

uname -m
nvidia-smi
docker --version
sudo docker ps

The architecture check should report aarch64. The NVIDIA command should display GPU and driver information. Docker should return its version and a container table; an empty table is fine.

Step 2 — Download the application

Use the Open WebUI image that bundles Ollama, following the same installation approach documented in NVIDIA’s DGX Spark playbook:

sudo docker pull ghcr.io/open-webui/open-webui:ollama

Wait for the download to finish. The :ollama tag is a rolling image, so its software can change. After you validate your setup, record the image digest and software versions before publishing or reproducing the project.

Step 3 — Start your local assistant

Copy the full command below into Terminal:

sudo docker run -d \
  --name gn100-doc-assistant \
  --restart unless-stopped \
  --gpus all \
  -p 127.0.0.1:3000:8080 \
  -v gn100-webui-data:/app/backend/data \
  -v gn100-ollama-models:/root/.ollama \
  -e OLLAMA_BASE_URL=http://127.0.0.1:11434 \
  -e RAG_EMBEDDING_ENGINE=ollama \
  -e RAG_EMBEDDING_MODEL=nomic-embed-text \
  -e RAG_OLLAMA_BASE_URL=http://127.0.0.1:11434 \
  ghcr.io/open-webui/open-webui:ollama

This configuration requests GPU access, keeps application data and models in persistent Docker volumes, and connects both chat and embeddings to the bundled local Ollama service. The web interface is bound to the GN100’s loopback address, so the published port is accessible from that machine.

Check the container:

sudo docker ps --filter name=gn100-doc-assistant
sudo docker logs --tail 80 gn100-doc-assistant

Allow time for the initial startup. If GPU initialization fails, fix the NVIDIA container runtime before continuing.

Step 4 — Create your local account

Open a browser on the GN100 and visit http://localhost:3000. Select Get Started and create the initial administrator account. Use a strong password, and keep authentication enabled.

Here, localhost means the GN100 itself. Opening that address on your laptop will not reach the workstation with this configuration.

Step 5 — Download the models

Return to Terminal and run these commands one at a time:

sudo docker exec gn100-doc-assistant ollama pull qwen3:8b
sudo docker exec gn100-doc-assistant ollama pull nomic-embed-text
sudo docker exec gn100-doc-assistant ollama list

Confirm that both models appear in the list. Refresh Open WebUI and choose qwen3:8b in the model selector. Do not select the embedding model for chat.

Send a simple test message:

Explain retrieval-augmented generation in three simple sentences.

While the model is generating, inspect its processor allocation:

sudo docker exec gn100-doc-assistant ollama ps

Check the PROCESSOR column for GPU use. A successful chat verifies inference, but does not yet verify document retrieval.

Step 6 — Check local document processing

Open Admin Panel → Settings → Documents, or the equivalent menu in your version. Confirm the following before uploading files:

Save the settings. Existing saved admin settings may override startup environment variables, so check the interface even if you used the command above. If you change the embedding model later, reindex the documents with the new model.

Step 7 — Make a document with known answers

Create a plain-text file named project-notes.txt in your text editor. Paste in this fictional example and save it:

Project Cedar — Internal Reference Notes

Project owner: Maya Chen.
The prototype review is scheduled for November 18, 2026.
The approved prototype budget is $7,250.
The team meets every Tuesday at 10:30 AM.
The final report must include a performance summary,
known limitations, and recommended next steps.

Using a small file with known answers makes it easier to tell whether the assistant is retrieving your content or guessing.

Step 8 — Create a knowledge base

• Open Workspace → Knowledge → Create.
• Name it My Private Documents and add a short description.
• Keep its access limited to your account.
• Upload project-notes.txt.
• Wait until processing has completed successfully.

A knowledge base lets you reuse the same collection across conversations. Uploading a document alone does not guarantee that a particular chat will use it; attach the collection in the next step.

Step 9 — Attach the knowledge and ask a question

For a predictable first experiment, use traditional retrieval: in the selected model’s settings or advanced parameters, set Function Calling to Legacy and keep File Context enabled. Depending on the version, these controls may be under Workspace → Models → Edit or the chat controls. This avoids requiring the model to invoke knowledge-search tools on its own.

Start a new chat with qwen3:8b. Type # in the message field, then select My Private Documents. Confirm that the collection appears as an attachment, and send:

Use the attached documents to answer this question:
What is the approved prototype budget for Project Cedar?
Include the source reference and the supporting passage.
If the information is missing, say so. Do not guess.

The expected answer is $7,250. Open any source reference and compare its supporting text with the original file. Source presentation varies by version; do not assume every file will have reliable page numbers.

Step 10 — Test accuracy and missing information

Test the last question explicitly:

Which vendor was selected for Project Cedar?
Use only the attached documents.
If the answer is missing, say:
"I could not find that information in the provided documents."
Do not guess.

A useful assistant should acknowledge missing information. Prompt instructions reduce guessing but do not guarantee accuracy, so verify important answers and their citations.

Step 11 — Add your own reference files

After the sample works, add one to three short, text-based PDFs to the same knowledge base. Choose documents you are authorized to process. Ask specific questions before attempting a broad summary of a large collection.

Using the attached manual, list the installation requirements.
Separate required items from optional items.
Include supporting source references.
If a requirement is unclear or missing, say so.

Complex tables, multiple columns, images, and scanned pages can affect text extraction. Inspect the extracted content when answers are incomplete. For this starter setup, keep documents and questions in English; evaluate multilingual retrieval separately before relying on it.

Step 12 — Verify offline operation

A local interface is only one part of a private workflow. Chat inference, embeddings, document extraction, and any optional tools must also remain local. Keep web search, hosted model connections, and external tools disabled for this experiment.

• Download both models and complete a successful document question while online.
• Temporarily disconnect the GN100 from the internet.
• Open a new conversation, attach the existing collection, and ask another question.
• Upload a new plain-text file and ask about its contents.
• Confirm that both existing-file retrieval and new-file processing still work.

This checks that the tested workflow can operate offline; it is not a complete network or security audit. Application data, documents, and indexes remain on disk, so account access, disk protection, and backups still matter.

Using the assistant again

Start or stop the existing container with:

sudo docker start gn100-doc-assistant
sudo docker stop gn100-doc-assistant

These are separate actions: use start to run the assistant and stop when you want to shut it down. Reopen http://localhost:3000 after starting it. Do not run the original creation command again for normal daily use.

Stopping the container preserves the named volumes. Back up application data and documents before upgrades. Removing a volume deletes its stored data.

Troubleshooting

Where to take the project next

Once the basics are reliable, try separate collections for research, technical manuals, and personal projects. Compare answers from different local models, evaluate multilingual embeddings, or add a local OCR pipeline for scanned documents. Measure answer quality on a set of known questions before expanding the collection.

The Acer Veriton GN100 gives developers room to explore local AI workflows beyond a single starter assistant. This project offers a practical first goal: bring your own documents, ask useful questions, and check the evidence behind the answers.

Explore the Acer Veriton GN100 at AAAWave →

Official references

• Acer VGN100-UD13 specifications
• NVIDIA: Open WebUI and Ollama on DGX Spark
• Open WebUI installation guide
• Open WebUI knowledge bases and document chat
• Open WebUI RAG configuration
• Open WebUI offline preparation
• Qwen3 8B model and nomic-embed-text model