Tips for getting the most out of Relatiq.
Short, practical notes on search profiles, index sizing, and getting recommendations to actually work.
Everyone has adopted AI. Most haven't made money from it.
Adoption and financial return have decoupled — and the gap isn't a model-choice problem. It's a data-readiness problem most companies skip past.
Read more →November 15, 2026Classification vs regression: the first decision in any ML project
Predicting a category versus predicting a number sounds like a small difference. It decides which models apply, how you measure success, and what a wrong answer even means.
Read more →November 13, 2026Why intermittent demand needs a probability, not a point forecast
Spare parts, low-volume SKUs and anything mostly zero break ordinary forecasting in a specific way — and the fix is reporting a range instead of a single confident number.
Read more →November 10, 2026Can a classifier actually build your product taxonomy
A classification model can place items in a hierarchy from their attributes and descriptions alone — but only governance rules keep thousands of individually reasonable decisions consistent as a whole.
Read more →November 5, 2026Why retrieval-augmented generation still needs hybrid search
A RAG pipeline is only as good as what it retrieves. Pure vector search misses exact codes, rare terms and multi-step questions — exactly where hybrid retrieval earns its cost.
Read more →October 27, 2026Turning ERP and CRM data into fast, meaningful search
Most internal search problems aren't a search problem first — they're a connector problem. Here's the pattern we keep seeing when customer, product or document search needs to run on data trapped in an ERP or CRM.
Read more →October 23, 2026Personalizing a newsletter means more than picking content
We're splitting newsletters into measurable blocks so personalization can change which content runs and how the newsletter itself is structured, not just the links inside it.
Read more →October 20, 2026Setting up Relatiq Search from a Google product feed
Paste a Google Merchant feed URL and Relatiq Search previews your products, suggests a field setup and starts indexing — no manual field mapping, no config file.
Read more →October 16, 2026Picking the right search profile
Default, Description, Mix or Name — what each profile actually changes and when to use it.
Read more →October 14, 2026Sizing your first index
How much vCPU and RAM a new Relatiq Search index actually needs, and what to watch as it grows.
Read more →October 10, 2026What Tracker needs before it can recommend anything
One script, a handful of real sessions, and what the first profiles look like before recommendations are worth trusting.
Read more →October 8, 2026How vector embeddings turn meaning into math
The idea behind semantic search, explained without the hand-waving — what an embedding actually is and why distance between vectors means similarity in meaning.
Read more →October 6, 2026Precision and recall: why search quality is always a trade-off
The two numbers that actually define search or classifier quality, why improving one tends to hurt the other, and what NDCG adds for ranked results.
Read more →October 4, 2026What makes time series forecasting hard
Trend, seasonality and noise — the three things every forecast has to separate, and why naive methods quietly fail once real-world data is involved.
Read more →September 30, 2026Why hybrid search beats pure semantic or pure keyword
Keyword matching and semantic embeddings fail in opposite, complementary ways. Hybrid search isn't a compromise between them — it's what covers both failure modes at once.
Read more →September 28, 2026How anomaly detection actually works
Anomalies are rare and unlabeled by nature, which rules out standard classification. Here's what statistical and model-based detection actually look for instead.
Read more →