How to Use AI Assistants
This guide walks you through creating an assistant, connecting it to your knowledge, tuning how it responds, and deploying it to your users.
What is an assistant?
An assistant is an AI agent configured for a specific job: it has its own instructions, its own knowledge base of uploaded documents, a selected AI model, and optional tools. You can run several assistants in one organization — for example one for customer support and another for internal documentation.
Set up an assistant step by step
- 1
Create an assistant
In the dashboard, open Assistants and click Create New Assistant. Start from an agent template for common roles (support, sales, internal helpdesk) or from a blank assistant, then give it a descriptive name.
- 2
Write instructions
Define the assistant's role, tone, boundaries, and output format. See the instructions vs. knowledge guide for what belongs in instructions and what belongs in uploaded files.
- 3
Add knowledge
Upload the documents the assistant should answer from — PDF, Word, Office, and CSV files are parsed automatically, and scanned documents go through OCR. See how files are organised.
- 4
Pick a model
Choose from OpenAI, Anthropic, Meta, and DeepSeek models. Basic models are cost-effective for routine tasks; Premium models handle complex reasoning. See choosing a model.
- 5
Set up predefined answers
Add predefined answers for questions you want answered the same way every time. Matching questions are served from the FAQ cache — instant, consistent, and without consuming model tokens.
- 6
Test in the playground
Use the playground to chat with the assistant before deploying. Try real questions from your users, compare models on the same prompt, and refine instructions until responses are reliable.
- 7
Deploy
Connect the assistant to the chat widget to put it on your website, or use it directly inside the platform.
Vector store query mode
This setting controls when the assistant searches your uploaded documents to answer a question.
Vector store query mode options (click to enlarge)
Auto (recommended)
The assistant decides per message whether the knowledge base is needed. Best balance of speed, cost, and accuracy for general conversations.
Always query
The knowledge base is searched for every message. Use it for document-heavy workflows where answers should always be grounded in your files.
Chat widget assistants
If an assistant is connected to a chat widget, this setting is locked — the query mode is optimized automatically for web interactions.
LLM parameters
Instructions define what the assistant should do; LLM parameters tune how it responds — creativity, length, and how much knowledge-base context it pulls in.
LLM parameter settings (click to enlarge)
Swipe sideways to see the full table
| Parameter | What it controls | Recommended |
|---|---|---|
| Temperature | Creativity and randomness. Lower values give consistent, focused answers; higher values give more varied ones. | 0.3 for business use, 0.7 for creative tasks |
| Max Tokens | Maximum response length. One token is roughly 3-4 characters. | 2000 for detailed responses, 500 for brief answers |
| Top P | Word selection diversity. Lower values stick to the most likely words. | 0.7-0.9 for balanced responses |
| Frequency Penalty | Reduces repetition of words and phrases within a response. | 0.3-0.7 |
| Presence Penalty | Encourages the assistant to move on to new topics instead of circling back. | 0.0-0.6 |
| Vector Search Top K | Number of knowledge-base documents retrieved per query. More context, but potentially less relevant. | 3-10 depending on document complexity |
Recommended parameter ranges (click to enlarge)
Quick presets
- Factual support and Q&A: temperature 0.3, top P 0.7 — precise, consistent answers.
- General conversation: temperature 0.7, top P 1.0 — a balanced starting point.
- Creative content: temperature 1.0+, top P 1.0 — varied, exploratory output.
Change one parameter at a time so you can see what each adjustment does.
Human handoff inbox
When a conversation needs a person, the assistant can hand it off to the live conversations inbox. Your team sees active chats in real time, can take over from the AI mid-conversation, and can hand the conversation back once the issue is resolved. Use it for escalations, complaints, or anything your instructions mark as human-only.
Best practices
- Keep knowledge current: use clear, well-structured documents with descriptive names, and remove outdated files.
- Start simple: get a basic assistant working end to end before adding tools or advanced settings.
- Test with real questions: use actual user phrasing in the playground, not idealized prompts.
- Iterate on instructions: when the assistant answers poorly, refine instructions or knowledge before changing model parameters.