A chatbot can only give a useful answer when it has relevant information and enough context to use it. This guide helps you distinguish a training problem, an AI response problem and a usage limit.
Start by selecting your chatbot and reviewing the example conversation in Chatlogs. Keep the visitor's exact question so you can repeat it after making a change.
Incomplete or conflicting answers
The chatbot retrieves relevant passages from your training data rather than reading every source for every answer. Its context window must also accommodate instructions and conversation history.
- Open Training and check that the source containing the answer completed training.
- Review the extracted content. A successful scan does not guarantee that every part of a page was readable.
- Remove or update conflicting versions of the same information.
- Ask a specific test question in Overview.
- If useful information still does not fit, review Settings > Model & Advanced and the available context sizes.
The available context settings are 8k, 16k and 32k, subject to your account's features. A larger context consumes more credits and does not repair missing or incorrect source material. See Training basics and limits and Extending chat context size.
Files and pages that cannot be read
The file uploader under Training > Files accepts up to 20 files at once and up to 20,000,000 bytes per file. Files are submitted individually; the server's 50 MB request limit is not a combined limit for the selected batch. Supported formats are PDF, TXT, MD, JSON, DOC, DOCX, CSV, XLS and XLSX. File size and the amount of extracted training text are different limits.
Common extraction problems include:
- Image-only content: text inside pictures or scanned pages may not be available to the text extractor. Supply a text-based version and check the imported result.
- Private pages: a normal website scan cannot sign in to a customer account or read private pages on your behalf.
- Dynamic or protected pages: JavaScript content and anti-bot measures can affect what is extracted. Check the result instead of assuming either complete support or complete failure.
- Audio and video: provide a text transcript when the information is only present in a recording.
For a failed source, use Adding plain text, Adding files or Adding questions and answers. Re-test the relevant question after training completes.
Training takes longer than expected
Duration depends on the number of sources, their size and the source website's response. Check the status under Training, including sources waiting for your selection or approval. You can leave the page while background processing continues.
When training finishes, test the chatbot in Overview. If a source fails, inspect that source and retry it after correcting the issue rather than repeatedly starting a full scan. See Training in the background.
Plausible but incorrect answers
AI can produce a confident answer that is wrong. A lower creativity setting or a restrictive prompt can help, but neither guarantees accuracy.
- In Settings > Role & Behavior, define the chatbot's scope and what it should do when information is missing, such as admitting uncertainty or offering a human contact option.
- In Settings > Model & Advanced, lower Creativity where the selected model supports it.
- Keep factual training sources current and remove contradictions.
- Add a correction for a specific bad response.
- Re-test that question and related questions in a new conversation.
See Improving chatbot responses for a more detailed workflow.
Questions about the entire knowledge base
A chat answer is not a reliable inventory of every file, page or topic you trained. The chatbot works with the information retrieved for the question, so asking it to list everything it knows may omit sources.
Use the Training tab to inspect the source inventory. Test representative questions to check coverage rather than relying on the chatbot's description of its own knowledge.
Rate limits and exhausted credits
Rate limits restrict how quickly visitors can send messages. Configure them under Settings > Security and check the displayed limit message when a visitor is blocked.
Message credits are a separate monthly usage allowance. They reset on the first day of the calendar month. Check Subscription and any per-chatbot limit when responses stop because of exhausted credits. Review the capacity options available to your account or wait for the reset.
Temporary service errors
Website sources, AI services and integrations can be temporarily unavailable. Check whether the problem affects one source or action, one chatbot, or every test. For an integration failure, also check the saved connection and the external account's permissions.
If the issue persists, use Support Request and include the bot ID, approximate time, exact question or failed step, and any error message. Do not include passwords or API keys.