Chat Over Your Documents, Inside the Product

Chat over one corpus, inside the product, with answers a person can check. One user is a two-week build. Access rules and refusals are the longer scope.

By Muhammad Saif and Aashan Khan · Techtribe co-founders

A folder of PDFs and a browser tab that answers questions about them is a demo. It is a useful demo. It is not a feature of your product. Your user already has an account. The document already belongs to that account, or to that company, or to that case. The chat has to sit on that screen, ask about that file, and refuse when the file does not say the thing. That is the build this note is about.

The public version of this idea is a site where anyone uploads a PDF and talks to it. The document chat guide is the product shape we mean instead: the upload, the question, and the answer live inside software you already sell, or inside the first version of it. The service that names the retrieval work is RAG development. This page is the scope: what the first chat includes, what it must not pretend to know, and which of the two published packages is the honest one.

One corpus, one question, one citation

The first version has one body of material. A handbook. A set of contracts of one type. The listings for one catalog. The policies a support agent is allowed to quote. Not "everything the company has ever written." Everything is a migration, a permissions project, and a search problem wearing a chat box.

The user asks in a sentence. The answer comes back with a place to look: a page, a section, a row. If you cannot click from the answer to the passage, you do not have a citation. You have a confident paragraph. Confident paragraphs are how a wrong policy gets repeated in a customer email. The click is the product. The prose is the convenience.

When the passage is not there, the answer is not found. Not a guess with a hedge. Not a general explanation of the topic from the model's training. The user asked about your document. If your document is silent, the product is silent, and it says so in the interface. That refusal is a feature you can test. Ten questions the corpus answers, and two it must not, are enough to see whether the build is the one you ordered.

A tidy set of files you can hand over is the scope. A scan that is a photograph of a photograph, a spreadsheet with three headers for the same column, or a wiki nobody has opened since a reorg, is data work. Data work can be larger than the chat. Chat over a tidy set is one feature, and a cleanup is a different conversation. Say which one you have on the call, before anyone offers you a date.

What "inside the product" rules out

The chat does not get its own login. The person is already signed in, or this is the first product and the sign-in is the same one the rest of the app will use. A second password for "the AI" is a second account-recovery problem and a second place to get the user wrong.

The chat does not get its own copy of the truth that drifts. If the document is updated, the next question sees the update, or the screen says which version it is reading. A hidden index from last quarter, still answering in the present tense, is worse than a slow index. People trust the voice. They do not check the date unless you show the date.

The chat does not send the document to a place you cannot name. The provider, the region, and whose key is being billed are decisions you make on purpose. A key in an account you cannot see is how a pilot ends in a surprise bill. Prefer your account. Put a spend alert on it before the URL is shared. The monthly side of that bill, after the build, is in AI app maintenance cost. This article is the build, not the token invoice.

The chat does not answer from the open internet because the open internet had a smoother sentence. If you wanted a general assistant, say so, and we will tell you that you do not need us to wrap a model you already pay for. The reason to build this is that the answer has to come from your material, on your screen, for your user.

The two-week shape

The short build is one core flow: the person opens the document that belongs to them, asks, and reads an answer they can check against a passage. One AI feature, not a platform. A clean interface. A live URL. The code in your repository from the first commit. Support on what shipped.

It does not mean a permissions matrix. It does not mean evaluations as a suite, monitoring, or fallbacks across providers. It does not mean your brand's design system, and it does not mean a second integration. If the only integration is "read this folder" or "read this bucket you already named," that can fit. If you also need the CRM, the inbox, and the billing tool, you have named three systems. A short build includes one. The longer description of that clock is an LLM MVP in two to four weeks. Bring the folder to the call and we will say whether the short build is honest.

Bring five real documents and five real questions, including two the documents cannot answer. If you cannot produce those, the two weeks will be spent discovering the job, and two weeks is the wrong place to discover the job. An agency can narrow a fuzzy request. An agency cannot invent what the handbook was supposed to say and then be graded on the invention.

One user is enough for this shape. You, or a design partner, on a URL you can send. A second user with a different set of files is a different product. It is a normal next step. It is not a sentence you add to the two-week proposal on the Thursday before kickoff.

