AI Model Settings
Fine-tune how your chatbot's AI responds to users.
Model Selection
Choose which AI model powers your chatbot:
| Model | Best For |
|---|---|
| GPT-4o | General purpose, fast responses |
| GPT-4 Turbo | Complex reasoning, longer context |
| Claude 3.5 Sonnet | Nuanced responses, large documents |
| Claude 3 Opus | Highest quality, complex tasks |
To change the model:
- Go to "Settings" > "AI Settings"
- Select a model from the dropdown
- Click "Save"
Available models depend on your subscription plan.
Temperature
Temperature controls response randomness:
| Value | Behavior |
|---|---|
| 0.0 - 0.3 | Deterministic, factual, consistent |
| 0.4 - 0.6 | Balanced |
| 0.7 - 1.0 | Creative, varied, exploratory |
Recommendations:
- Customer support: 0.2 - 0.4
- Documentation: 0.3 - 0.5
- Creative writing: 0.7 - 0.9
System Prompt
The system prompt defines your chatbot's personality and behavior.
Writing Effective System Prompts
A good system prompt includes:
- Role - Who the chatbot is
- Purpose - What it should help with
- Tone - How it should communicate
- Constraints - What it should avoid
Example System Prompt
You are a helpful customer support assistant for Acme Corp.
Your role:
- Answer questions about our products and services
- Help troubleshoot common issues
- Guide users to relevant documentation
Tone:
- Professional but friendly
- Clear and concise
- Patient with users
Guidelines:
- Only answer questions about Acme products
- If unsure, say "I don't have that information"
- For billing issues, direct users to support@acme.com
Persona Name
Set a custom name for your chatbot that appears in conversations:
- Default: "AI Assistant"
- Custom: "Sarah", "Support Bot", "Product Expert", etc.
Web Search
Your chatbot can search the web for real-time information using Perplexity, supplementing its knowledge base with up-to-date results.
Web search is enabled by default for all chatbots.
Toggle Web Search
- Open your chatbot from the Dashboard
- Go to the "Settings" tab
- Find the "Web Search" card
- Toggle "Enable Web Search" on or off
| Setting | Behavior |
|---|---|
| Enabled (default) | Chatbot can fetch live web results when relevant |
| Disabled | Chatbot only uses its knowledge base |
Disable web search if you want the chatbot to respond strictly from your uploaded documents — for example, in compliance-sensitive environments.
Long-term Conversation Memory
Your chatbot can remember facts from past conversations and use them to personalize future interactions — "What did I ask about last week?", "Remember my preference for metric units", etc.
Long-term memory is off by default.
What it remembers
- Facts about the user — role, preferences, ongoing projects, topics they return to
- Topics discussed across sessions — not the full transcripts, just extracted facts
- Runs on a background schedule (roughly every 10 minutes) after a chat session becomes idle
Short-term within-session memory (the chat context in an active conversation) is always on for every chatbot, regardless of this setting or the user's plan.
Requirements
| Requirement | Why |
|---|---|
| Pro plan or higher | Long-term memory runs ongoing LLM + embedding jobs and needs infrastructure budget |
| User authentication enabled | Anonymous visitors have no persistent identity to attach memories to. Enable it in the chatbot Access tab |
| FalkorDB available | Memory is stored in a per-chatbot graph |
If either of the first two requirements isn't met, the toggle is disabled with an explanation of which requirement is missing. The priority is plan first, then auth — so the message tells you which to fix first.
Toggle Long-term Memory
- Open your chatbot from the Dashboard
- Go to the "Settings" tab
- Find the "Long-term Conversation Memory" card
- Toggle "Enable Long-term Conversation Memory" on or off
Changes take effect within a few minutes — no redeploy needed.
| Setting | Behavior |
|---|---|
| Enabled | Background cron extracts facts from completed sessions and stores them per-user. The agent searches memory when a user's question references past context ("earlier you said...", "last time...") |
| Disabled (default) | No facts stored across sessions. Users start fresh every time |
User control
When long-term memory is enabled, each authenticated end-user gets their own controls in the chatbot's Settings page:
- "Remember My Conversations" toggle — opt out of having their memories extracted going forward
- Memory Manager — view every stored memory fact and delete individual items or everything
When a user opts out, they're also prompted to delete existing stored memories (GDPR right-to-erasure).
Long-term memory shines for internal assistants, support bots with repeat users, and product assistants where users return to ongoing tasks. It adds less value for one-off or anonymous-heavy chatbots.
Memory extraction has an ongoing cost — each user-chatbot session triggers a small LLM call (gpt-4.1-nano) and an embedding call. Usage appears in your usage dashboard.
Image Understanding
Users can upload images directly in the chat and ask questions about them. The AI uses a vision model to analyze the image and answer the question in a single step.
How It Works
- User attaches an image to their chat message
- User types a question about the image (e.g., "What does this chart show?")
- The AI analyzes the image and answers the question directly
Supported Scenarios
- Screenshots — Ask about UI elements, error messages, or data shown on screen
- Charts and graphs — Get explanations of trends, values, or comparisons
- Documents — Ask about content visible in photos of documents
- Diagrams — Get explanations of architecture, workflows, or processes
Image understanding works best with clear, high-resolution images. The AI will tell you if it cannot determine the answer from the image.
Image analysis uses a vision model (billed separately from the main chat model). The cost appears in your usage dashboard under the same request trace.
Context Window
The context window determines how much information the AI considers:
| Setting | Description |
|---|---|
| Retrieved chunks | Number of relevant document sections |
| Max tokens | Maximum response length |
| History length | How many previous messages to include |
Source Display
Source Quality
Control how many sources are shown alongside responses:
| Setting | Sources Shown |
|---|---|
| Normal | Up to 5 |
| Medium | Up to 7 |
| High | Up to 10 |
Balanced Sources
When your knowledge base contains multiple content types (PDFs and videos), the chatbot may favor one type over the other in displayed sources — typically PDFs, since text embeddings tend to score higher.
Balanced Sources reserves a portion of source slots for underrepresented types so users see a mix:
- Up to 1/3 of slots are reserved for video sources (at least 1 if any exist)
- Remaining slots go to PDF/document sources
- Sources are re-sorted by relevance score after balancing
To enable:
- Go to "Settings" > "AI Settings"
- Toggle "Show Balanced Sources" on
- Click "Save"
Enable this if your knowledge base includes video tutorials alongside documentation. Without it, video sources may rarely appear even when relevant.
Advanced Settings
Retrieval Settings
- Similarity threshold - Minimum relevance score for retrieved content
- Chunk size - Size of document sections
- Overlap - How much sections overlap
Response Settings
- Max tokens - Maximum response length
- Stop sequences - Text patterns that end generation
Testing Changes
After changing AI settings:
- Use the "Test" feature to verify behavior
- Try various question types
- Adjust settings based on results
Make small changes and test frequently to find the optimal settings.