Spot a new kind of complaint before it becomes a fire
Clusters incoming reviews and tickets by meaning and alerts you when a genuinely new complaint theme starts to accumulate.
- How it works
- Every incoming review and ticket is embedded with BGE-M3. New items that sit far from all existing complaint clusters get grouped together, and the small model writes a one-line label for each new cluster ('customers say the new lids crack in transit'). When a genuinely novel theme accumulates a handful of instances you get an alert, rather than discovering it in next month's report. It catches the things a fixed tag list has no label for — that list's blind spot by construction.
- Data you need
- A steady inbound stream of reviews and tickets in one place — a helpdesk API or a scheduled export — plus a few months of history to establish what normal complaint space looks like. In practice this means stitching sources together (helpdesk, Trustpilot, marketplace reviews), which is one-off integration work, and some marketplaces make review export awkward. Below roughly a few hundred messages a month, a person reading the inbox already does this job better.
- What to expect
- Expect false alarms, especially early, until the cluster-size threshold is tuned to your volume. Embedding distance measures 'worded differently', which is not the same as 'a different problem' — the same issue described in two languages can look like two clusters until the threshold is tuned; BGE-M3 is multilingual, which helps but is not perfect. This narrows the window on emerging issues; it does not guarantee catching them. Getting all feedback channels into one stream is the real setup cost for most shops.
- Where people stay involved
- Every alert is a suggestion, not a verdict — a person reads the clustered messages and decides whether it is a real emerging issue, one confused customer, or a phrasing artefact. Alerts should reach someone empowered to actually check the warehouse or the product batch.
Which model, and what it costs to run
Qwen3-8B
This job runs in bulk rather than to a waiting person, so size is not the constraint — we take the strongest independent score on sticking to the document that we may serve freely and that fits on a single card.
- Licence
- Apache-2.0
- Weights at 4-bit
- 5 GB
- Context
- 32K tokens
- Publisher
- Alibaba (Qwen Team)
What the hardware costs
One 48 GB card holds it
- Rent in the EU
- $1.60/hrScaleway, Paris (PAR2)
- Buy the card
- $7,569new, one-off
- Or rent it by the token
- $0.04 / $0.04per M in / out · DeepInfra
Hardware only, third-party prices from 2026-07. The figure excludes the KV cache, which grows with context length and how many people use it at once — sized properly in a conversation, not guessed here. Renting by the token is cheaper up front; why our customers still self-host is below.
Sticking to the document
Measured on public documents, by a model acting as judge. Read it beside the answer rate: the lowest hallucination rates on this board belong to models that simply decline more often.
Independent measurement · Vectara · board updated May 11, 2026
The API is cheaper per token. Here is why our customers don't use it.
We will not pretend otherwise: renting a model by the token from a serverless API costs less per million tokens than a card we run for you. We show that price on every use-case page. What it does not include is the part a shop with a customer database actually pays for.
- 01
Your data never leaves hardware you can point at
A serverless "we don't retain your data" is a clause in a contract. Running the model on a card in Amsterdam is a fact of architecture: your catalogue, tickets and customer records are never sent to a third party at all. For a GDPR audit, that is the difference between a promise and a floor plan.
- 02
The price cannot move without your say-so
A serverless rate card is somebody else's lever. The provider can raise the price, retire the model, or change the terms, and your cost moves with it. The same model on the same card costs the same next year — you own the number.
- 03
The model cannot be taken away
Hosted APIs deprecate models on their own schedule; the one you built on can be gone in a quarter. An open-weight model on your own hardware runs for as long as you keep the lights on. No vendor can end-of-life it out from under you.
And the price gap closes with volume: past a card you keep busy — very roughly four billion tokens a month — owning is cheaper outright, even before the three reasons above.
Getting a case like this one from a conversation to production takes about two months, and you can stop at the end of any phase.
Also in voice of customer
All 15 →Turn a quarter's worth of product reviews into a short themes report
Turns exported product reviews into a short plain-language themes report, with every claim linked back to actual review text.
Tag every review and ticket with your own fixed label list, so you can finally count things
Applies your own fixed label list to every review, ticket or survey answer, turning free text into a column you can pivot and chart.
Summarise what your NPS promoters and detractors actually wrote
Summarises the free-text answers behind your NPS score, band by band — what promoters praise and detractors cite, backed by direct quotes.
A Monday-morning digest of what customers complained about last week
A one-page weekly digest of support tickets — top issues, movement against last week, anything unusual — in the inbox before Monday stand-up.
Next
Tell us what your team does by hand
Describe the process that takes the most time. We will say plainly whether a model is the right tool for it.