Not the internet, not the model's imagination — your PDFs, policies, and product docs, retrieved passage by passage and cited inline.
When you upload a document, its text is split into overlapping passages and each passage gets an embedding — a mathematical fingerprint of its meaning. When a customer asks a question, the question gets the same treatment, and the agent retrieves the handful of passages whose meaning genuinely matches — then writes its answer from those passages, citing them inline. If nothing in your documents matches well enough, the agent has no sources and says so, plainly, instead of inventing an answer. That honest failure can also be a trigger: flip one switch and "not in the knowledge base" escalates the conversation to your team.
Website crawling isn't supported yet — export the pages that matter to PDF or Markdown and upload those. You'll know exactly what the agent knows.
text-embedding-3-small, computed once at upload.The agent is explicitly instructed to answer only from retrieved sources, to cite them, and — when the answer genuinely isn't there — to say so without citations rather than bluff. Combined with the knowledge-not-found escalation trigger, the failure mode of a hard question is a human with a transcript, not a confident wrong answer.
PDF, Word (.docx), plain text, Markdown and CSV are the primary formats — JSON, HTML and XML text files work too. Each document can hold up to 200,000 characters of extracted text.
Not yet — today the knowledge base is built from uploaded documents. Exporting key pages to PDF or Markdown and uploading those takes minutes and gives you explicit control over what the agent knows.
Re-uploading a document replaces its chunks in one atomic step — the next question is answered from the new content. There is no cache of old answers.
Retrieval only surfaces passages that are genuinely similar to the question; below the similarity bar, the agent has no sources and is instructed to say so rather than improvise. You can also make “answer not found in the knowledge base” automatically escalate to your team.
The document’s chunks and their embeddings are deleted with it, immediately — deletion cascades at the database level.
10 minutes from sign-up to your first answered customer message. Free forever plan, no card required to build.