AI server for product development: searching lab reports and analysing test data in-house

What an AI server does for an R&D department

An AI server for product development runs a language model over the R&D department's own test reports, lab notebooks, measurement files and earlier experiments: engineers ask which tests were done on a material or how two test series compare, and it answers with the documents it used. It also analyses test data by writing and running a short program on the server and drafts report sections for a person to check, and nothing leaves the building, which matters for results not yet protected by a patent.

Typical questions are what came out of an endurance run three years ago, or what a specification or a paper in the literature collection says about a test method.

We wrote this for the head of R&D, development or testing at a manufacturer with a department of 5 to 20 people: mechanical and electrical engineering, plastics, chemicals, medical and consumer products, and their suppliers. The tasks come first, then the questions of confidentiality, then the hardware in one section with a table, and notes for whoever runs the IT. It differs from our guides to a GPU server for chemistry and materials research, which is about simulation codes, to RAPIDS on RTX PRO workstations, which is about data science libraries, and to formulation R&D with AI on your own lab data, which is about predicting recipes. Here the subject is the everyday knowledge work of a development department.

Three terms come up throughout. A language model is the AI that reads text and writes answers; gpt-oss-120b from OpenAI, under the Apache 2.0 licence, is one that runs on your own hardware. Search with source, which engineers call RAG (retrieval-augmented generation), means the system first finds the relevant passages in your documents and the model answers only from them, naming the document. A GPU, or graphics card, is the part of the machine that runs the model; its memory decides what fits.

Four tasks in plain words

Our summary of typical uses; the example questions are illustrations.

Each task saves a different kind of time. Finding earlier work avoids repeating a test somebody did before. Comparing and summarising cuts reading. Data analysis replaces the hour of spreadsheet work before a meeting, and a draft saves typing while the engineer still decides what the report says.

Search lab reports with AI: test reports and lab notebooks

A development department collects its knowledge in many formats: PDF test reports, scanned protocols with handwritten notes, spreadsheets with measurement series, electronic lab notebooks, emails with suppliers and slides from design reviews. The search finds what it can read, so the first project step is to decide which sources go in and to bring them into a readable form. Our guide to preparing company documents for an AI assistant covers that step, and our guide to PDF parsing for RAG covers the tools that turn reports with tables and charts into text.

Three rules make the answers useful to engineers.

• Every answer names its source. The answer shows the report number, date and page of every passage it used, so the engineer opens the original before relying on a value. A model can summarise a passage wrongly; the source makes that visible.
• Metadata does the filtering. Test reports carry a product, a material, a test standard, a date and a status. Stored as fields, they let the search answer "only released reports for product line X since 2022", and our guide to RAG on company data explains why search quality, more than the model, decides the answer.
• Access follows projects. Not everyone in the department may see every customer project. The access check happens when documents are retrieved, not by an instruction in the prompt, so a person only gets answers from reports they may open.

Before the system goes live, write down 50 to 100 real questions from the department with the right answer and the report it comes from. That list shows whether the search works and whether a later change made it better, as our RAG guide recommends.

Analysing test data with code on the server

A figure computed inside the text of an answer cannot be checked; a figure computed by a short program can. For data analysis, the model writes Python code, the server runs it on the uploaded measurement file, and the engineer sees the code, the result table and the chart. The calculation can be repeated and checked, which matters when a figure goes into a report.

Open WebUI, a common chat front end for models on your own server, describes its options for this in its documentation (read 25 October 2026). Pyodide runs Python in the user's browser with a small, fixed set of packages, and "Nothing else can be installed at runtime." The Jupyter backend runs on the server but is a "Shared environment: all users share the same Python runtime and filesystem" and is "Not sandboxed by default", so the documentation recommends it only for single users or trusted set-ups. Open Terminal, the backend the documentation recommends, runs code in a Docker container on the server with any package installed. Files attached to a message are placed in /mnt/uploads/ before the code runs.

For a development department that means: measurement files stay on the server, the code runs in a container that sees only those files, and the person who asked checks the result. Keep the container separate from the document store and the rest of the network, and keep Open WebUI current, as our guide to Open WebUI security describes. gpt-oss-120b lists Python code execution among its native capabilities on its model card, and a second model that reads images, such as Qwen3.8-27B under Apache 2.0, which its model card describes as a native vision-language model, can read charts and photos from older reports.

We supply workstations with RTX PRO 6000 Max-Q cards and servers with Server Edition cards, and our Private AI/ML service sets up the search and the code sandbox; list the documents and data your department works with, and our team replies with a proposal.

Standards and published literature

Engineers want the assistant to know the test standards they work to. Purchased standards come with licence terms, and these often exclude AI use. DIN Media's support page states that "It is currently not permitted to use standards in artificial intelligence (AI)" under its terms and offers a separate AI licence for internal, isolated AI environments; ISO's current terms for its Online Browsing Platform exclude embedding, prompting or querying ISO content with AI unless a separate licence allows it. Our guide to standards in your company AI sets out the terms and how licensed standards are kept in a collection of their own.