The four-to-six-week shape

The longer build is the version a second person can sign in to. Their documents are not the first person's documents. The check that enforces that is the same kind of check the rest of the product uses, and it runs before the model sees a passage. A prompt that says "only use this user's files" is a wish. The query either filters, or it does not.

This scope can include an evaluation set: the questions you care about, the answers you would accept, stored so a prompt change can be checked instead of remembered. It can include a way to see latency and whether the refusal is still firing. It can include more than one integration. Auth and the admin basics are part of why this is longer than a single flow. The product around the model is described in an LLM-powered SaaS MVP. Which of the two shapes you are in is a question for the call, after we have seen the files, not a guess you make from a menu.

Design in your brand belongs on the longer build, not on the short one. If the chat is going in front of a customer who will compare it with the rest of your company, the screen has to look like the company. If the chat is for you and an investor next Tuesday, a plain screen is the right cut.

Monitoring is not a souvenir. If nobody is depending on the answer yet, start short and learn whether anyone asks the question. If a support team will paste the answer into a customer thread, you need the version where you can see a bad day. Skipping the eval and discovering the bad day in the thread is how a short build becomes a second project. Book the call before you pick. We would rather tell you the short one is enough than let you commission a platform for a folder.

Access is the part the model will not do

Two accounts. Account A has the employee handbook. Account B has a single vendor contract. B asks for the vacation policy. The correct result is not a summary of A's handbook. The correct result is that B's corpus does not contain it. If the demo was indexed as one bucket for convenience, B just read a document B was not allowed to open, in complete sentences, with a citation.

This is ordinary application work. It feels slow next to a chat that already "works" on a laptop. The laptop worked because the laptop had every file. Production is the moment the laptop is no longer the user. The architecture for that boundary is the same class of problem as any other multi-tenant feature. Add AI to an existing SaaS is the note for when the product already has those rules and the chat must obey them. If the product does not have those rules yet, do not add a fluent way to skip them.

Write the rule in a sentence before the index exists. "A user may retrieve a passage only from documents their account is allowed to open." If you cannot write the sentence, you are not ready for a second user. You may still be ready for the two-week build, with one account and a folder you chose. Say that. The dangerous version is the sentence no one wrote, implemented as "search everything, the model will be careful."

Citations, tables, and the pages that are empty

A citation that points at a file name is half a citation. The user still has to hunt. Point at a page, and where the page is a scan, say so. Some pages are photographs. Some are tables that a text extract turned into soup. Some are covers and blank sheets. The first build should show its uncertainty on those pages instead of reading soup with authority.

A health pass on the corpus belongs in the work even when it is small. How many files opened. How many pages had no text. How many tables broke. You are not commissioning a document-intelligence lab. You are refusing to answer from garbage without telling the reader it was garbage. The document chat guide lists citations you can click and a report of scans, tables, and empty pages for that reason. A chat that hides the bad page will be blamed for an answer the page never contained.

Numbers are the usual failure. A table of prices, a dosage, a notice period. The model is fluent around them and wrong by a row. The citation is how a person catches the row. If your users will not click, the feature is a draft for a person who will, not a source you put in front of a customer alone. Decide that before launch. A draft with a human is a good two-week product. An unattended answer about a price is a production decision, and it needs the eval.

What we will not call a document chat

A rewrite of your knowledge base. A new wiki. A tagging program for ten years of drives. Those can be wise. They are not this feature, and a proposal that starts with them has changed the project. The chat can sit beside the mess and only index the folder you named. The mess can be a later scope. It should not be discovered as the critical path in week two.

A general company assistant that also books meetings, queries the warehouse, and files tickets. That is an agent with tools. The note for an agent that works in a demo and fails on customers is your AI agent is not ready. A document chat can be one tool that agent calls later. Building the agent because the chat sounded small is how the scope doubles without a second yes.

A promise that every answer is correct. A serious build includes evaluations so you can see when an answer is wrong. It does not include a guarantee that the first prompt is the last one. Anyone who sells the guarantee is selling a feeling. Ask them what happens on the question the corpus does not contain. If the answer is "it will use its knowledge," you are buying a search engine that lies when the search misses.

