Train AI on your knowledge
RAG and chunking.
Company documents become meaningful chunks, and the chatbot answers only from those chunks. Every reply names the document and section, so your team can match it to the original.
From a document to a customer reply
Knowledge sources
Training starts from the files your team already uses: instruction PDFs, site pages, price spreadsheets, product catalogs, FAQs, and closed support tickets. Each source is stored with a date and an owner, so a later reply can be traced to a file.
Duplicates, outdated versions, and files with no readable text are left out. When a price or a rule changes, only that document is updated. The model does not have to be rebuilt.
Chunking
A long document is split into pieces that each carry one complete topic. The split follows headings, paragraphs, and tables. Neighboring chunks overlap, so an answer that sits on a boundary is still found in search.
Chunk size follows the document. FAQs stay short. Contracts and technical manuals stay longer. The model reads these chunks on each question, not the whole file.
Embeddings
Each chunk becomes a numeric vector that holds the meaning of the sentence. Search is no longer limited to the exact word. A question about a warranty period can match a passage that says “eighteen-month guarantee” even when the word warranty is absent.
The embedding model is chosen to read Persian properly. Product names, SKUs, and numbers stay in the chunk text, so semantic search and exact lookup work together.
Vector database
The vectors live in a vector database. Each question returns the closest passages by meaning: a handful of relevant chunks, not the whole archive. Retrieval can be limited to one product, one language, or one document set, so a sales answer stays separate from a support answer.
When a document changes, only its chunks are rebuilt and replaced. The index is sized for catalogs and ticket archives, and it keeps the lookup fast.
Writing the answer
The language model reads the retrieved chunks and writes the reply in the customer’s language. If those chunks do not contain an answer, it does not guess. It says the topic is not in the current sources, or it hands the chat to a teammate.
GPT, Claude, Gemini, or an open model is chosen by how hard the question is. A simple question can use a lighter model. A technical one uses a stronger model. The customer sees one reply. Routing happens underneath.
The reply to the customer
The same answer, with the document name and section, goes out on the site widget, WhatsApp, Telegram, Bale, Rubika, and Eitaa. The customer sees a normal conversation. Your team checks the source in the panel.
When the model’s confidence is low, the chat reaches an operator with a summary and the passages that were found. The operator does not start from an empty thread.
Alongside the training
One API for every model
ChatGPT, Claude, Gemini, DeepSeek, and open models share one connection. Switching models does not mean a new contract with each lab.
Routing and cost control
Each question goes to a model that fits its difficulty and price. If one provider is down, the next model answers, and token use shows up in the panel.
Cited, and on your servers
Answers are built only from your sources, with the document attached. On the enterprise plan the model can run on your own servers, so the data stays inside the company.
How it happens
Collect the sources
We take the files, the site address, and a sample of tickets, and set aside duplicates and outdated copies.
Build the knowledge base
The text is chunked, embedded, and stored in the vector database. We test it with a set of real questions.
Publish on the channels
The same knowledge connects to the site widget and the messengers. Your team sees the source of each reply and can take over the chat.