How document-based AI answers work
Understand retrieval, generated answers and source references.
Your documents supply the knowledge
RAG stands for retrieval-augmented generation. It connects an AI assistant to your approved content by finding relevant passages for each question and giving that text to the model as context.
Bulbul uses documents only. You supply Markdown or text files, review their extracted sections and approve what visitors may use. The AI generates answers from retrieved text; owners do not enter manual question-and-answer pairs.
Search by meaning, then generate
Approved sections and visitor questions are turned into embeddings, which let Bulbul search by meaning instead of exact words. It finds the most relevant sections for that website and passes their text to the AI model.
For example, a visitor might ask ‘Can I try it before paying?’ while your document describes a ‘14-day free trial’. Semantic retrieval can connect the wording. The model composes an answer from the retrieved passage and returns supporting source references.
Uploading content does not permanently train the model. The text remains knowledge that Bulbul retrieves when answering.
Keep the answer within the evidence
The answer flow is designed to use approved knowledge, ask for clarification when needed and offer support when facts are missing. References help you review the evidence; they do not replace checking whether the answer is correct.
Test direct questions, different wording, exact product names and questions the documents cannot answer before using the widget with customers.
Demo versus your assistant
The public demo is Bulbul’s own website assistant, answering live questions about its features, setup and pricing with sources from its public product guide. The homepage animation uses excerpts from that guide. Your own assistant answers from the documents you approve for your website.
Your assistant remembers the last few messages so follow-up questions make sense, but every fact still has to come from your current approved documents. Test it in the Playground before you publish.