A week picture, not a platform roadmap

The first days are the corpus and the questions, not a model comparison. We look at the files you named. We try the questions you named. We write down what a good answer points at, and what a refusal looks like on the screen. If the files are not the files, we stop and say so. Pretending a dump is a corpus wastes the part of the calendar that was supposed to be the product.

The middle is the flow on a URL. Upload or select, ask, read, click the passage. One user. The refusal visible. The spend alert on. Your repo, not a preview you cannot leave. If your own team will change which document is in the set after we leave, they should see that change while we are still in the engagement. The questions for that handoff are in an AI product your team can maintain. A chat only the original builder can retarget is a rented answer.

The end of a two-week build is a decision you can make with a link: keep going toward a second user, stop because nobody asked the question, or quote the production scope now that the passage click is real. The end of a production build is the second account test still holding, and the examples still passing when the prompt moves. Widening access before that test is how a citation becomes a leak with a footnote.

Questions that change the scope

How many document types. One type, one schema of "a passage in a file," stays a short build if the user is singular. A handbook plus the signed contracts plus the customer's private uploads is three corpora with three rules. That is the longer build, or a separate conversation if the rules conflict. Bring the count to the call.

Who uploads. If only you upload, the flow is short. If every customer uploads, you need abuse limits, a size cap, and a way to delete. Deletion includes the index, not only the file icon. A file that is gone from the screen and still answered from the index is a bug you will hear about from the customer who asked you to delete it.

What the answer is allowed to do next. Show a passage: this feature. Create a task, send an email, or write back into the system of record: a second feature, with a person in the loop on the first release. The accept button stays yours. The model proposes a sentence. It does not file the sentence.

Whether the product already exists. An existing app with real permissions should be extended, not replaced, which is the point of the add-AI note. A new product whose whole job is this chat can still be a short first build. Do not buy a new platform because a generator preferred different nouns for the same upload.

What to bring to the call

The corpus, or a sample that is honest about the mess. The one user. Five questions with the page you would accept, and two questions that must fail. The name of the system if there is one system. The sentence about who may see a file, even if today the answer is "only me."

Leave the roadmap of every department's archive at home, or mark it as later. The first chat is the one a specific person will use next month. The others can be a second proposal after that person has clicked a citation and told you the page was the right page. If they will not click, you learned something cheap. That is what the short package is for.

If you already tried this in a builder, bring the URL. A generated chat that answers from every row is a brief. It is not evidence that the two-week package should also include the repair of that index. We can keep the screen and replace the retrieval. We should not keep a retrieval that has no user on it.

If this is your app, send the repo.

A 30-minute scope call. We read what is live, name what is dangerous, and quote a fixed scope.

How Techtribe builds document chat

We scope one corpus and one flow: open the document, ask, read an answer that points at a passage, and show not-found when the passage is missing. A single user on a tidy set is a short build. A second user, access checks before retrieval, and a way to see a bad answer are a longer one. A continuing program of archives is a team that stays. A dirty corpus is its own conversation, not a footnote on the chat.

The scope call is 30 minutes with a founder. That is where we say which of those you are in, or tell you the corpus is not ready. Bring the sample files, the questions, and the sentence about who may read them. The build guide is document chat. The service page is RAG development. If the chat has to obey an app you already run, start from add AI to an existing SaaS.

Questions

Can the first version be chat over one set of documents?

Yes. One corpus, one signed-in user, and an answer that cites a page is a complete first build. A second corpus, several permission levels, and a monitored refusal are a longer scope, not a tweak to the same two weeks.

Will the chat invent an answer when the document is silent?

Not if the product is built to refuse. A missing passage should come back as not found, with nothing presented as a policy. A prompt that says 'be honest' is not that behavior. The refusal has to be a result the interface can show.

Who is allowed to open a document through the chat?

The same person who could open it in the product without the chat. The model does not get a wider view than the user. If two accounts should see different files, that test is part of the longer scope, and it is run before anyone else is invited.

How do we find out whether this chat is the right size?

Book the scope call. Thirty minutes with a founder. Bring the sample files and the questions. We will say whether this is one user and one corpus, or a second user with access checks, and we will say if the archive is too messy to pretend it is a tidy folder.

Book a free call

Have a vibe-coded app in production?

30-minute free call. We'll read what you have, name what is dangerous, and quote a fixed scope.