AI server for fragrance and flavour houses: formula search, IFRA and customer documents

What an AI server does in a fragrance or flavour house

An AI server for a fragrance or flavour house is a machine in your own building that runs a language model on your formula library, your panel notes and your regulatory files: it finds earlier formulas that match a brief, shows where a restricted material was replaced before, fills customer questionnaires from your approved data and drafts IFRA certificates and cover letters for the regulatory team to check and sign. Because the model runs on your own hardware, no formula goes to an outside AI service. A server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards serves the whole company of a small or mid-sized house.

This guide is written for the owner, the head of R&D or the regulatory manager of a fragrance or flavour company that has no large IT department. It starts from the work, explains each technical term the first time it appears, and keeps the hardware in one section with a table. The regulatory statements are general information on EU rules and IFRA's published guidance as of October 2026; your regulatory adviser confirms what applies to your products.

Four jobs, in plain words

For a fragrance or flavour house the formulas are the core of the business. Everything the AI does sits around them, and none of it changes a formula by itself.

Our summary of the tasks; IFRA's page on certificates (read 25 October 2026) for who issues the certificate.

A language model is software that reads and writes text. On its own it knows nothing about your formulas; it answers from documents and data you give it at the moment of the question. That method, called retrieval, is how the server works on your library without the model being trained on it, as our guide to RAG on company data explains. The server therefore needs a clean source: the formula database, the raw-material files and the documents customers received.

Searching the formula library

A house with decades of work has thousands of formulas and many more trials, each with a brief, panel notes, stability results and the reason it was accepted or dropped. Most of that knowledge sits in the formula software and in documents nobody reads twice. Two kinds of search help here, and they need different computers.

Similar by composition. Finding the formulas closest to a given one is a comparison of numbers: which materials, in which shares. That runs on an ordinary processor inside the formula software or a small script, with no GPU.

Similar by meaning. A brief says "fresh, green, for a shower gel in Asia", a panel note says "metallic in the dry-down", a customer complaint says "too sweet after six weeks". Matching such texts to earlier work needs an embedding model, a small model that turns texts into numbers so that similar meanings land close together, and a language model that reads the hits and writes a summary with links to the formulas. This is GPU work, and it is where staff save time: a question in plain words returns the five trials that matter and why they were stopped.

The same search finds substitutes. When a material is restricted or becomes hard to buy, the assistant lists the formulas that contain it and the earlier formulas where it was replaced, with the replacement and the panel result. The perfumer or flavourist decides; the assistant only shortens the search. Our guide to formulation R&D with AI on your own lab data covers property prediction and planning lab batches, which is a separate step.

IFRA certificates and the allergen data customers ask for

IFRA, the International Fragrance Association, publishes the IFRA Standards, which restrict or prohibit materials for given uses. Its page on certificates is clear on who writes the Certificate of Conformity: it is "a document only prepared by companies that formulate fragrance mixtures", and "IFRA does not create the certificates". Raw-material suppliers generally do not issue one; they may tell their clients that a material conforms. The certificate states that a mixture complies with the Standards for the intended use, and IFRA adds that it does not replace a safety assessment. IFRA provides a template and lists independent service providers, none of which issues certificates on IFRA's behalf.

The Standards change on a cycle. IFRA normally issues an amendment every three years, with off-cycle amendments when an important change requires it. The 51st Amendment was notified on 30 June 2023, and the consultation on the 52nd closed on 12 June 2026. IFRA's FAQ gives implementation periods counted from the notification: for prohibitions 2 months for new creations and 13 months for existing ones, for restrictions and specifications 9 and 28 months. Each amendment therefore means checking the library and reissuing certificates for every customer and every use.

Where the AI fits. The limit check is arithmetic: the concentration of each restricted material in the mixture, compared with the limit for each use. That belongs in the formula or compliance software, which already does it. The assistant takes the result and drafts the certificate in your template, the cover letter and the list of formulas an amendment touches; it never computes a limit itself. A model that rewrites a number creates an error.

The cosmetics allergen change. Commission Regulation (EU) 2023/1545 amends Annex III of the cosmetics regulation and adds fragrance allergens that cosmetic labels must name above 0.001 % in leave-on and 0.01 % in rinse-off products. Products that do not meet the new rules may be placed on the market until 31 July 2026 and made available until 31 July 2028. By our reading, a cosmetics maker can only decide its labels if its fragrance supplier tells it how much of each listed allergen the fragrance contains, so the wave of requests lands on the fragrance house. The assistant answers each request from the allergen data the software has computed and keeps the answers consistent across customers; the cosmetics side of this change is covered in our guide to an AI server for cosmetics manufacturers.

Questionnaires. Customers send their own forms for regulatory status, allergens, origin, certifications and, more and more, for how you use AI. The assistant fills them from approved data, names the source of every answer and marks the questions it cannot answer. Our guide to AI in customer audits and supplier questionnaires covers what to say about your own AI use.

We build RTX PRO 6000 Max-Q workstations and Server Edition servers, and Private AI/ML connects the assistant to your formula data; send us your number of formulas and users and the documents you answer most often, and we reply with a proposal.

