Knowledge and context
Making the Aimogen Pro chatbot answer from your own content using embeddings retrieval, page context, internet access and uploaded files.
Out of the box the chatbot knows two things: whatever is in Bot Context, and whatever the model learned during training. It does not know your products, your prices or your documentation.
There are four ways to give it your content.
1. Bot Context#
The simplest. Put the information directly into the system prompt.
Good for: a small, stable body of facts — opening hours, contact details, return policy, the three things you always get asked.
Not good for: anything large. Bot Context is sent with every message, so a 5,000-word context costs you on every turn and eventually crowds out the conversation.
Bot Context supports placeholders, so you can inject the current page:
The visitor is reading: %%post_title%%
Page content:
%%post_content_plain_text%%
Answer questions about this page.That single technique makes a page-aware assistant with no infrastructure at all.
2. Embeddings retrieval#
The proper solution for a real knowledge base. Your content is indexed as embeddings; each visitor question retrieves the most relevant chunks, and only those are injected into the prompt.
Enable it:
- Configure an embeddings provider and a vector store — see Knowledge base and RAG.
- Index your content — see Adding content to the index.
- Under Settings › Embeddings › Enable Embeddings For, tick Chatbot Shortcodes
(
embeddings_chat_short). - Optionally set a namespace so this bot only searches part of the index.
Per instance:
[aimogen-chat-form embeddings="enabled" embeddings_namespace="support-docs"]Namespaces are what let a support bot search documentation while a sales bot searches product pages. See Retrieval and prompt injection.
3. Internet access#
Lets the bot search the web and use the results.
Enable Settings › AI Internet Access › Chatbot (internet_chat_short). See
Search and SERP providers.
Good for: questions about current events, or topics outside your own content.
Costs: a search plus several page fetches before every affected reply. Noticeably slower, and more tokens.
Show Disable Internet Access Button gives visitors a toggle, and
Disable Internet Access Button Default State sets where it starts.
4. Uploaded files#
Visitors can upload a document and ask questions about it. See File uploads, vision and PDF chat.
Scoped to the conversation, not added to your knowledge base.
Combining them#
A well-configured support bot typically uses three:
| Layer | Carries |
|---|---|
| Bot Context | Identity, rules, and the handful of always-relevant facts |
| Page placeholders | What the visitor is currently looking at |
| Embeddings retrieval | Everything else — documentation, products, policies |
Internet access is usually left off for a support bot: you want answers from your own content, not from a web search.
Keeping the bot honest#
Retrieval reduces invention but does not eliminate it. Instruct explicitly:
Answer only using the information provided to you in this conversation and in
the retrieved context.
If the answer is not in that information, say: "I do not have that information.
You can reach our team at support@example.com." Do not guess, and do not use
general knowledge to fill gaps.
When you use retrieved information, keep to what it actually says. Do not
extrapolate prices, dates, compatibility or availability.Conversation memory#
Separate from knowledge. Max Chat Messages To Send As API Context and
Max Chat Context Size (Characters) control how much of the current conversation is resent each turn.
For memory across sessions, see Persistent chat, history and logs.
Troubleshooting#
The bot does not know something that is indexed. Check Chatbot Shortcodes is enabled under
Settings › Embeddings, and that the namespace matches. See
Retrieval troubleshooting.
The bot invents answers. Tighten Bot Context with an explicit "say you do not know" instruction, and lower the temperature.
Replies are slow. Internet access is on, or too many retrieval chunks are being injected. Reduce the top-K.
Replies cost more than expected. Every retrieved chunk is input tokens on every message. See Controlling cost.
Related#
Still stuck? Open a support ticket and include the diagnostics from Aimogen Pro › System & Logs › System Info.