ITMTB · Enterprise search
Search that finds what you already have.
Domain-tuned search for catalogues, report libraries and knowledge bases. Measured on your own content before it ships, deployed beside your existing application, and fine-tuned on your users' real queries for the first month.
On a 25,000-report catalogue, searches returning nothing fell from 53% to 5%.
From the live deployment for The Business Research Company, a market research publisher. Success story.
Why it works
Most "no results" are retrieval failures, not content gaps.
In the catalogue behind the numbers above, 470 of 500 sampled failed searches had a relevant item that already existed. The engine is built around the ways people actually ask.
Finds what people mean, not what they typed
Spelling corrected from your own vocabulary, so specialist terms are never 'fixed' into something else. Longer queries narrow instead of failing. Abbreviations, hyphenated names and two-letter terms are indexed and kept whole.
Measured on your catalogue before anything ships
A known-item test set generated from your items scores every change per variation class. A class dropping while the overall figure holds is a regression and does not ship. Live, a console shows every search, its quality, its latency and its clicks.
Safe to put in front of revenue
Runs as a separate service with its own schema, user and connection pool, so it cannot touch orders or enquiries. The existing search answers if the new one fails or times out. Rollback is one configuration value.
One answer for people and AI assistants
The same engine is exposed as an MCP tool, so an assistant asking for an item gets what the search box would show. Semantic rescue and related items run behind a per-customer switch, gated so nonsense still returns nothing.
What you get
An engine, a console and the operations around them.
Search engine
Lexical retrieval tuned to your corpus: catalogue vocabulary, convention-word demotion, progressive relaxation, exact-title and exact-subject promotion, geography and year qualifiers.
Search console
Every search with its results, a quality label (zero, weak, ok, rescued), latency per surface with sample sizes, error counters, clicks per surface, a query explorer and CSV export.
Business rules
Pin, exclude or redirect specific items for specific queries from the console, with a preview, without a deployment.
MCP tool
The same search available to AI assistants and agents through the Model Context Protocol, scoped per project.
Operations
Nightly index rebuild, fifteen-minute incremental refresh, drift check against the source catalogue, health endpoint, alerts when the index falls silent, tagged releases with approval.
Semantic channel
Meaning-based rescue for queries the lexical path cannot answer, and related items after a short confident answer. Small local model, no external API, no per-query cost. Off by default, on when it earns its place.
How an engagement runs
Evaluate first. Deploy where you want it. Fine-tune for a month, included.
Evaluate on your catalogue
Before any commitment, ITMTB generates the test set from a sample of your items and your real failed queries, runs it against your current search, and returns the per-class table. The decision rests on numbers, not a demonstration.
Deploy on-premises or as SaaS
On your infrastructure beside your existing application, or hosted by ITMTB. Either way the engine indexes your data into its own schema, is called through a local proxy with a timeout, and falls back to your existing search at every step. Your search box and results page do not change.
One month of fine-tuning, included
The first month after launch is where a search becomes yours. It is part of the engagement, not a change request.
- •The console is read daily; every weak, zero and rescued search is reviewed by a person.
- •Every real failed query becomes a committed test case before it is fixed, so it cannot regress.
- •Ranking fixes from your users' own queries: relaxation limits, convention words, qualifiers, compound terms.
- •Vocabulary and synonym candidates drawn from your catalogue and your search logs, reviewed with your team.
- •Business rules set up with your team for the queries where commercial priority should decide.
- •Latency budgets and timeouts tuned to your traffic; counters confirmed at zero.
- •Semantic rescue calibrated on your data, switched on only if the nonsense guardrail holds.
- •Your team trained on the console: reading demand in zero results, clicks and the explorer.
Measure every month
Catalogue reach, demand visible in honest zero results, click behaviour per surface, and the join from search to the actions that matter in your business. Engine releases stay gated on the evaluation.
Success story
The Business Research Company
A global market research publisher with 25,700 reports was telling more than half of its searchers that the report they wanted did not exist. Within three weeks of design sign-off, ITMTB's engine was in production on their site, with the previous search as a fallback at every step.
- 53% → 5%of searches returning nothing
- 98.8%known-item accuracy across 20 variation classes
- 1 in 10searches rescued by catalogue-vocabulary typo correction
- 0fallbacks, degraded searches and dropped events, all counted
Ready to Transform Your Business?
Join industry leaders already scaling with our custom software solutions. Let’s build the tools your business needs to grow faster and stay ahead.
FAQ
Frequently asked questions
Reading on enterprise search
The method, the measures and the deployment, in full.
Everything on this page is written up with its evidence. Start with the success story.
The Business Research Company: zero-result searches from 53% to 5%
What was wrong, what was built, how it was measured and what changed for a 25,700-report catalogue.
Read
How to measure the success of an enterprise search tool
Twelve measures across offline tests, live telemetry and business outcome, and five ways the numbers can mislead.
Read
Why enterprise search is important
Six reasons search decides more than it is given credit for, and a short check for your own catalogue.
Read
A realistic dataset for comparing enterprise search solutions
The 20 query classes, the nonsense guardrail, real failed searches, and how to score vendors fairly.
Read
Improving enterprise search for people and AI agents
The retrieval failures found in a real catalogue's logs, and why fixing retrieval came before any interface.
ReadStart with your own catalogue.
Send a sample of your items and a week of search queries. ITMTB returns the per-class evaluation of your current search and what a tuned engine would recover.