The flavour side: Regulation (EC) No 1334/2008

Flavours sold to food makers carry their own label rules. Regulation (EC) No 1334/2008 defines flavourings as products not intended to be consumed as such, which are added to food to impart or modify odour or taste. Under Article 14, flavourings not sold to the final consumer may be marketed only with the labelling of Articles 15 and 16, easily visible, clearly legible and indelible, in a language the buyer easily understands. Article 15 lists what the packaging or containers show, among them the word "flavouring" or a more specific name, the statement "for food" or a more specific intended use, special storage conditions where necessary, a mark identifying the batch or lot, and the maximum quantity of each component subject to a quantitative limit. Article 16 allows "natural" together with a named source only if the flavouring component comes at least 95 % by weight from that source.

The assistant helps in two ways. It checks a draft specification or label against a list built from these articles and reports what is missing; the regulatory team decides. And it answers the food customer's questions about a flavour from the same approved data, including allergens and origin, as our guide to an AI server for food and beverage producers describes from the customer's side. Safety data sheets for compounds are a task of their own, set out in our guide to safety data sheets with local AI.

Why confidential formulas stay in the building

Many people in a house ask whether they may paste a formula or a customer's brief into a public chatbot. That is a question for the management and its legal counsel, and our article on trade secrets, NDAs and AI tools explains what the EU's Trade Secrets Directive asks of a company that wants its secrets protected and what business AI subscriptions state about inputs. A server in your own building removes the question for the daily work: the model, the library and every draft stay on hardware you control.

Inside the company, access rights do the rest. Perfumers and flavourists see full formulas; regulatory sees the compliance data; sales sees the approved documents for its customers. The assistant answers each person only from what that person may open, and a query log shows who asked what.

Which machine, for how many people, and what it needs in the building

Only part of the staff uses the assistant at the same moment. With the example values of our guide to an AI server for small companies, 100 employees send about four requests at the same moment in the busiest hour. One RTX PRO 6000 card, Max-Q or Server Edition, holds 96 GB, of which the driver reports 95.6 GiB. Our users-per-card rule takes 90 per cent of that, less 3 GiB per card, less the model: gpt-oss-120b, an open model under Apache 2.0 whose weights take 60.8 GiB, leaves 22.2 GiB, enough for 19 conversations of 32,768 tokens at once (a token is a word or part of a word).

Usable memory by our users-per-card rule (0.9 × 95.6 GiB − 3 GiB − weights); gpt-oss-120b weights about 65 GB = 60.8 GiB (Hugging Face files page, read 25 October 2026); 1.125 GiB of 16-bit cache per 32K conversation; power by our estimate (cards, processor or processors and 15 per cent for the rest), confirmed by a measurement on the delivered system.

Two RTX PRO 6000 Max-Q in a workstation. The NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition is a 300 W card with a blower fan, and NVIDIA allows up to four in a workstation. Two of them run the assistant on one card and search, embedding and translation models on the other, at about 1.1 kW at the wall on an ordinary 230 V socket. It suits a house that starts with the lab or the regulatory team and has a side room with a door, since the machine is too loud for an office. Our page on buying a two-card Max-Q workstation lists the parts.

Two RTX PRO 6000 Server Edition in a rack server. The entry configuration 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. The Server Edition has no fan of its own and is cooled by the server, which is built for continuous operation and remote management. This is the machine we would size first for a house that wants one assistant for R&D, regulatory and sales. If one card fails, gpt-oss-120b fits on the other until the replacement arrives, by our estimate.

Four RTX PRO 6000 Server Edition. Four cards suit a house that wants a larger model, the assistant on a pair of cards for more conversations, and separate cards for search, translation into the languages of its customers and document reading. At about 3.4 kW, or about 2.7 kW with the cards capped at 450 W, it needs a 32 A or three-phase feed, as our guide to GPU server power on 16 A, 32 A and three-phase feeds works through.

Practically all the power becomes heat, so the room needs cooling that runs whenever the machine runs. 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.

Running it without an IT team

A small house needs one person who owns the assistant: someone who decides which documents and data go in, who may use it, and what to do with feedback. Installation, updates and repairs can sit with a supplier. Our guide to an AI server without an IT department splits the tasks between that person, an existing IT service provider and a support provider. On request the machine arrives with the operating system, the NVIDIA driver, CUDA and a container runtime installed; models, search, user accounts and the connection to your formula software are our Private AI/ML service, with our engineering partner Vixen.UNO.

Start with one job and measure it. A good first pilot is the questionnaires of the last six months: run them through the assistant, compare its answers with the ones your team sent, and count the corrections. Our guide to what a private AI pilot should measure sets out the method.

Notes for the IT team

1. Data access. The formula software stays the system of record. The assistant reads it through an export or a read-only database account and never writes back; drafts go to a review queue.
2. Serving. Run the models with a current vLLM release (0.32.0 of 18 October 2026 at the time of writing) behind a gateway with keys and a query log. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere, so keep the engine on an internal network.
3. Permissions. Pass the user's identity to the search, and filter documents by the user's rights before the model sees them.
4. Versions. Pin model revisions in safetensors format, keep the NVIDIA driver at a release that NVIDIA's bulletin of 30 September 2026 lists as fixed, and record the model version in every draft.

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 sets up the models, the search over your library and the connection to your formula software; tell us which job you would start with, and we propose a pilot.

FAQ