Build a studio knowledge assistant (RAG)
Stop the model guessing from the open internet. Point it at your own past projects, specs and standards — so every answer is grounded in documents you actually trust.
Twelve years of projects, all locked in folders nobody can search. RAG hands you the key.
A mid-size Bengaluru firm had done the same kind of school building four times across a decade. Each time, a junior asked the same questions — what stair width did we use, which fire consultant, what was the spec for the sports floor — and each time someone hunted through old folders for an hour, or just re-decided from scratch. The knowledge was all there, on the server, completely unsearchable. Then they built a small assistant that sat on top of those folders. Now a junior types 'what stair detailing did we use on the last CBSE school?' and gets the answer, drawn from the firm's own past drawings and specs, with the project named. The model didn't get smarter. It got grounded.
Retrieval-augmented generation — the model that reads your filing cabinet first
Look it up first, then answer — in plain English
A normal LLM answers from what it absorbed during training — the open internet, frozen at a point in time, with no idea about your studio. That's why it hallucinates your specifics: it has never seen them.
RAG — retrieval-augmented generation — changes the order of operations. Before the model answers, the system first retrieves the most relevant chunks from a collection of documents you provide, then hands those chunks to the model and says, in effect, 'answer using these.' So the model isn't guessing from the open internet; it's summarising passages pulled from your own files.
The plain-English version: imagine an intern who, before answering any question, runs to your filing cabinet, pulls the three most relevant pages, reads them, and then answers from those pages — and tells you which folder they came from. That's RAG. The answer is grounded in your documents, and a good setup cites which document each part came from, so you can verify it. It dramatically reduces hallucination on your specifics, because the model is reading rather than recalling.
Your past projects, specs and standards become a searchable brain
The value of a studio assistant is entirely in what you feed it. Good corpora to ground it on: past project specs and BOQs, completed drawings and their notes, your office standards and detail library, consultant reports, your own SOPs and templates, and product/material data you've vetted. The richer and more yours the corpus, the more the assistant becomes institutional memory you can talk to.
Conceptually the approach is the same everywhere: collect the documents, split them into searchable chunks, index them so the system can find the relevant ones, and connect that index to an LLM that answers from the retrieved chunks. As of 2026 you can build this without coding — Claude and ChatGPT both support pointing a 'project' or custom assistant at a set of your own files; Microsoft Copilot grounds answers in your firm's SharePoint and Office documents; Google Gemini does the same over Workspace. These are RAG under the hood, wrapped in a friendly interface.
Start tiny. One project type, a handful of clean documents, a dozen real questions you ask repeatedly. Prove it answers from your files and names its sources before you scale it across the office.
RAG is only as good as the documents you feed it. Garbage in, confidently-cited garbage out — curate the corpus like you'd curate a portfolio.
Your client documents are confidential — never feed them to a public model carelessly
Here is the part that matters most, and it's the easiest to get wrong in the rush. A studio knowledge assistant is built on your clients' confidential data — their drawings, briefs, budgets, sometimes their personal details. Feeding that into a public AI model without care is both a confidentiality breach and, in India, a matter for the Digital Personal Data Protection Act, 2023 (DPDP Act).
So treat data-handling as a first-class design decision, not an afterthought. Know where your documents go: a public consumer chat tool may use inputs differently from an enterprise tier with a no-training, data-isolation agreement. Prefer setups that keep your corpus private and contractually ring-fenced; strip or anonymise personal and commercially sensitive details where you can; and get client consent before their project goes into any assistant. The rule from the ethics module applies in full force here: never feed confidential client data or drawings into a model whose data terms you haven't read.
Done right, RAG is one of the safest AI uses there is — because it answers from documents you control. Done carelessly, it's a leak with a search box.
Your firm is sitting on the most valuable training data you'll ever have — a decade of solved problems. A grounded assistant over your past specs, details and consultant reports turns that into searchable institutional memory and onboards juniors in days, not months. Build it on an enterprise tier with a no-training, data-isolation agreement, anonymise client-identifying data, and insist the assistant cites the source project for every answer so a professional can verify before relying on it. The assistant recalls the firm's precedent; you still decide whether the precedent fits this project.
Point it at your past FF&E schedules, your vetted vendor list, your finish specs and your project photos with notes. Then 'what was that brass tap we loved on the Koramangala flat, and who supplied it?' gets answered from your own records instead of your memory. It's a real edge on repeat clients and fast pitches. Guard the privacy line hard: client budgets and home layouts are personal data — anonymise, get consent, and use a tier whose terms you've read before any client's project enters the assistant.
You may have only a handful of past projects — but you also have no team to remember them for you, so a small grounded assistant is pure leverage. As of 2026 you can stand one up with no code: a Claude or ChatGPT 'project' pointed at a folder of your own clean documents. Start with one project type and the ten questions you keep re-answering. On privacy you're both the studio and its only data-protection officer, so read the data terms yourself and keep genuinely confidential client material out of any consumer-tier model.
Claude Projects (Anthropic)
RAG over your own files
Point a Claude project at your specs, briefs and standards; its long context handles large documents well. Use an enterprise tier with clear no-training data terms before any confidential client material goes in.
ChatGPT custom GPTs / projects
RAG over your own files
Upload a curated document set and ask grounded questions, with no coding. Convenient and fast to prototype; check the data tier's training and retention terms for client confidentiality.
