AI service assistant for machine builders: error codes, parts by photo, service reports

What an AI service assistant does for a machine builder

An AI service assistant for a machine builder answers the questions of service technicians and hotline staff from the company's own service knowledge: manuals, earlier service reports, error-code lists, spare-part catalogues and the history of each machine in the installed base. It turns an error code into the next repair steps, finds a spare part from a photo, drafts the service report from a technician's voice note and answers in the customer's language, with the source of every answer shown. For a service organisation of up to a few hundred people, a server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards carries it; four cards add capacity, a second model for languages or images, and room for other departments.

We wrote this for the head of service or aftersales at a machine or plant builder, who decides whether such an assistant pays off, and for the IT team that will connect it to the ticket system and the ERP. The tasks and the configurations come first; the technical notes follow after them.

Four tasks of a field service team

A machine builder's service team supports machines at customers' sites, often for many years after delivery. Four tasks take most of its time, and each needs a different part of the assistant.

Our summary of the design; models named in the sections below with the pages they come from.

The assistant finds sources and drafts text. The technician still diagnoses the machine and the service manager still releases a repair that touches a safety function. What changes is the time spent searching: in an earlier report, in the right revision of the manual, in the parts list of a machine built twelve years ago.

This article covers the builder's aftersales for machines at customers' sites. An assistant for the maintenance team inside one plant is part of our article on an AI server for manufacturing, and a chatbot for customers on a website is the subject of our guide to a server for a customer service chatbot.

Machine error codes: troubleshooting from the code to the next step

A machine control reports a code, the operator calls the hotline, and someone has to find what the code means on this machine, with this software version, and what helped the last time. A language model guessing from a code number gives poor answers. The better design splits the work.

1. Lookup. The model calls a tool that looks up the code in the error-code table for the machine type and software version, and fetches the machine's record from the installed-base data: build year, options, earlier tickets. vLLM's documentation makes --enable-auto-tool-choice mandatory for tool calls and pairs it with a parser per model family, openai for gpt-oss.
2. Retrieval. The search index returns the manual section for the code and earlier service reports with the same code or the same symptom. This is RAG, retrieval-augmented generation, as our article on RAG on company data explains.
3. Answer. The model writes the next steps, with the manual page and the report numbers it used, so the technician can open the source.

Old service reports are the most useful source and the least consistent one. They are short, full of abbreviations and sometimes wrong. A service engineer should decide which report types enter the index, mark reports whose fix turned out to be wrong, and own a fault database, the team's knowledge base that grows from confirmed cases. The embedding model that makes the reports searchable is small: Qwen3-Embedding-0.6B, Apache 2.0, names more than 100 languages, takes 32K tokens and has a 1.19 GB file, and a matching Qwen3-Reranker-0.6B exists. Our guide to embedding and reranker GPU requirements sizes them.

Spare parts by photo

A technician at a customer's site sees a worn part without a readable number, takes a photo and needs the part number. Two techniques help, and both need a test with your own parts.

• Image search. An image embedding model turns every picture into a vector, a list of numbers, so similar pictures lie close together. Google's SigLIP 2 (siglip2-so400m-patch14-384) is published under Apache 2.0 for "zero-shot image classification and image-text retrieval"; its model card shows about 1B parameters, and its file takes 4.54 GB in 32-bit. You index the part images from the catalogue, renderings from CAD and photos from earlier cases; the technician's photo returns the closest matches.
• Reading the type plate. A vision-language model reads text in a photo, such as a type plate or a label. Mistral Small 4 (Mistral-Small-4-119B-2603) accepts text and images under Apache 2.0, as its model card states.

The parts list of the machine in question narrows the result: a match is only useful if the part was built into that serial number. The output is a shortlist of candidates with their images, and the technician confirms the part. Renderings look different from a dirty part in a hall, so add photos from real service cases to the index as they come in. Images of documents, such as scanned parts lists, are a different task, covered in our guide to ColPali visual document retrieval.

Service reports from voice notes

After a job, the technician speaks a short note into a phone or tablet: what was found, what was replaced, what is still open. A speech-to-text model transcribes it, and the language model fills the report template from the transcript, with part numbers checked against the parts list. The technician reads the draft and sends it; the customer's sign-off follows your existing process.

OpenAI's Whisper repository puts its code and model weights under the MIT licence and gives about 10 GB of GPU memory for Whisper large and about 6 GB for turbo, so speech takes a fraction of one card. Our guide to Whisper GPU requirements covers throughput for larger volumes of audio.

Answers in the customer's language come from the same platform. The model card of gpt-oss-120b states no language list. The model card of EuroLLM-22B-Instruct-2512 lists 35 languages, among them all 24 official EU languages, under Apache 2.0, and Mistral Small 4's model card names 11 languages, German, French, Spanish, Italian, Portuguese and Dutch among them, and support for "dozens of languages". Test the languages of your largest markets with real service questions; our comparison of open LLMs for European languages has more detail.

We build servers with two or four RTX PRO 6000 Server Edition cards and set up the models and retrieval; tell us how many technicians and hotline staff would use the assistant and where your service data lives, and our team replies with a configuration.

Whose data it is: customers, GDPR and the Data Act

This is general information, not legal advice; your legal counsel or data-protection officer confirms what applies to your contracts. Service data comes from several owners. Manuals and parts lists are yours. Service reports and tickets contain the names of customer staff and details of the customer's production, so the GDPR applies to the personal data in them, and service contracts or non-disclosure agreements may limit how you reuse a customer's information for other customers. A model on your own server keeps documents, prompts and answers under your access rules, and the retrieval index can filter by customer, so a technician sees reports from other customers only where your contracts allow it.

