AstorX Insights

Why keyword search fails on a 300-page tender

Every tender team has a story about the requirement nobody found until it was too late. Usually it was findable, just not by the word anyone searched for.

The vocabulary problem

Search for "penalty" and you will not find the clause written as "liquidated damages," "service credits," or "price reduction in case of delayed delivery." Search for "shelf life" and you will miss "residual validity at the time of delivery shall not be less than two thirds." Tender authors are not trying to hide obligations; they are lawyers and procurement officers writing in their own registers, across documents drafted years apart. The obligation is real. The keyword is whatever that author happened to type.

This gets strictly worse across languages. A Czech tender's binding text does not contain your English keywords at all, and machine-translating 300 pages so you can Ctrl+F the translation introduces its own errors precisely where precision matters: units, negations, modal verbs.

The structure problem

Even when the word matches, keyword hits arrive as a list of disconnected locations. Hit 14 says delivery within 30 days. Hit 29, in a different document, says 14 business days. A keyword tool shows you both and understands neither; spotting that they conflict is left to whoever reads the hit list, usually quickly and usually late in the day.

Obligations also split across sentences and sections. "The supplier shall comply with the requirements of Annex 4" is meaningless without Annex 4, which modifies a table in Annex 2, which an amendment later revised. No keyword query reconstructs that chain. People can, if they have the week it takes. Mostly they do not have the week.

What semantic analysis actually changes

Modern document AI reads the way a careful human does, at machine speed: it understands that "residual validity" is a shelf-life requirement, that a figure in a Slovak annex and a figure in the English master refer to the same delivery obligation, and that two passages contradict each other even when they share no vocabulary. The output that matters is not a chat answer. It is a structured view of the tender: every requirement, deadline, value, and obligation, each one carrying a citation to the exact source passage so a human can verify it in one click.

The citation point deserves emphasis. The failure mode of naive AI is confident summary without provenance, which in tender work is worse than useless. You cannot take "the AI said so" into a bid defense. If a system cannot show you page and paragraph for every claim, it is a liability with a nice interface.

A fair test for any tool, including ours

Take a tender you know intimately, one where you already found the painful requirement the hard way. Run it through the tool and check three things: Did it surface that requirement unprompted? Does every finding cite its source? Did it catch the cross-document inconsistencies you know are in there? Twenty minutes with a tender you have already lived through beats any demo script a vendor can prepare. We hold that bar for AstorX, and we would encourage you to hold every vendor to it.