OrhanAI supports 11 direct training source types. Here is what each one is good for, how crawling and indexing work, and how to tell when a source is missing.
Let AI handle the repetitive questions. Free your team to focus on real conversations.
Get Started FreeSetup in minutes · No credit card required
A chatbot is only as good as what it was trained on. When an answer is wrong or vague, the cause is almost always a missing or stale source, not the model.
OrhanAI accepts eleven kinds of training source. Each one solves a different problem:
Website for crawling pages from a site
Sitemap for pulling a known page list instead of discovering it
PDF for policies, manuals, and price lists
DOCX for internal documents that never made it to the website
TXT for plain notes and exports
CSV for structured rows such as products or locations
Image for screenshots, menus, and diagrams
Audio for recorded briefings and calls
Video for walkthroughs and demos
Manual Q&A for exact question and answer pairs you want answered a specific way
Plain text for pasting content directly
Image, audio, and video sources are indexed for multimodal retrieval, so a question can be answered from a diagram or a recording, not only from written text.
Website and sitemap URLs are normalized before anything runs, and a duplicate crawl source for the same chatbot is rejected rather than silently added twice.
During the crawl the source records what it found: the normalized input URL, discovered URLs, crawled URLs, and how many pages were skipped. If a crawl looks short, that metadata usually explains why, whether the site blocked requests or the page limit for your plan was reached.
Every source is split into chunks and embedded. Chunks carry the embedding profile, content kind, and a content hash so identical content is not embedded twice. Retrieval is filtered by workspace and chatbot, so a source from one chatbot never leaks into another.
At answer time the chatbot retrieves from your indexed content first and falls back only when retrieval is weak or the information is genuinely missing. That is what keeps answers grounded in your sources instead of invented.
A common mistake is adding everything at once. A smaller, accurate set beats a large, contradictory one.
Crawl the pages a customer would read before contacting you
Add the documents your team keeps sending by email
Add manual Q&A for the handful of answers that must be worded exactly
Only then widen to archives and long tail material
You do not have to guess at gaps. Visitor questions that ended in a fallback are clustered into knowledge gap records, and eligible chatbots can generate FAQ entries from real conversation clusters.
Read the gaps weekly. Each cluster is a specific page or document that does not exist yet.
Promote a gap into a training source, retrain, and the next visitor asking that question gets a real answer.
Deleting a training source removes its stored file and its knowledge chunks together, so removing outdated content actually removes it from answers. When a policy changes, replace the source rather than layering a correction on top of it.
Everyone assumes you need a developer to build an AI agent. You don't. If you have a website, a no-code AI platform can turn it into a working conversational AI agent in minutes, one that handles support, captures leads, and never clocks out. Here's exactly how to train, test, and launch it right.
