Construction /Retrieval (RAG) live field guide · 8 min

Project Knowledge Assistant

Ask a question and get the answer straight from the firm's own specs, head contract, ITPs, and standards — with the source passage, a link, and a confidence flag — instead of interrupting whoever last touched the file.

Open the live demo

theater/demos/construction_project-knowledge-assistant.html · sandbox · read-only

Open

FIG. 1

The live demo, running on fabricated data. Open it to step through the full flow — every output is shown for a person to approve before anything happens.

How it would work

Reads only your approved project documents, answers in plain language with the quoted clause and a match score, raises a caution beat where judgement is needed, and hands every answer to a person to accept, refine, or flag.

Input 01

The question + the indexed corpus

A staff question in plain language, run against a bounded set of approved documents for one job — the current project spec, the executed head contract, live ITPs, and referenced standards like AS 3600:2018.

Agent 02

Retrieves, quotes, scores

Matches the question to the indexed passages, drafts an answer from those passages only, attaches the source clause and section reference, and scores how well the answer is grounded in the retrieved text.

Output 03

A cited answer, for a person to accept

A plain answer with the quoted source, a match-to-documents score, and a caution beat where judgement applies — for a site engineer to accept, refine, or flag as wrong before anyone builds to it.

Where it works well

It turns "ask whoever remembers" into a cited answer in seconds, with the clause attached so the asker can check it.

Best for site engineers, document controllers and project managers on mid-to-large jobs where the file count is high enough that no one holds it all.
It earns its keep when the corpus is genuinely the source of truth and the questions are factual and document-answerable, not judgement calls.
The recaptured hours go back to the people who were fielding the same questions by hand — redirected to the work that needs their judgement.

The slow, invisible cost is the interruption tax: every time someone needs the minimum cover, the right ITP revision, or what a head-contract clause obliges, they break someone's concentration to ask — and the answer arrives in minutes, hours, or wrong.

Where it works badly

It is confidently wrong when the corpus is stale or ambiguous — and the match score measures grounding, not currency.

Leave three revisions of one document in a folder with no superseded marking and it can retrieve the wrong one with high confidence.
Scanned drawings that were never OCR'd are invisible to it, so it answers "that isn't in your documents" when in fact it is.
It is the wrong tool for judgement — "should we vary the pour given the forecast?" is informed by documents but not answered by them.

The honest test

Pick a question your team asks weekly and check whether one document, in one current revision, definitively answers it. If the real answer is "it depends, ask the engineer," no retrieval tool fixes that.

Index last month's spec instead of the current revision and it will quote superseded reinforcement detail at 94% — because the score says how well the answer tracks the indexed text, not whether that text is still correct. That is the trap.

What it doesn't do — and shouldn't

It surfaces and quotes. A person decides what is authoritative and what to build to.

WHAT IT DOES

Quotes the spec clause and the standard's table, with the section reference

Shows a match-to-documents score and the documents it retrieved from

Raises a caution beat where a decision has physical or contractual consequence

WHAT IT WON’T

Approve a pour, sign an ITP, or interpret a contractual entitlement

Resolve a conflict between two documents — it shows both and stops

Confirm a figure is right for the as-built condition on the ground

The consequences are physical and contractual: getting cover, lap lengths, or a head-contract notice period wrong has safety, durability and liability consequences a confidence bar cannot carry. The accountable person stays on the decision because the consequence lands on them — not the tool.

What your data has to look like

One governed copy of each document, at its current revision, with structure a clause can be cited from.

Typical readiness

across orgs we see, before the first job

One authoritative copy per document

Needs shaping

Citable structure inside each document

Usual weak point

Searchable text, not image-only scans

Needs shaping

Access roles defined

Usual weak point

Re-index on document change

Needs shaping

The real first job

Getting to one governed, current, searchable source per document is usually the real first job — bigger and more valuable than the retrieval layer on top. It is a question of how documents are captured and controlled, not of buying a tool, and it is the work that makes the answers trustworthy rather than merely fast.

Right fit if…

The corpus is genuinely the source of truth — current spec, executed contract, live ITPs, the standards you build to

Mid-to-large jobs where the file count is high enough that no one person holds it all

The same factual, document-answerable questions recur across a long build

You can point to one current revision of each document, with superseded versions out

Walk away if…

Your drawings register is three weeks behind and superseded revisions sit unmarked

Half your drawings are image-only scans that were never OCR'd

A small single-project job where everything fits in one folder and one person knows it

The questions you ask are really judgement calls — "it depends, ask the engineer"

Open questions

The worried-buyer questions, answered straight

It can surface a wrong figure if the corpus is wrong — which is exactly why every answer carries the quoted source passage, a link to the document, and a match-to-documents score, and the demo shows an amber caution beat telling you to confirm the exposure classification with the project engineer before a pour. The assistant surfaces what the spec and the standard say; a competent person confirms it against the as-built condition and signs off. It makes checking faster, not optional.

Only as well as the corpus you point it at. If three revisions of the spec sit in one folder with no superseded marking, it can quote the wrong one with high confidence; image-only scans it can’t read at all. Getting to one governed source per document — current revision, superseded versions removed, scans run through OCR — is usually the real first job, and it is the work we help with before the AI layer earns its keep.

No. The document controller still owns which revision is current and what is authoritative; the project engineer still makes the engineering judgement. The assistant recaptures the time those people lose answering the same question by hand and redirects it to the work that needs their judgement. It draws the line deliberately: it surfaces and quotes, a person decides.

Current enough that the newest authoritative revision is the one indexed and superseded ones are out. The corpus shows each document’s section count and last-updated date so you can see what it is answering from. Re-index on a stale snapshot — last month’s spec, a head contract amended since execution — and it will answer confidently from the old text. Re-indexing on document change is part of the setup, not an afterthought.

It stays inside your own tenancy — your SharePoint or Drive — and retrieval runs against your private corpus, not a public model trained on your files. Access roles mean a site labourer and a commercial manager don’t see the same documents. Nothing in the head contract or a past tender leaves your environment to answer a question, and the demo here runs entirely on fabricated data for a project that does not exist.

It says so and refuses to guess, rather than inventing a plausible number. The demo’s second question — annual leave for site labourers — is answered by declining, because that lives in the HR system, not the project knowledge base, and it then offers to escalate to a named human. A logged gap also lets the document owner decide whether that material should be indexed.

What it takes to build

3–4 weeks · 4 phases

Reused from template~70%

Bespoke to this skin~30%

stack · Private RAG · SharePoint/Drive · LLM

What it would cost

Fixed scope, fixed price, fixed dates.

Bite-sized first piece

One contained change, low risk

Pilot build

Most builds land here

Embedded support

Scale on proof

Considering this for your project knowledge?

The honest place to start is a bite-sized first piece — one job's corpus, one set of recurring questions. Tell us where the interruption tax hurts; we'll play it back, scope it, and show you what's possible.

Book a call How we work