Your own test procedures, which cite a standard and describe how your lab applies it, are your documents and can go into the search. For published papers and datasheets, the licence or subscription terms of each source apply; your legal counsel confirms what a licence allows.

Confidential until the patent is filed

Development results are worth most before anyone else knows them. Two sets of rules make that concrete.

Patents. Article 54(1) of the European Patent Convention treats an invention as new "if it does not form part of the state of the art", and Article 54(2) holds the state of the art to comprise "everything made available to the public" in writing, orally, by use or in any other way "before the date of filing", as our guide to an AI server for patent attorneys quotes. Whether entering an unfiled invention into an external AI service makes it available to the public depends on the service's terms and is a question for your patent attorney. On a server of your own, the question does not arise.

Trade secrets. Under Article 2(1) of Directive (EU) 2016/943, information is a trade secret only if it is secret, has commercial value because it is secret and "has been subject to reasonable steps under the circumstances, by the person lawfully in control of the information" to keep it secret. Test results, process windows and failed approaches are typical examples. Access by project, a query log and a server off the internet are steps you can document; our article on trade secrets, NDAs and AI tools covers the business side, including customer data under an NDA.

This is general information as of October 2026 and no legal advice; your patent attorney and legal counsel confirm what applies to your inventions and contracts.

Which machine for a team of 5 to 20, and what it needs in the building

A development department of 5 to 20 people rarely sends more than a few questions at the same moment. With the example values of our guide to an AI server for a company of 10 to 100 people, 40 per cent of staff active in the busiest hour, 6 requests an hour, 30 seconds each and a peak factor of 2, 20 people come to about one request at a time by our arithmetic. Long reports are the larger load, since each session keeps its text in GPU memory.

Usable memory by our users-per-card rule (0.9 × 95.6 GiB − 3 GiB − weights): gpt-oss-120b 60.8 GiB of weights, 1.125 GiB of 16-bit cache per 32K session; reference configurations on our AI servers page; power by our estimate (cards, processors, plus 15 per cent), to be confirmed by a measurement on the delivered system.

A workstation for the department. The NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition has 96 GB of memory at 1,792 GB/s and a 300 W blower cooler; NVIDIA allows up to four in one workstation. Two of them draw about 1.1 kW at the wall by our estimate, which an ordinary 230 V socket carries. By our users-per-card rule, which takes 90 per cent of the 95.6 GiB the driver reports, less 3 GiB per card and less the weights, gpt-oss-120b leaves 22.2 GiB on the first card: 19 sessions of 32,768 tokens or 9 of 65,536 tokens at the same moment. On the second card, Qwen3.8-27B, whose FP8 weights take 28.75 GiB (30.9 GB), reads charts and scanned pages and leaves room for the search models; MIG can split that card into two isolated parts of 48 GB or four of 24 GB, for example one for the code sandbox's GPU jobs. This is the machine for a department that wants its own system in its own building.

A server for the company. The entry server on our AI servers page has two NVIDIA RTX PRO 6000 Blackwell Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives, and draws about 1.8 kW by our estimate on one 230 V 16 A feed, which carries 2.94 kW at 80 per cent. It belongs in a server room, where the IT team looks after it, and other departments can use it too, with access rights keeping development documents apart. The Max-Q is the workstation card: in a rack it belongs only in a chassis whose maker lists it, and for continuous operation we recommend the Server Edition.

Four cards. When quality, sales and service also want their own assistants, the inference node with four cards gives each a card. Simulation work belongs on its own machine, as our guides for chemistry and for automotive R&D describe.

Notes for the IT team

1. Search models. Embedding models and rerankers are small next to the language model; Qwen3-Embedding-0.6B takes about 1.2 GB in BF16, as our guide to embedding and reranker servers for RAG lists, so they share a card with other models.
2. Access filter. Map project and customer permissions from your directory to document collections, and filter at retrieval time.
3. Code sandbox. Run Open Terminal in its own container without access to the document store or the internet, give it to named R&D users only, and log what runs.
4. Serving. Run the models with vLLM behind an LLM gateway with keys and a query log; vLLM's --api-key protects only some paths (vLLM 0.32.0 security guide), so keep the engines on an internal network.
5. Software on delivery. The machine arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime; the RTX PRO Blackwell cards are compute capability 12.0, as our guide to RTX PRO Blackwell software support explains.
6. Backups. The document index can be rebuilt from the sources; the sources themselves, the lab notebooks and the query log belong in the backup plan.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build workstations with two RTX PRO 6000 Max-Q and servers with two or four RTX PRO 6000 Server Edition cards to order, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the search over your reports, the access rights and the code sandbox; tell us which document sources you would start with, and we propose a first step.

FAQ