Newsletter
Stay Ahead
Occasional, engineering-led notes on applied AI — what we're building, running, and learning in production.
Occasional, engineering-led notes on applied AI — what we're building, running, and learning in production.
Use case: Group reviews, tickets, and survey responses into themes
Reading through thousands of reviews, support tickets, or open-ended survey responses one at a time doesn't scale, and a keyword tag cloud misses anything phrased unusually. Webanto's clustering use case embeds each piece of feedback with POST /api/v1/embeddings (model nomic-embed-text), groups the resulting vectors with a standard clustering algorithm such as k-means or HDBSCAN run in your own analysis code — Webanto returns vectors, not clusters — and then calls POST /api/v1/chat (model qwen) once per cluster to generate a short human-readable label from a sample of that group's feedback. The output is a ranked list of themes, such as 'shipping delays' or 'checkout confusion', instead of a spreadsheet of raw text nobody has time to read. Cluster quality depends on picking a sensible number of clusters and reviewing edge cases by hand; the pipeline surfaces patterns for a human to act on, it doesn't replace that judgement.
Thousands of open-ended responses reduced to a ranked list of themes instead of a wall of unstructured text nobody reads in full.
Themes surfaced by actual semantic similarity in the feedback, catching complaints phrased in different words that a keyword search would miss.
A repeatable process that reruns on each new batch of feedback, so theme tracking doesn't mean re-reading everything from scratch every quarter.
Faster triage — a support or product team can jump straight to the largest or fastest-growing cluster instead of scanning tickets in arrival order.
Export the feedback you want to analyse — reviews, tickets, survey free-text — into flat records with a stable ID and the raw text.
Call POST /api/v1/embeddings with model "nomic-embed-text" for each record. 8,000 survey responses ÷ 100 per credit = 80 credits.
Run a clustering algorithm — k-means with a chosen k, or HDBSCAN if you'd rather not pick k upfront — against the returned vectors in your own analysis environment; Webanto's API returns vectors only, not cluster assignments.
Inspect cluster sizes, merge clusters that clearly represent the same theme, and set aside singleton outliers before labelling.
For each cluster, sample 10-15 responses closest to the cluster centroid and send them to POST /api/v1/chat with model "qwen", asking for a short theme label and one-sentence summary.
Chat calls here are small and infrequent — one per cluster, not per response — so even a run producing 40 clusters costs only a handful of credits against the 2,000-tokens-per-credit chat rate.
Review the generated labels against a broader sample of each cluster's actual text before publishing results; an AI-written label can be plausible but slightly off, especially for a cluster that turns out to mix two related themes.
Just the embeddings. POST /api/v1/embeddings returns a vector per piece of feedback; running k-means, HDBSCAN, or any other clustering algorithm over those vectors happens in your own code. That keeps you free to choose whatever clustering approach and cluster count fits your data, rather than being locked into a fixed method Webanto picked for you.
There's no universally correct number — it depends on how granular you want the themes to be. Start with a rough guess, around 10-20 for a few thousand responses, look at cluster sizes and separation, and adjust. HDBSCAN avoids picking a fixed count upfront but can produce a large 'noise' bucket that needs separate review.
Yes. Qwen is summarising a sample of each cluster, not the whole thing, so if a cluster is genuinely mixed — two distinct issues that happened to embed close together — the label may only capture one of them. Always spot-check labels against a broader sample of the cluster's actual text before treating them as final.
Probably not. At that volume a human can read every response directly in less time than it takes to build and validate a clustering pipeline. Clustering earns its cost once volume is high enough that manual reading is no longer realistic, typically in the thousands.
Keyword tagging is cheaper to build but brittle — it only catches feedback using your chosen keywords and needs constant maintenance as language shifts. Embedding-based clustering catches semantically similar complaints regardless of exact wording, at the cost of needing a clustering step and some manual review of the output that keyword tagging doesn't require.
Shoppers who spot a product in the wild, on someone else's feed, or on a competitor's site rarely know the right keywords to find it in your catalogue. Visual search closes that gap: a shopper uploads
Large media libraries accumulate redundancy fast — the same product shot re-exported at a different crop, a stock photo uploaded twice under different filenames, a banner re-saved at a lower compressi
Reverse image search answers a narrower question than catalogue browsing: given one photo someone just handed you — a screenshot, a competitor's product shot, a customer's snapshot of a damaged item —
Start a 14-day free trial on any paid plan. No credit card required.