RAG chatbot development
RAG knowledge bot development, with evals you can check
We build a knowledge bot that answers staff or customer questions from your own documents, with a source link on every answer. You also get the test set and an eval report you can rerun, so you see the quality instead of trusting a demo.
From $4,000 · 10 working days
Built by Narek Maloyan and Tigran, a founder-led team in Yerevan, Armenia. Narek built the agent that reached #1 on SWE-bench Verified (70%, May 2025), and has 7+ years of production machine learning and a PhD. You talk to the engineers who build your project.
Why most document chatbots disappoint
A chatbot over your documents is easy to demo and hard to trust. It answers the five questions from the sales call well, then invents a policy that does not exist. Nobody measured it, so nobody knows how often it is wrong. We fix that by writing the test first and building to pass it.
The package: $4,000, 10 working days
One knowledge bot on one agreed document set. The standard package is $4,000. Bigger scopes, such as several bots or many sources, get a fixed quote before work starts. Half at kickoff, half when the acceptance tests pass.
What you get
- Ingestion of your agreed sources, with updates when documents change
- Answers with a link to the source passage
- "I don't know" when the documents do not cover the question
- A chat in Teams, Slack, or your website, whichever you use
- A test set of 30 to 50 real questions, written with your team
- An eval report you can rerun, plus a handover session
Acceptance tests
- At least 90% of the test questions answered correctly, each with a source link
- Every out-of-scope test question gets "I don't know", never an invented answer
- Where your sources have permissions, a user never gets content they cannot open
- Your team can rerun the eval without us
14-day warranty. For 14 days after acceptance, we fix for free any case where the bot fails an acceptance test on the same kind of question it passed on.
How the evals work
- We collect 30 to 50 real questions from your team, with the answer and the source document for each.
- We add questions the documents do not answer, to test that the bot refuses instead of guessing.
- Every answer is scored: correct or not, right source or not, refused when it should.
- You get the report and the script. Rerun it after a new model, new documents, or a prompt change.
Ten days, step by step
- Day 0, scoping call. Free, 30 minutes. We agree the sources, the users, and the acceptance tests in writing.
- Days 1 to 3. Ingestion and the test set. You can ask the first questions on day 3.
- Days 4 to 9. We tune retrieval and answers against the eval, with a short written update every day.
- Day 10, acceptance. We run the tests together. You pay the second half when they pass.
Questions
What is a RAG knowledge bot?
RAG means retrieval-augmented generation. The bot first finds the right passages in your documents, then writes an answer from those passages only, with a link to the source. It answers from your content, not from what the model happens to remember.
What is an eval, and why do I need one?
An eval is a fixed set of real questions with the expected answers, scored automatically. It tells you how often the bot is right, how often it cites the correct source, and whether it says "I don't know" when the answer is not in your documents. You can rerun it after every change, so quality never drifts in silence.
Which documents can it use?
Common sources are SharePoint, Google Drive, Confluence, Notion, PDFs, help center articles, and your website. We agree the sources on the scoping call.
Where does it run, and is our data used for training?
It runs in your cloud account or your Microsoft 365 tenant, with the model provider you choose. We do not train models on your data.
What if we need more than one bot or a very large document set?
The package covers one bot on one agreed document set. Bigger scopes get a fixed quote before any work starts.
Related: a knowledge agent inside Microsoft 365 with Copilot Studio, MCP servers, and how we handle your data.
Bring 20 questions your team asks every week
Book a scoping call. We will tell you if a knowledge bot can answer them from your documents, and how we will prove it.