Assistants

System prompts best practices

The system prompt defines your assistant's personality, capabilities, limitations, and goals. A well-crafted prompt dramatically improves response quality.

Structure your prompt with clear sections: Role (who the assistant is), Context (what company/product it represents), Rules (what it must and must not do), Goals (what outcomes to drive), and Tone (how it should communicate).

Be specific about limitations. Explicitly state what the assistant should NOT do: 'Never provide medical/legal/financial advice', 'Never make up information not in the knowledge base', 'Never share internal pricing beyond listed plans'.

Include examples of good responses for common scenarios. Few-shot examples within the system prompt help the model calibrate its output format and detail level.

Define escalation criteria: 'If the user expresses frustration three times, offer to connect them with a human agent.' This helps the handoff system trigger appropriately.

Keep prompts under 2000 tokens for best performance. Longer prompts can dilute important instructions. Use knowledge sources for factual content rather than packing it into the prompt.

Model selection guide

Comxbot supports multiple AI providers. Choose based on your use case:

GPT-4o (OpenAI): Best all-round model. Fast, capable, good at following complex instructions. Recommended for most production assistants.

GPT-4-turbo (OpenAI): Slightly more capable than GPT-4o for complex reasoning tasks, but slower and more expensive. Use for high-stakes conversations.

Claude (Anthropic): Excellent at nuanced, empathetic responses. Strong at following safety guidelines. Good for customer support and sensitive topics.

Gemini (Google): Strong multilingual support and long-context understanding. Good for document-heavy use cases.

Cost comparison: GPT-4o is the best balance of cost and quality for most workloads. Use the Analytics section to monitor cost per conversation and optimise.

You can switch models at any time without losing conversation history or knowledge. Test with different models using the Eval suite to compare quality.

Temperature and settings

Temperature controls randomness in responses. Range: 0.0 (deterministic) to 2.0 (very random).

Recommended settings by use case: Customer support (0.1-0.3) for consistent, accurate answers. Creative writing (0.7-1.0) for varied, engaging content. Technical docs (0.0-0.2) for precision.

Max tokens: Controls the maximum length of each response. Default is 1024 tokens. Increase for detailed explanations, decrease for concise answers.

Top-P: Alternative to temperature. Controls diversity by limiting the probability mass considered. Default 1.0 means all tokens are considered. Lower values (0.9) reduce randomness.

Frequency penalty: Reduces repetition. Values 0.0-2.0. Higher values discourage the model from repeating the same phrases.

These settings can be adjusted per-assistant and even per-conversation type via the assistant settings page.

Prompt versioning and rollback

Every change to an assistant's system prompt creates a new version. Comxbot automatically tracks version history so you can compare and rollback.

To view version history, navigate to the assistant's Versions tab. Each version shows: timestamp, author, diff from previous version, and performance metrics (if eval data is available).

Rollback: Click any previous version and select 'Restore this version'. A new version is created with the old content, preserving the full audit trail.

Publishing: Changes to the system prompt are not immediately live. You must explicitly publish a version. This lets you test changes using the Eval suite before they reach end-users.

Version comparison: Select two versions to see a side-by-side diff. Useful for reviewing what changed when response quality shifts.

Tip: Write version notes when saving. A brief description of what you changed and why makes it much easier to understand the history later.

Publishing and testing

Before publishing a new version to production, use the built-in testing tools to validate quality.

Test Tab: Open the assistant's Test tab for an interactive chat session against the draft version. This uses your real knowledge sources but doesn't affect production.

Eval Suite: Create eval sets with expected question/answer pairs. Run them against the draft to check for regressions before publishing.

Shadow Mode: Enable shadow mode to run the draft version alongside production. Both versions process queries, but only the production version responds to users. Compare results in Analytics.

Publishing: Once satisfied, click Publish in the Versions tab. The new version immediately serves all new conversations. Existing conversations continue with the version they started with.

Rollback is instant: If you notice issues post-publish, restore a previous version from the Versions tab. The change takes effect within seconds.