RAG Chatbot Developer — An AI Assistant That Answers From Your Own Documents

RAG over your own data is my AI Integration Professional package: $2,499, delivered in three to five weeks. It covers retrieval over your documents, help centre or database, up to three automated workflows such as ticket triage or reply drafting, human review steps, up to two integrations with tools such as Slack, HubSpot, Zendesk, Notion or Google Workspace, an evaluation set with quality monitoring, an LLM cost dashboard, guardrails with PII handling, and two weeks of support after launch.

If you only need a simple assistant inside your product that does not read private documents, the Starter package is $799 over one to two weeks. If the assistant has to serve more than one team, with an internal gateway for budgets and audit, private model access and permission-aware retrieval, the Enterprise package is $6,999 over six to ten weeks. And if the chatbot is the product you plan to sell, with tenants, plans and metered usage, that is an AI Agent SaaS build from $4,999 instead.

Running costs are separate and I estimate them in the proposal: model API fees, a vector-capable database and hosting. For most internal and support chatbots these are small next to the build, and cost caps are part of every package so a busy month cannot surprise you.

Ingestion comes first: pulling in PDFs, web pages, help-centre articles or database rows, cleaning them, splitting them into passages that make sense on their own, and embedding them into pgvector on PostgreSQL. Retrieval is hybrid, combining keyword and semantic search with reranking, because pure vector search misses exact terms like product codes and policy numbers. Where users have different permissions, retrieval respects them, so nobody gets an answer built from a document they are not allowed to open.

Then comes the part most demos skip. In week one we write an evaluation set together: real questions with the answers you expect. Every change to prompts, chunking or models is scored against it, so you can see whether the assistant got better or worse instead of guessing. Every run is traced, so a wrong answer can be replayed and fixed at its cause.

Be honest with yourself about this one, because a custom build is not always the answer. If you need a widget on a marketing site that answers from a few dozen pages, a hosted builder such as Chatbase will get you there in an afternoon for a monthly fee, and I will tell you so on the first call.

A custom build earns its price when one of four things is true: the content is private and must stay in your own infrastructure; different users must see different answers; the assistant has to act inside your systems, such as opening a ticket or updating a record; or the monthly fees and per-message limits of a hosted tool have started to cost more than owning it. With a custom build the code sits in your GitHub organisation, the data in your database, and the model-provider accounts in your name.

Virtual Client is a multi-tenant voice coaching platform where teams rehearse conversations against AI personas and receive rubric-scored feedback from OpenAI and Anthropic models. It runs in production across several organisations with roughly 2,500 practice sessions delivered, and the hard problem was the one RAG shares: keeping each organisation's data strictly separate. Contently AI is a social content SaaS that puts OpenAI, Anthropic and Groq behind one interface, which is the same routing pattern I use to keep a chatbot's cost and quality under control.

Founders is a platform where an AI grades structured product reviews from 0 to 100 on Groq, so scoring quality had to be measured, not assumed. The assistant on this website also answers from my live services, projects and articles through retrieval rather than from the model's memory. Ask it something specific and check the answer against the page.

I work remotely from Bangladesh, UTC+6 all year. Riyadh is three hours behind me, so the whole Saudi working day overlaps mine, and both countries work Sunday to Thursday, which means no lost days at either end of the week. London is five hours behind in summer and six in winter, so I am online for the full UK afternoon. For the US I cover East Coast mornings and evenings, with a fixed weekly call that stays in the same slot.

You sign a short contract with your company name on it: scope, milestones, price, confidentiality and full assignment of intellectual property to you on payment. I will sign your NDA instead if you prefer. Payment is by milestone in USD through Wise, Payoneer, international bank transfer, or a Stripe invoice if your finance team needs to pay by card. The usual split is 40% to start, 40% at the mid-point demo and 20% at hand-off.