Food & Beverage /Retrieval (RAG) live field guide · 8 min

Spec & Allergen Concierge

Answers spec, allergen, and COA questions from your approved documents only, quoting the source document and section on every line — and refusing to answer when the approved pack is incomplete, so customer service and sales stop waiting on one experienced person.

Open the live demo

theater/demos/food-bev_spec-allergen-concierge.html · sandbox · read-only

Open

FIG. 1

The live demo, running on fabricated data. Open it to step through the full flow — every output is shown for a person to approve before anything happens.

How it would work

Reads only your approved specs, allergen matrix and COAs, drafts a cited answer to a buyer's question, and surfaces the source document and section for a person to approve before it goes back to the customer.

Input 01

The approved document pack

Product specifications, the allergen matrix, certificates of analysis, label files and the SOPs you have signed off as current — the only documents the tool is allowed to answer from.

Agent 02

Searches, cites, or refuses

Searches the approved pack for the buyer's question, drafts an answer with the exact document, version and section quoted on every claim, and says it cannot answer when the pack does not fully cover the question.

Output 03

A cited draft, awaiting sign-off

A draft answer with every source shown and the grounding strength flagged, held for a customer-service officer or QA to read, correct and approve before it is sent to the buyer.

Where it works well

It finds the right clause in the right version every time, and never sends the answer on its own.

Done by hand it is a forward-the-email-and-wait loop that lands on the one QA lead who knows where everything lives.
Best where the same kinds of questions recur across a large range: allergen declarations, shelf life, COA requests, 'does this contain…'.
At dozens of buyer queries a week, the recaptured hours go back into QA judgement and customer relationships, not document archaeology.

The slow, invisible cost of a spec query is the hunt — tracing one allergen or shelf-life claim across a spec sheet, an allergen matrix and a COA, in whichever version is current, usually inside one experienced person's head.

Where it works badly

It is confidently wrong when the approved pack is stale — and a cited answer looks more trustworthy than the document behind it.

Weak on anything the pack does not contain — a dietary or certification claim (low-FODMAP, organic) it has no document for. It should refuse, not infer from a related sheet.
If your specs and allergen statements are not version-controlled, the tool cannot tell a current document from a superseded one, and neither can the reader trusting the citation.

The honest test

If you cannot say, right now, which version of the spec and allergen matrix this answer was drawn from — this tool makes your wrong answer faster and more citable, not safer.

Point it at last quarter's allergen matrix, or a spec that was reissued after a recipe change, and it drafts a clean, sourced answer quoting a version that no longer holds. The citation makes the wrong answer look right. That is the trap.

What it doesn't do — and shouldn't

It drafts a cited answer. A person approves what goes to the customer. That boundary is deliberate.

WHAT IT DOES

Surfaces the document name, version and section it answered from

Shows the grounding strength and how many sources it matched

Refuses and routes to QA when the approved pack does not cover the question

WHAT IT WON’T

Send the answer to the buyer, auditor or sales contact

Make an allergen or dietary claim the documents do not support

Decide whether a spec or COA is current and correct

An allergen answer is a safety statement. Under the Food Standards Code, allergen declarations now follow Plain English Allergen Labelling, and a wrong "contains / does not contain" answer can drive an incorrect label or a buyer's purchasing decision — the kind of error that ends in a FSANZ food recall. The accountable person stays on the decision because the consequence lands on them, not the tool.

What your data has to look like

Approved specs and allergen statements that are current, version-stamped, and searchable as text.

Typical readiness

across orgs we see, before the first job

Product specs as searchable text

Needs shaping

A version-controlled allergen matrix

Usual weak point

An owned "approved pack" boundary

Needs shaping

Supplier COAs filed against the right product and batch

Usual weak point

SOPs and label files kept in step with the specs

Needs shaping

The real first job

The weak point is almost always currency and version control — specs reissued after a recipe change, an allergen matrix updated in one place and not another, COAs filed loosely. Getting the approved pack into clean, version-stamped, searchable form is usually the real first job — larger and more valuable than the retrieval layer on top. Once the pack is trustworthy, every cited answer after that is fast and right by default.

Right fit if…

You field daily spec, allergen and COA queries from buyers, foodservice and export customers

A large product range where the answers live across many documents and one person knows them all

Your specs and allergen matrix are signed off and version-controlled today

The same kinds of questions recur — allergens, shelf life, COA requests

Walk away if…

Your specs live as scanned-image PDFs nobody has digitised

Your allergen matrix is updated ad hoc with no version anyone owns

Most queries are genuinely novel and need a QA judgement call each time

You want a tool that signs off the allergen answer so a person doesn't have to

Open questions

The worried-buyer questions, answered straight

It can draft a wrong allergen answer — which is exactly why nothing is sent on its say-so. It answers only from your approved allergen matrix and specs, quotes the document, version and section behind every claim, and refuses outright when the pack does not cover the question rather than inferring. A customer-service officer or QA reads the cited answer and approves it before it leaves. Under the Food Standards Code an allergen statement is a safety claim, so the person stays on the decision.

It works from documents it can read as text and version-stamp. If your specs are scanned-image PDFs, or your allergen matrix is updated in one copy and not another, the tool cannot tell a current document from a superseded one — and a citation to a stale version looks just as trustworthy as a correct one. Getting the approved pack into clean, current, version-controlled form is usually the first piece of work, and the piece that pays off across every answer after.

No. It removes the document hunt — the forward-the-email-and-wait loop that lands on the one person who knows where everything is — so QA spends their time on the calls that genuinely need judgement and on keeping the approved pack current. The answer is still signed off by a person before it goes out. The capacity it frees goes back into QA judgement and customer relationships, not headcount.

Current to your latest signed-off version. Specs get reissued after recipe or supplier changes, and allergen declarations had to meet the Plain English Allergen Labelling requirements under the Food Standards Code, whose transition period has now closed — so an out-of-date matrix is both wrong and non-compliant. The honest test: do you know, today, which version of the spec and allergen matrix an answer was drawn from?

Your specs, allergen data and supplier COAs are commercially sensitive. Any deployment runs against your own environment and document store, not a shared pool — retrieval is bounded to the approved pack you control, and we scope where the documents sit and who can see them as part of the build. The demo here runs entirely on fabricated data; NorthLink AI and its Anzac Biscuit spec are not real.

No. It answers from the documents you give it and refuses where they fall short, but it does not certify that your specs or labels meet the Food Standards Code. Plain English Allergen Labelling, precautionary “may be present” statements and your food recall obligations are decisions a person owns against the current Code. The tool gets the cited answer most of the way; the accountable person closes the gap.

What it takes to build

3–4 weeks · 4 phases

Reused from template~70%

Bespoke to this skin~30%

stack · Microsoft Copilot · private RAG · review UI

What it would cost

Fixed scope, fixed price, fixed dates.

Bite-sized first piece

One contained change, low risk

Pilot build

Most builds land here

Embedded support

Scale on proof

Considering this for your org?

The honest place to start is a bite-sized first piece — one contained change, low risk. Tell us where it hurts; we'll play it back, scope it, and show you what's possible.

Book a call How we work