add_knowledge_text
ChatGPTAdd free text to a chatbot's knowledge — policies, opening hours, a product description, anything that isn't on a web page. Embedded in the background. Give it a title: the title becomes a heading on the text, which makes retrieval noticeably better than an untitled wall of prose. Demo chatbots accept up to 50000 characters of training content in total.
add_knowledge_url
ChatGPTAdd one web page to a chatbot's knowledge. The page is fetched, extracted, and embedded in the background — no need to call anything else afterwards. Use this for a page the original crawl missed, or a page on a different site (a docs host, a help centre, a pricing page). Demo chatbots accept up to 5 pages in total.
add_qa
ChatGPTAdd an exact question-and-answer pair to a chatbot. Use this when the answer must be worded a particular way — pricing, refunds, guarantees, anything where paraphrasing from a page would be risky. question_variations lets you list other phrasings of the same question so retrieval catches them all. Demo chatbots accept up to 50000 characters of training content in total.
apply_improvement
ChatGPTAdd the exact information supplied by the user for this recommendation and queue training. Show the exact question and answer to the user and obtain explicit confirmation first. Never invent facts.
create_chatbot
ChatGPTCreate an AI chatbot trained on a website and return links to try it, embed it, and keep it. The chatbot is a free demo: it crawls up to 5 pages of the site, answers 10 messages, and expires after 8 days unless somebody claims it. Claiming is a human step — open the returned claim_url, sign up or log in, and the chatbot moves into that account permanently. Training runs in the background and usually takes a minute or two. The links work immediately; answers improve once training finishes. Call get_chatbot with the returned token to check progress. No account is needed for this. If the person connects their ChatbotGen account, this creates the chatbot directly in it instead, with no claim step.
diagnose_question
ChatGPTTest a question and assess the answer using a bounded sample of evidence, including possible retrieval misses or conflicting sources. Consumes one message; external actions are disabled. Evidence and model assessments require human review.
get_chatbot
ChatGPTLook up a chatbot created with create_chatbot: how far training has got, how many of its free messages are used, and the same chat, embed and claim links. Pass the token from the create_chatbot response (the chatbot_id works too). Once somebody claims the chatbot it belongs to their private account and is no longer visible here.
get_conversation
ChatGPTRead messages and the sources stored with each answer to diagnose a customer conversation. Account connection required. Sources are retrieval evidence, not a model reasoning trace. Paginated, 20 messages per page.
get_improvement
ChatGPTRead one recommendation and its evidence before proposing a change.
get_integration_guide
ChatGPTStep-by-step guide to connect the chatbot to the owner's systems for a goal: order status, returns/refunds, account lookups, helpdesk tickets, a remote MCP server, webhooks, Zapier/Make/n8n, or signed-in customer identity. Returns what to ask the owner, the steps, a ready action draft and the current setup. Call it before proposing any integration; omit goal to list them.
get_knowledge
ChatGPTRead a source before correcting it. Content is paginated in 12000-character pages.
get_knowledge_summary
ChatGPTSummarize stored source counts, source health issues and a bounded selection of labels and processing states. These observations do not establish topic coverage or answer correctness.
get_whatsapp_test
ChatGPTGet the ChatbotGen test number, a WhatsApp link with this chatbot's code prefilled, and a QR code. The QR opens a chat with our shared test number; it does not connect the customer's own number. The first message joins this chatbot, then the person can send questions from their phone. Use after selecting a chatbot and offer it as a way to try the bot. The sandbox may be unavailable; never invent a test number or code.
list_chatbots
ChatGPTList every chatbot in your ChatbotGen account, with its status, message count, and links — plus how many more your plan allows. Start here when the person says "my chatbot" without saying which one.
list_improvement_suggestions
ChatGPTList persisted quality recommendations with their evidence and status, 20 per page.
list_knowledge
ChatGPTList knowledge sources with IDs, status and excerpts. Paginated, 20 per page.
refresh_knowledge
ChatGPTRecrawl an existing URL in place and retrain. Does not create a duplicate page.
resolve_improvement
ChatGPTAfter showing the latest answer, record the owner's explicit confirmation that it is good, or dismiss the recommendation. A confirmed answer is the owner's judgment, not independent verification.
review_conversations
ChatGPTRead recent customer conversation summaries to find unanswered questions and poor answers. Account connection required. Excludes plugin tests; page through the sample before claiming comprehensive coverage.
search_knowledge
ChatGPTRetrieve trained evidence for a topic using the chatbot's actual retrieval. Matches are not proof of complete coverage; test answers too.
test_chatbot
ChatGPTTest a real answer using the bot's model, instructions and retrieval. Consumes one message. Pass the returned conversation_id for follow-ups. Test conversations are isolated from customer events. External tools, bookings and handoff actions are not executed.
update_chatbot
ChatGPTReplace chatbot instructions to adjust tone, boundaries and behavior. Read current instructions with get_chatbot and preserve unrelated rules.
update_knowledge
ChatGPTReplace text or Q&A and retrain, or set excluded on a URL to disable/re-enable it as evidence. Read the source first and preserve unrelated facts. Use refresh_knowledge to recrawl a changed page. Files must be replaced in the dashboard.
verify_improvement
ChatGPTRepeat the original question after training, consuming one message. Returns before and after evidence for human review; retrieved sources do not prove the answer is correct.
