Custom AI Assistant on Your Company Data
An AI assistant that answers from your own documents and shows where each answer came from. Built with access control, testing and a routine that keeps it current.
A custom AI chatbot for business that answers from your own documents
A custom AI chatbot for business, built as retrieval-augmented generation (RAG), answers by first searching your documents and then writing a reply from what it found, with a link back to the source. We build that whole system: document ingestion, search, the chat interface, permissions, a test set and a refresh routine so answers stay current. We do not pre-train or fine-tune a model on your data. That is a deliberate choice, not a gap: for company knowledge, retrieval is easier to update, cheaper to run, and lets a reader check every answer against the page it came from.
Everything you need, nothing you don't.
Document ingestion for PDFs, Word files, web pages, help articles and shared drives, cleaned and split into passages that keep titles, headings, page numbers and file names, so every answer can point to its exact source.
A search layer that mixes keyword matching with meaning-based (vector) search and re-ranking, tuned on your documents and your wording rather than left at default settings.
A chat interface placed where people already work: your website for customers, or an internal page or Slack for staff, with sources shown beside each answer and a way to open the original file.
Cited answers with a refusal rule: when the documents do not contain the answer, the assistant says so and offers a contact, instead of filling the gap with a plausible guess.
Access control at folder or document level, so a person only receives answers drawn from files they are allowed to open, plus separate public and internal assistants where the audiences differ.
An evaluation set of real questions with agreed correct answers and source pages. We re-run it after each change, report the results to you, and hand it over so you can keep testing.
A refresh pipeline that re-indexes documents when they change, removes deleted or superseded files from answers, and alerts a named person if a sync fails or a source stops updating.
Question logs and review tools: ratings, unanswered questions and the documents most often cited, which show what is missing from your knowledge base as well as how the assistant is performing.
Our Process
First we list your sources and the twenty or so questions people ask most, and agree who may see what. We index a small slice of documents and test search before building anything around it, because poor retrieval cannot be rescued by a better prompt. Then we write the evaluation questions with you, tune chunking, search and prompts against them, and add permissions and the chat interface. A pilot group tries it on real work while we read the logs. Only after the test results and the pilot feedback look right do we open it to everyone. Scope and price are agreed on a free call.
Technologies We Work With
Claude or OpenAI language models for writing answers, an embeddings model for search, and a vector store such as pgvector in PostgreSQL or a managed service, chosen to match your hosting. Laravel or Node.js for the application and permissions, n8n for scheduled ingestion from Google Drive, Notion, SharePoint or your website, and Slack or an embedded widget for the front end.
What we build, and what we do not
Being specific about scope avoids most disappointments in this kind of project.
What we do
- Connect your documents to a language model through search, so replies are written from your own material.
- Show the source passage and file for each answer.
- Enforce who can see which documents.
- Build a test set and report accuracy against it, before launch and after every significant change.
- Set up the refresh routine and hand over documentation your team can follow.
What we do not do
- We do not pre-train a model from scratch, and we do not promise that fine-tuning would make it know your facts. Fine-tuning changes style and behaviour; it is a poor way to store facts that change. Our comparison of RAG vs fine-tuning for business covers the reasoning.
- We do not claim zero errors. Language models can still misread a passage.
- We do not clean up decades of contradictory documents for you, though we will show you where they conflict.
Custom GPT or a RAG build: which one do you need?
A custom GPT is a configured assistant inside ChatGPT. A RAG build is an application you own. Both can answer from files, so the difference lies in control and scale.
| Factor | Custom GPT | RAG assistant (custom build) |
|---|---|---|
| Who uses it | People with ChatGPT accounts | Anyone you choose: customers on your site, staff in Slack, or an internal page |
| Document volume | Limited number and size of uploaded files; see OpenAI's current help page | Sized to your document set, with search built for it |
| Citations | Varies by setup and cannot be customised deeply | Passage-level citations designed in |
| Permissions | Everyone with access sees the same files | Per-user filtering of sources |
| Staying current | Re-upload files by hand | Automated sync |
| Testing | Informal | A repeatable evaluation set |
| Effort and cost | Low; an afternoon for a small set | A real project with running costs |
If your material is a few dozen files, only your own team uses it, and a wrong answer is cheap, try a custom GPT first. It costs you almost nothing and shows what is possible. Come to us when it stops being enough.
How is one answer produced?
- The user asks a question. The system first checks who they are.
- It searches the index, keeping only passages from documents that person may open.
- The best passages go to the language model with instructions to answer only from them.
- The model writes the reply and marks which passages it used.
- The interface shows the answer with links to the sources. If nothing relevant was found, it says so.
How do you know the answers are good?
We write the test before the assistant. Together we collect real questions from staff or customers, note the correct answer and the page it lives on, and add some questions the documents cannot answer. Each version is scored on whether the right passage was retrieved, whether the answer is correct and supported, and whether it refuses when it should. Scores guide the work, and you see them. They are not a guarantee about every future question, which is why logs matter after launch.
