Everything Matters. Not Everything Gets Read.

Construction projects generate a staggering amount of paper. Drawings, specs, submittals, RFIs, contracts, schedules — a single mid-size project can easily hold hundreds of documents and tens of thousands of pages. Our platform ingests all of it, and our AI answers questions on top of it.

That gives us a rare vantage point: we can measure not just what a project contains, but what its people actually ask about. So we instrumented every retrieval call our AI makes in production — live capture, no sampling — over an extended window, and traced each one down through the data hierarchy we ingest: projects → documents → pages → chunks.

This post is what we found. Every percentage is unmodified production data; customer organizations are anonymized to their trade, and we've kept absolute figures to rough scale on purpose.

The experiment

Against that corpus, thousands of queries were executed against hundreds of live projects — making tens of thousands of retrieval calls under the hood. That's the dataset for everything below.

The anatomy of a question

Start at the smallest unit. A typical retrieval call is remarkably narrow: it reads about 10 chunks, from 7 pages, in a single document. Even a typical query — which usually makes several retrieval calls under the hood as it reasons — settles at 29 chunks across 19 pages in 3 documents. And of what gets retrieved, only about a third survives into the final answer as a citation.

Here's what that looks like against a real project. Watch the questions come in — and how little of the corpus each one opens:

Each question is a needle-in-a-haystack exercise: a tiny, focused slice of an enormous corpus. Which raises the obvious question — after thousands of needles, how much of the haystack has been touched?

All the queries combined barely scrape the corpus

Pool every retrieval event across every active project and weigh it against those projects' combined corpus, and the aggregate answer is: still a small slice. How small? Before we show you — care to place a bet?

The funnel narrows fast as you go down the hierarchy: 14.2% of documents were retrieved from at least once, but only 8% of pages and just 2% of chunks. Citations narrow it further — 6.7% of documents, 3.2% of pages, 0.7% of chunks ended up cited in an answer.

The bigger the project, the smaller the slice

Pooled totals weight every chunk equally, so a handful of giant projects dominate them. The honest way to look at it is one project at a time: for a typical project of a given size, how much of its own corpus do its queries end up touching? Time for your second bet.

Two patterns jump out:

1. In small and mid-size projects — about three in five projects — every single document gets queried. 100% document coverage, again and again. Meanwhile the typical share of pages touched drops to the teens, and chunks to low single digits.
2. The bigger the project, the thinner its queries spread. As projects grow, the share of their own corpus that questions touch collapses — a typical large project sees only a few percent of its documents and around 1% of its pages queried. And it's exactly those large projects that hold roughly four-fifths of everything we ingest. The corpus's weight sits precisely where usage density is lowest — not because that data doesn't matter, but because a large project's questions at any given phase concentrate on the working set that phase needs.

That asymmetry is the core finding. Coverage is total at the document level and sparse below it — which means the ingestion question is never whether to process a document, only how deeply.

Contractors focus on their trade — but read everyone's documents

Retrieval isn't just structurally concentrated; it's semantically concentrated too. We classified every retrieved chunk by the discipline of the document it came from, and broke it down per organization:

The diagonal you'd expect is there: electrical contractors retrieve most heavily from electrical documents (34–49% of their activity), mechanical contractors from mechanical documents (21–59%), the flooring contractor from architectural finishes, the civil contractor from civil drawings.

But the more interesting finding is everything around the diagonal. No organization's primary trade exceeds 59% of its retrieval — for most it's a third or less. General/project-wide documents pull 8–46% everywhere, and nearly every row has meaningful activity across plumbing, structural, electrical, and mechanical alike. A mechanical contractor bidding coordination questions needs the electrical drawings. An electrical estimator needs the architectural set.

And this matrix carries an ingestion lesson of its own. You cannot know, before ingesting, which document will end up mattering — a mechanical contractor's next question may land on an electrical drawing. But you can know what kind of information each document carries: its discipline and type are visible on day one, and usage patterns like this matrix arrive right after. That's the difference between treating all information equally and ingesting it intelligently.

What this means for ingestion

Three lessons, straight from the telemetry:

1. Never skip a document. The typical small-to-mid project ends up querying 100% of its documents — full awareness from day one is non-negotiable.
2. Never spend equally. The overwhelming majority of pages and chunks are never touched — uniform deep processing is pure waste.
3. You can't know upfront — but you can get close, then adapt. No classifier can predict the pieces of data nobody will ask about. What does work is a combined approach: predict the probably-less-touched data from feature signals — discipline, document type, and a few ingredients of our secret sauce — to set each document's initial processing tier, then let live usage adapt those tiers as the questions actually arrive.

So our pipeline is tiered: every document is ingested, indexed, and retrievable on day one — and depth is earned. Here's the whole idea in one interactive picture — ask it a few questions and watch what happens:

We're already feeding these diagnostics back into how we prioritize ingestion work. The corpus keeps growing — but growth is exactly why intelligence about how data gets used matters more than raw processing power ever will.

Methodology: live capture of all production retrieval events over an extended recent window, measured per query. Organization names are anonymized, absolute figures are deliberately kept to rough scale, and all percentages are unmodified production data.