Machine data adds a newer rule. The Commission's Data Act page states that Regulation (EU) 2023/2854 "entered into application on 12 September 2025" and gives users "the right to access the data produced through their utilisation of smart objects". Recital 25 of the Regulation says that where the manufacturer is a data holder, its basis for using non-personal data should be a contract with the user. Recital 31 adds that a data holder cannot, in principle, refuse an access request solely because the data is a trade secret, and should identify trade secrets before disclosure. If your assistant learns from operating data of connected machines at customers' sites, check with your legal counsel that your contracts cover that use.

If customers' own operators use the assistant directly, Article 50 of the AI Act has required since 2 August 2026 that people are told they are interacting with an AI system, unless that is obvious. Our article on the EU AI Act for companies using LLMs gives general information; national practice may differ.

Configurations by card count

The options follow the reference configurations on our AI servers page and NVIDIA's guide for certified systems: two processors, at least six cores per GPU and system memory of twice the GPU memory. Our users-per-card rule takes 90 per cent of the 95.6 GiB the driver reports, less 3 GiB per card, less the weights.

Reference configurations on our AI servers page; conversations by our users-per-card rule with gpt-oss-120b (60.8 GiB) and a 16-bit cache; users at an example value of 8 requests in flight per 100 technicians and hotline staff at the peak; power by our estimate with the cards at 600 W, processors and 15 per cent for the rest of the server.

Two cards. The entry server has two RTX PRO 6000 Server Edition, two AMD EPYC processors, 384 GB of DDR5 ECC and two NVMe drives. gpt-oss-120b, Apache 2.0, with 117B parameters of which 5.1B are active per token, takes 60.8 GiB and leaves 22.2 GiB of cache on the first card. With the 18 full-attention layers and 8 key/value heads of dimension 64 listed in its config.json, a conversation of 32K tokens needs 1.125 GiB of 16-bit cache by our arithmetic, and 19 fit at once. With an example value of 8 requests in flight per 100 users at the busiest moment, that covers about 235 technicians and hotline staff; your gateway log after a pilot replaces the example value. MIG, NVIDIA's way of splitting one card into isolated parts, divides the second card into four 24 GB instances: SigLIP 2 for image search, the embedding model and reranker, Whisper for voice notes, and one for tests.

Four cards. The inference node with 768 GB and four NVMe drives runs gpt-oss-120b as one copy split over two cards with tensor parallelism, which keeps 2 × 83.0 − 60.8 = 105.3 GiB for the cache: 93 conversations at 32K, about 1,160 users with the same example value. The third card takes a second model. EuroLLM-22B, 42.2 GiB in 16-bit, leaves 40.9 GiB of cache on one card for answers in the smaller EU languages. Mistral Small 4 in NVFP4 takes 70.8 GB, 65.9 GiB, and leaves 17.1 GiB on one card by our rule; its model card's own example runs it on two GPUs, and NVFP4 support on RTX PRO Blackwell depends on the engine release, so test it on the target card before you plan on it. The fourth card carries search and speech in MIG instances.

Two servers. If technicians at customers' sites depend on the assistant during a machine stoppage, plan two four-card servers that each carry the peak, so one can fail or be updated while the other answers, as our guide to LLM high availability with two GPU nodes describes.

The RTX PRO 6000 Max-Q is the workstation card, a 300 W blower card; in a rack it belongs only in a chassis whose maker lists it, and for continuous operation we recommend the Server Edition.

Notes for the IT team

1. Interfaces. vLLM serves an OpenAI-compatible API and has pooling endpoints for embeddings. The ticket system, the field-service app and the hotline desktop call one interface behind an LLM gateway with keys and a query log.
2. Tools, not exports. Error-code lookups, installed-base records and parts lists per serial number are tool calls into the systems that own them. The backend identifies the user and passes only the data that user may see; the model never sees credentials, as our guide to prompt injection and LLM security explains.
3. Access from the field. Technicians reach the assistant over your VPN from laptops and tablets. Where a site has no connection, export the manual and the parts list of that machine to the device before the visit; the assistant itself runs on the server.
4. Index per customer. Every chunk in the index carries the customer and the access list of its source, and retrieval filters before it ranks. A vector database with metadata filters handles this.
5. Network and versions. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere, while routes such as /invocations and /metrics answer without the key (vLLM 0.32.0 security guide), so keep the engines on an internal network behind the gateway. Run vLLM 0.31.0 or later, the NVIDIA Container Toolkit 1.17.8 or later and a driver release that NVIDIA lists as fixed in its bulletin of 30 September 2026.
6. Software on delivery. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime; our guide to RTX PRO Blackwell software support lists the engine versions for these cards. The field-service app and its user interface stay with your team or your software provider.

Power and room

By our estimate, the two-card server draws about 1.8 kW at the wall with two EPYC 9135 processors and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent. The four-card server draws about 3.4 kW with the cards at 600 W and needs 32 A or three-phase, or about 2.7 kW with the cards capped at 450 W. Both belong in a server room or a data centre, as our guide to GPU server power on 16 A, 32 A and three-phase feeds explains. Without a server room of your own, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU.

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 the servers to order with two, four or eight RTX PRO 6000 Server Edition cards, assembled and tested before delivery, delivered within 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 models, retrieval with access per customer and the tool connections to your service systems. Manuals written with AI feed the same index, as our guide to machine manuals with AI under the Machinery Regulation shows; describe one product line and its service data, and we propose a pilot and the server for it.

FAQ