Can NVIDIA GB10 Analyse Thousands of Company Documents Without Sending Them to the Cloud?

Yes. Indexing and questioning thousands of documents is well within what a single NVIDIA GB10 system can do, and every step can run on the box itself, so nothing goes to a cloud service. The usual method is retrieval-augmented generation (RAG): the system extracts text from your PDFs and files, turns it into searchable vectors, and passes the most relevant passages to a language model when someone asks a question. The parts that take the most care are not raw speed but extracting text from difficult PDFs, scanned pages and tables, keeping document permissions intact, and making sure no component quietly calls an external service. The ASUS Ascent GX10 starts at S$8,100 and the NVIDIA DGX Spark is S$8,800 at SourceIT (listed 28 Oct 2026).

What does a local document pipeline look like?

What has NVIDIA published for DGX Spark?

NVIDIA's DGX Spark playbooks include several document-related recipes. "Build and Deploy a Multi-Agent Chatbot" lets users upload documents for retrieval, using GPT-OSS-120B as a supervisor with a coding model and the Qwen3-Embedding-4B embedding model; NVIDIA notes it uses about 120 GB of the 128 GB in its default configuration. There are also playbooks for an agentic RAG application in AI Workbench, for turning text into a knowledge graph, and for video search and summarisation. One caution: the AI Workbench RAG example can be configured to use NVIDIA-hosted endpoints, so check every component is pointed at local models if data must stay in the building.

NVIDIA's larger RAG Blueprint, its multimodal PDF extraction stack, is not an official DGX Spark playbook. One published attempt, by Classmethod in August 2026, got it running but found several container images were built only for x86 and had to be rebuilt for Arm, and the OCR component would not start. The working stack used about 22 GiB, handled 16 simultaneous connections and was estimated at about 10 comfortable users. Expect some engineering if you want that full stack.

How long does indexing thousands of documents take?

NVIDIA has not published embedding throughput for DGX Spark, so here is the arithmetic you can check against your own test. Take 5,000 documents averaging 20 pages, at roughly 500 tokens per page: about 50 million tokens. If your embedding model processes 5,000 tokens per second, that is under three hours; at 2,000 tokens per second, about seven. For comparison, DGX Spark processes prompts for an 8B language model at several thousand tokens per second in published tests, and embedding models are typically smaller. In practice, PDF extraction and OCR on scanned pages are often slower than embedding. Index once, then update incrementally as files change.

Storage is not a constraint at this scale. The vector index for millions of passages is measured in gigabytes, small next to the 1 to 4 TB of NVMe in GB10 systems. Our 1TB vs 2TB vs 4TB guide explains where storage does matter.

The hard part of document AI is not reading fast. It is reading badly scanned PDFs correctly and only showing people what they are allowed to see.

What about privacy and PDPA?

A local pipeline removes the biggest question in cloud document AI: where your documents are processed and stored. To make that real:

• Block outbound traffic from the AI box except for approved updates, so no component can call an external API unnoticed.
• Use local models only for extraction, embedding and generation, and check default configurations for hosted endpoints.
• Carry permissions into retrieval, so a search only returns passages from documents the user could already open.
• Decide what not to index, such as personal data that has no business purpose in the assistant.
• Use the drive encryption. NVIDIA documents the DGX Spark's NVMe drive as self-encrypting; check the equivalent on other GB10 brands.
• Log and retain queries according to a policy agreed with whoever handles PDPA in your organisation.

The wider risks are covered in Your AI Assistant Might Be Your Biggest Security Risk and Where Does Company Data Go When Using AI?

Which model should answer the questions?

For document Q&A, the language model mainly needs to read the retrieved passages carefully and quote them accurately. A 20B to 30B model is often enough and much faster than a 120B one, which matters because RAG prompts are long. On DGX Spark, llama.cpp's benchmark shows GPT-OSS-120B processing prompts at 2,444 tokens per second at a 2,048-token prompt and generating at 58.7 tokens per second; at a 32K-token context those fall to 1,567 and 42.8. Use the larger model for analysis across many documents, and the smaller one for everyday questions.

How to start

1. Pick one collection, such as policies, contracts or technical manuals, with clear ownership.
2. Test extraction on your worst 20 files first: scans, tables, forms.
3. Choose an embedding model and measure indexing speed on a sample.
4. Build retrieval with permissions from day one.
5. Check answers against the source and measure how often they are wrong.
6. Expand to more collections once quality is acceptable.

Turning this into a company assistant is covered in Can NVIDIA GB10 Run an Entire Company's Private AI Assistant? and Is NVIDIA GB10 Powerful Enough to Build Your Own Corporate ChatGPT?

Buying in Singapore

SourceIT supplies the ASUS Ascent GX10 (1 TB S$8,100, 2 TB S$8,300, 4 TB S$10,635), the NVIDIA DGX Spark (S$8,800), the HP ZGX Nano G1n (S$11,499) and other GB10 systems with local warranty and a GST invoice, and can commission a unit with your chosen models. Browse the NVIDIA DGX Spark collection. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Can a GB10 analyse company documents without the cloud? Yes. Extraction, embedding, the vector database and the language model can all run locally on a GB10 system, so documents and questions stay on your network.

How long does it take to index thousands of documents? NVIDIA has not published embedding throughput for DGX Spark. As arithmetic, 50 million tokens at 5,000 tokens per second takes under three hours; PDF extraction and OCR are often the slower steps.

Does NVIDIA provide a document AI recipe for DGX Spark? Yes. Its playbooks include a multi-agent chatbot with document upload and an agentic RAG application. The full NVIDIA RAG Blueprint is not an official DGX Spark playbook.

How do I keep confidential documents confidential? Carry document permissions into retrieval, block outbound traffic from the AI box, use only local models, and log queries under an agreed retention policy.

How much storage does a document index need? Usually a few gigabytes even for millions of passages, which is small next to the 1 to 4 TB in GB10 systems.