An AI chatbot that answers from your own documents, with sources.
Your documents are split into passages, indexed for semantic search, and searched for each question. The LLM answers only from what it finds and links back to the source, so people can check it. See the full pipeline in the diagram below.
For most company chatbots, RAG is the right choice: your documents change, people need sources, and nothing has to be retrained when you add a file. Fine-tuning is for teaching a model a format or tone, not for storing facts. More in RAG or fine-tuning?
Before launch we collect real questions from your team or customers and use them as a test set. Changes to prompts, models or documents are checked against it, so quality goes up rather than drifting.
Proof
Real products I built for real clients: the problem, what I built, and the stack.
RAG and AI chatbots
A UK contract compliance and invoice assurance platform. AI extracts contract clauses, people confirm them, and every invoice is scored against contract and PO.
RAG and AI chatbots
An AI analytics chat for e-commerce, built with Next.js, Claude Code, Supabase, pgvector and GPT-4. Ask about any Shopify store, get answers with charts.
How it's built
Your documents
PDFs, docs, wikis, help centre
Chunk and embed
Passages turned into vectors
Vector index
Semantic search over passages
Retrieve
The best passages per question
LLM
Answers only from what was found
GPTCited answer
Links back to the source
Why this stack
An evaluation set of real questions is re-run on every change.
Typical starting points. The final stack is chosen per project and written into your scope.
RAG and AI chatbots2 min read
Most assistants that answer from company documents need retrieval, not a trained model. The difference in plain words, and how to tell which you need.
Every answer is grounded in retrieved passages and shows its sources, and we test against a set of your real questions. When the documents don't contain the answer, it says so.
Usually not. A RAG chatbot reads your documents at question time, so there is no model training and documents can be added or removed at any time.
Yes. The same assistant can serve your team internally or your customers on your website, with different document sets for each.
It's chosen per project based on answer quality, cost and your data rules, typically Claude (Anthropic) or GPT (OpenAI).
A 30-minute call is enough to tell whether it fits, what the first version should include, and how long it takes.
Prayag Dalal · Top Rated on Upwork, 34 jobs completed
