Where Context Lives on a Job Site

Abstract

Retrieval for field work is usually framed as a document-search problem. Our ongoing inventory of where construction project information actually lives — plans and revisions, photo streams, text threads, schedules, estimates, and unrecorded expertise — suggests the retrieval problem is operational, not semantic: the right answer depends on job, phase, revision state, and crew, not on textual similarity. This note documents the inventory structure and the retrieval design constraints we are deriving from it for the Construction Field Assistant.

What we investigated

Before building retrieval for a construction assistant, we asked a prior question: where does project context actually live, and in what form? This note is a structured inventory of the information surfaces on residential and light-commercial projects — drawn from workflow conversations with people in the trades and from the document sets of real projects we have been given access to.

Why it matters

Every AI retrieval demo assumes a corpus. Construction does not have a corpus; it has a scatter: the current plan set (and its superseded revisions), photo streams keyed to nobody’s filing system, text threads that contain binding decisions, schedules that drift from reality, and estimates whose line items encode scope. An assistant that treats this as “documents to embed” will confidently retrieve the wrong revision — which on a job site is worse than retrieving nothing.

Method

Qualitative and ongoing: semi-structured conversations with contractors and field staff about how specific, recent questions actually got answered (critical-incident style), plus document-set analysis of completed projects. We record, for each incident: where the answer lived, how it was found, how long it took, and what would have made it wrong (revision, scope change, verbal override). The sample so far is small and regional; we are explicit about that below.

Findings

Preliminary, from a small sample:

Recency and revision state dominate relevance. Most retrieval mistakes people described were right document, wrong version — a semantic match and an operational miss.
Binding context is conversational. Decisions that override documents frequently live in text threads and phone calls; a retrieval system blind to communication surfaces will authoritatively cite superseded plans.
Questions are short; context requirements are deep. Field questions average a few words (“what’s the header height on the west opening?”) and depend almost entirely on implicit job/phase context the asker will not state.

Limitations

The sample is small, regional, and skewed toward residential work; findings may not generalize to heavy civil or large commercial projects. Incident recall is retrospective and subject to memory bias. No quantitative retrieval evaluation exists yet — that requires the corpus assembly this note is scoping. This is a working analysis intended to be revised.

Product implications

For the Construction Field Assistant, this inventory converts directly into architecture: retrieval must be scoped by job and phase before ranking by similarity; revision supersedence must be modeled as first-class metadata; and communication threads must be ingestible context, not an out-of-scope inconvenience. It also sets the evaluation bar: the correct metric is not “found a relevant document” but “found the operative one.”