The search debate that's forgetting where everything is

TL;DR: AI search covers meaning and connections but often misses location. Spatial search provides that third search component. FME MCP tools let AI assistants call published workspaces to answer questions such as 'whats near here?' from real data, without requiring GIS expertise.

There is quite the debate at the moment about whether AI should find information by meaning (Semantic) or by connections (Graph) but these camps are missing the spatial dimension?

A grid planner queries safety and maintenance records:
'Find inspection reports with severe tree encroachment or vegetation risk, on high-voltage overhead lines with active landowner wayleaves, where the clearance corridor directly crosses railway or road right-of-ways.'

Three questions in one coat:

• Meaning: 'Severe vegetation risk / dangerous tree overhang' - this is a subjective field assessments and photographic notes in patrol logs.
• Facts: 'Voltage: 33kV–132kV, Asset type: Overhead Line, Wayleave agreement: Active' - structured record attributes.
• Space: 'Corridor intersection with linear transport infrastructure (Spatial Polygon Overlay & 3D Proximity)' - line-of-sight easements, buffer polygons and spatial intersection against 3rd-party transport network layers.

The argument everyone is having

There is a loud argument in the AI world right now about how a computer should find information when you ask for it.

One camp says: search by meaning or semantic search. Store what text/images are about and match your question to whatever comes closest. This works beautifully for fuzzy questions ('something cosy, near water'), it is cheap to query and is language agnostic.

The other camp says: search by connections or graphs. Build a web of what relates to what: this flat is in this building, this building is served by this station, this landlord owns these properties. This works well for questions with many steps but can be expensive to build and maintain.

The question not being asked: where?

Data often has a spatial element and semantic or graph based search cannot be used to support searching spatially. 

This is not a niche gap. The questions that contain a place outnumber the search niche everyone is arguing about, and the money follows. Industry estimates put search-by-meaning software at around $2–3 billion a year; the software and services built around questions with a place in them are measured in tens of billions. Estimates vary with how the categories are drawn; the point is the order of magnitude, not the decimal places.

Why this is suddenly practical

Two things changed recently.

First, AI assistants learned to call out to other software. Your question stays in plain language; behind the scenes the assistant phones a specialist (the map, the property listings, the street network) and folds the answers together. The same way you would phone a friend who knows the area, except the friend never sleeps and never forgets the 22-minute uphill walk.

Second, small focused tools beat one giant do-everything tool. An assistant armed with three small tools (search by meaning, check exact facts, measure place) can answer questions no single tool could. It stops sounding plausible and starts being right, which is the difference between an assistant a regulator would trust and one that makes things up.

Analysts remain rightly sceptical: a large share of AI projects will be quietly shelved in the next couple of years, usually because the value was unclear or the cost crept up. Assistants that answer real questions with real measurements are the surviving half of that statistic. Assistants that amplify guesswork are not.

A note of realism

We have built these stacks, and the lessons were ordinary:

• Silence is the enemy. When a search breaks, the assistant must say so. An assistant that cannot tell 'nothing found' from 'search broke' will confidently tell you the archive is empty when it crashed.
• Exact details must be copied exactly. Names, codes, dates: if the assistant is expected to reuse an identifier, it must copy it character for character.
• Dates must mean the same thing to everyone. The calendar the question uses and the calendar the data uses must agree; half-working filters quietly return wrong answers, which is worse than failing.

None of this is glamorous. It is the ordinary discipline of making data solid, and it decides whether your assistant is trusted or tolerated.

One honest caveat: any tool that searches your data and sits on the public internet needs the same care as any public service. We published the security lessons from exposing search tools online, alongside a plain-English guide, as companions to this post: The Pipeline Is the Program: Exploiting and Hardening Transformers in FME-Powered MCP Servers and Keeping Your FME Tools Safe When You Share Them Online.

So what?

The search debate will keep generating engagement but if your organisation runs on assets that have a location, a search strategy without attention on spatial search will not .

The question worth asking: if someone asked your assistant 'what's within five kilometres of this, and what do the records say about it?', could it answer, or would it make something plausible up?