About Services Tools Blog Contact us

Hotspot Search by Meaning

Lets a visitor search groups of places by mood or activity as well as by place name, and searches by name still work.

  • Semantic Search & RAG
  • Leisure and sport
  • Tool
  • Product
  • Local/private
  • 2026

The problem

A visitor who searches for a mood or an activity misses groups of places whose names do not contain the words they typed. A search by place name still has to work. And the map area is a separate limit from the words typed.

What it does

Each group of places keeps a numeric summary of its meaning, built from its name, tagline, region and categories. A search is turned into the same kind of summary and compared with them. The result is combined with a plain name and region match, so place searches still hit. A summary is rebuilt only when the text or the model changes.

Outcomes

  • A mood or activity query finds groups that never use those words in their name.
  • A place-name query still matches directly.
  • Groups are not re-embedded unless the text or the model changes.

Simplified mockup

Interactive preview with sample data. Nothing is sent to a server.

Search clusters by meaning

quiet food nearby meaning
1
Lantern Yard
Canterburysupperevening

One room, low tables, supper that runs late. The name never says quiet.

0.82
2
River Shed
Maidstonesuppercovered

A short menu under a roof, beside the path along the river.

0.74
3
Orchard Room
Ashfordsuppergarden

Shared plates under the trees. Conversation stays at the table.

0.67

None of the names contain “quiet” or “food”. The rank is cosine similarity on the embedded tagline, region and categories.

Place name, then the others

Query Canterbury

A pure place search still hits. Lantern Yard takes a name-match badge and stays first. The other two clusters remain, with lower scores, because the lists are unioned rather than replaced.

1
Lantern Yard
name match
0.91
2
River Shed
Maidstone
0.41
3
Orchard Room
Ashford
0.36
Map bounds stay separate

The text rank above is not clipped by the map.

A 12 km window around Canterbury would keep Lantern Yard and drop River Shed and Orchard Room. That cut happens as an area limit, before anyone reads the similarity.

Bounds are not stored inside the embedding.

Re-embed only what changed

Stored text and hash
ClusterSource textHashNext run
Lantern Yardname, tagline, region, categoriesa91cunchanged
River Shedname, tagline, region, categoriesc22eunchanged
Orchard Roomtagline edited todayoldtext changed
Ingest
Ready. Two clusters already match the stored hash.

How it works

  1. Run a setup script that builds a meaning summary for each group of places with a local Ollama model (nomic-embed-text).
  2. Store the summary, the source text and a fingerprint of it, and rebuild only when they change.
  3. Turn a search into a summary and rank groups by how closely they match.
  4. Add a plain name and region match so place searches still work.
  5. Return a closeness score with each result, and a badge when the hit came from the name.
  6. Keep the map area as a separate limit.

Pipeline

flowchart LR
  Hotspot[Name and tagline] --> Embed[Embed and hash]
  Query[Visitor query] --> Vector[Vector rank]
  Embed --> Vector
  Query --> Name[Name and region match]
  Vector --> Union[Union the lists]
  Name --> Union
  Union --> Pins[Ranked clusters]