Where does your data go? Privacy options
Three setups are common, and we help you choose by sensitivity and budget.
| Option | How it works | Trade-off |
|---|---|---|
| Provider API with standard terms | Passages are sent to a model API for each answer | Simplest and cheapest. OpenAI states it keeps API abuse-monitoring logs for up to 30 days by default and does not train on API data unless you opt in; Anthropic states it does not train on commercial inputs or outputs by default |
| Provider API with tighter controls | Regional processing or zero data retention where the provider offers it | Needs provider approval for some controls and can cost more; not every feature is eligible |
| Self-hosted open model | The model runs on servers you control | Data stays inside your environment, but you carry the hosting cost, maintenance and usually lower answer quality |
Whatever the option, index and logs live in infrastructure you own or approve, and personal data is kept to what the task needs. Provider terms last checked November 2026 on OpenAI's data controls page and Anthropic's privacy centre. Please read the current versions and have your own adviser review anything confidential or regulated.
What do you need to provide?
- Read access to the documents, or an export of them, and a person who knows which ones are authoritative.
- A list of who should see what, even if it is only staff versus public.
- Someone to supply real questions and judge the answers during testing. Their time matters more than any tool we choose.
- A decision on what the assistant should do when it does not know: show a phone number, open a ticket or send an email.
What happens after launch?
An assistant that was right in March can be wrong in June, because prices, policies and procedures change. The refresh pipeline handles the mechanical part, but a person still needs to own the content. We set up a short monthly review of the unanswered-question list and the lowest-rated answers, so you can see which document to write or correct. Often the fix is a clearer paragraph in a source file rather than a change to the assistant. If you would rather not run that review yourself, we can discuss ongoing support when we scope the work.
Who should not buy this?
- Teams with a short list of FAQs. A simple site chatbot, such as our AI chatbot integration, is quicker and cheaper.
- Teams whose documents are out of date or contradict each other. Fix that first or the assistant will repeat it with confidence.
- Anyone who needs the assistant to take actions, such as issuing refunds. That is a different, riskier design.
- Anyone looking for a guarantee of perfect accuracy.
Not sure whether your goal needs retrieval, fine-tuning or just a better prompt? Start with how to train ChatGPT on your own data, then bring your document list to a call.
Client feedback
What clients say
“Usama made an old first time app builder very confident with his professional approach, attention to details and a complete understanding of what I was trying to convey. Usama exceeded my expectations on our first meet. Would highly recommend Usama and his team.”
“Great experience working with Usama. He delivered a well-structured, conversion-focused product page and was very responsive to feedback. The final result is much more professional and aligned with improving sales performance. Communication was smooth, and he implemented all requested adjustments quickly. Would definitely recommend and work with him again.”
“Usama went above and beyond to help me, especially as I changed directions quite a few times, however his professionalism throughout made the whole experience, for an inexperienced person, very easy and informative and enjoyable. I would highly recommend Usama and his team.”
“Great job as always.”
“Very nice working with Usama. Will continue using his services.”
“It was a pleasure working with Usama. I am very satisfied with the communication, the final results and the expertise in this field. Looking forward to future collaborations.”
Real reviews from real clients — verify these on Fiverr →
How pricing works
We quote every project individually, because the price depends on scope: how many pages, systems or integrations are involved, how much is built from scratch, and how fast you need it. Small, well-defined jobs cost far less than custom builds. Tell us what you need, or use the quote request guide, and we will reply with a written quotation.
Frequently Asked Questions
What is a RAG chatbot for company documents?
It is a chatbot that looks things up before it answers. When someone asks a question, the system searches your documents for the most relevant passages, gives them to a language model, and the model writes a reply based only on that material, with links to the sources. RAG stands for retrieval-augmented generation.
Can you train ChatGPT on our company data?
Not in the sense of changing the model itself, and we would not recommend it for company knowledge. What people usually want is an assistant that answers from their documents, which retrieval does better: updates take minutes, and answers carry citations. Our guide to training ChatGPT on your own data explains the three routes in detail.
Will our documents be used to train AI models?
That depends on the provider and the plan you use. As of November 2026, OpenAI states that API data is not used to train its models unless you opt in, and Anthropic states the same for its commercial products including the API. Terms change, so we check them with you before launch and can recommend stricter options for sensitive material.
How do you stop the assistant making things up?
We reduce it, but we cannot promise it never happens. The assistant is limited to retrieved passages, must cite them, and must say when nothing relevant was found. We then measure with the evaluation set, which includes questions the documents cannot answer, and we read real logs after launch to catch what slips through.
Can different staff see different documents?
Yes. We store each document's permissions alongside its passages and filter search results by the person asking, before the model sees anything. A staff member never receives an answer built from a file they cannot open. Permissions are synced from your source system where it exposes them, or set by folder.
How does the assistant stay up to date when documents change?
A scheduled or event-driven sync notices new, edited and deleted files, re-indexes only what changed and removes what is gone. Each answer shows the document version or date it used. If a sync breaks, a named person is alerted rather than the assistant quietly serving stale information.
Guides related to Custom AI Assistant on Your Company Data
Custom GPT vs RAG Chatbot vs AI Agent: A Buyer's Guide to Cost and Ownership
Choosing between a Custom GPT, a RAG chatbot and an AI agent is a buying decision. Compare build effort, who maintains it, where your data goes and wh...
RAG vs Fine-Tuning for Business: Which One Should You Start With?
Start with RAG when the model lacks your facts, and consider fine-tuning only for a measured behaviour problem. Decision table, worked scenarios and e...
How to Train ChatGPT on Your Own Data: Projects, RAG or Fine-Tuning?
Most businesses cannot truly train ChatGPT, and do not need to. Here are the three real options, what each costs in effort, what never to upload, and...
Ready to start your custom ai assistant on your company data project?
Free consultation, no pressure — just a clear plan for what it'll take.
Get a Free Consultation