Large language models write excellent text but know nothing about your company: prices, terms, internal procedures. Plug a model straight into a chat and it will answer confidently — and sometimes wrongly. RAG (retrieval-augmented generation) solves this: the assistant first finds the relevant passages in your documents and then answers only on their basis. It is one of the core scenarios in our AI platforms service.
In short: about 80% of a RAG assistant's quality comes from the knowledge base and the test question set. The assistant should cite sources, honestly say "I don't know" and hand complex cases to a person.
How RAG works
- Preparation: documents are split into chunks and indexed for semantic search.
- Retrieval: for each question the system finds the most relevant chunks.
- Answer: the model answers only from the retrieved chunks and cites the source.
- Control: if no suitable chunks are found, the assistant says so and offers to connect a staff member.
Where a RAG assistant adds value
| Scenario | Example |
|---|---|
| Customer support | Answers on terms, pricing, timelines, documents |
| Sales | Matching a service to the client's task, booking a consultation |
| Internal knowledge base | Answers to staff on procedures and instructions |
| Onboarding | An assistant for new employees |
| Compliance | Searching policies and requirements with clause references |
What to prepare
- Current documents: FAQ, prices, terms, procedures — with no contradictions between versions.
- A test set of 50–200 real questions with reference answers.
- Rules: what the assistant does not discuss, when it hands over, what tone it uses.
- Access rights: which documents customers see and which only staff see.
How to control quality
- Re-run the test set after any change to the knowledge base or model.
- Conversation logs with answer ratings and error reviews.
- Metrics: share of answers with a source, share of human handovers, user satisfaction.
Frequently asked questions
Is it safe to upload company documents?
With the right architecture, yes: access control, storage in controlled infrastructure and agreements with the model provider not to train on your data.
Can the assistant work on the website, Telegram and phone?
Yes. One assistant can work in a website chat, a Telegram bot and by voice, logging leads in your CRM.
How much does it cost?
Our AI projects start from $1,500. The total depends on the knowledge base size, channels, integrations and security requirements.
Want an assistant that answers from your documents and does not make things up? Send a request on the AI platforms page and we will assess the task and data, build a test set and launch a pilot.