Microsoft Copilot
Grounded in firm SharePoint/Office
Answers from your firm's existing SharePoint, Word and Outlook content — RAG over documents already in your tenant. Good fit if your office already lives in Microsoft 365; enterprise data terms apply.
Google Gemini (Workspace)
Grounded in Google Workspace
Long context plus retrieval over your Drive, Docs and Gmail. Convenient if your studio runs on Workspace; same rule — read the data terms and ring-fence client data before relying on it.
“If I build a RAG assistant over my own documents, it can't hallucinate any more — every answer is grounded in my files.”
RAG dramatically reduces hallucination on your specifics, but it doesn't eliminate it. The model can still misread a retrieved passage, blend two documents, or answer confidently when retrieval pulled the wrong chunk or no chunk at all. That's why a good assistant cites its source for every claim and why you verify load-bearing answers against the named document. Grounded means 'much more likely to be right and checkable' — not 'guaranteed true.' The human stays in the loop here too.
Workshop — stand up a tiny grounded assistant on your own files
You'll build a minimal studio knowledge assistant over a handful of your own (anonymised) documents and test whether it answers from YOUR files, with sources, instead of guessing. About an hour, a free or trial tier that supports custom projects.
Free/trial: a Claude project or a ChatGPT custom GPT. 5-10 of your own clean, anonymised documents (specs, BOQs, standards) and a list of questions you ask repeatedly.
GROUNDED-ASSISTANT SYSTEM PROMPT (paste into the project's instructions):
You are my studio knowledge assistant. Answer ONLY from the
documents I have provided in this project.
Rules:
1. Base every answer on the uploaded documents, not general
knowledge.
2. After each answer, CITE the source: name the document
and the section it came from.
3. If the documents do NOT contain the answer, say exactly:
"Not found in the provided documents" - do not guess.
4. Never invent a code, product, price or supplier that is
not in the documents.
Test questions to ask it (replace with your own real ones):
- "What stair width did we specify on [project]?"
- "Which fire consultant did we use, and on what project?"
- "What was our standard spec for [a finish]?"- 1Curate and anonymise 5-10 of your own documents — strip client names, personal details and anything you wouldn't want a model to retain. This curation IS the build.
- 2Create a Claude project or ChatGPT custom GPT, upload the documents, and paste the system prompt above into its instructions.
- 3Ask your real repeat questions. For each answer, check that it cites a real source document and that the cited content actually says what the assistant claims.
- 4Probe the edges: ask something your documents genuinely don't cover. A good setup says 'Not found in the provided documents'. If it invents an answer instead, tighten the prompt — that's RAG failing loud, which is what you want.
- 5Verify one load-bearing answer against the original document yourself. Grounded is not the same as guaranteed.
- 6Note the privacy posture: write one line on where these documents now live and the data terms of the tier you used — then decide whether real (non-anonymised) client work could ever go in.
You’ll walk away with
A working minimal knowledge assistant that answers questions from your own documents and cites its sources, plus a clear-eyed note on its privacy posture — the seed of institutional memory you can talk to, built the safe way.
Two quick probes, five minutes.
- 01Ask your assistant the same question twice, worded differently, and check it cites the same source both times — consistency is a sign retrieval is working.
- 02Ask it a deliberately out-of-scope question (something not in your docs) and confirm it admits it doesn't know rather than inventing an answer.
RAG flips the order: the system retrieves the relevant chunks from YOUR documents first, then the model answers from them and cites the source. That turns a decade of past projects into searchable institutional memory and slashes hallucination on your specifics. It's one of the safest, highest-value AI uses — provided you guard the privacy line and verify load-bearing answers against the named source.
RAG = retrieve from your own documents first, then generate a grounded, cited answer. Point it at past specs, BOQs, standards and reports; build it no-code with Claude projects, ChatGPT GPTs, Copilot or Gemini. Curate the corpus, cite every source, verify load-bearing answers — and never feed confidential client data into a model whose data terms you haven't read (DPDP Act, 2023).
What is RAG in simple terms?
RAG — retrieval-augmented generation — means the AI looks things up before it answers. Instead of replying from what it absorbed during training (the open internet), the system first retrieves the most relevant passages from documents you provide, then the model answers using those passages and cites them. Think of an intern who pulls the right pages from your filing cabinet, reads them, then answers from them — grounded in your files, not its memory.
Can I build an AI assistant over my own projects without coding?
Yes, as of 2026. Claude projects and ChatGPT custom GPTs both let you upload a set of your own documents and ask grounded questions with no code. Microsoft Copilot grounds answers in your firm's SharePoint and Office files, and Google Gemini does the same over Workspace. Start small — one project type, a handful of clean documents — confirm it cites real sources, then expand.
Is it safe to put my client documents into an AI knowledge assistant?
Only with real care. Client drawings, budgets and personal details are confidential, and in India fall under the DPDP Act, 2023. Use an enterprise tier with a no-training, data-isolation agreement rather than a public consumer tool, anonymise personal and commercially sensitive details where possible, and get client consent before their project enters any assistant. Never feed confidential client data into a model whose data terms you haven't read.
An assistant that answers questions is powerful. But what if the AI could also DO the repetitive tasks — sort the email, draft the follow-up, file the document — not just answer about them? That's the leap to agents and automation, and its own set of guardrails.