Do you have any examples of real voice apps using few shots?

Thanks for reaching out, Mario. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

few shots, as in, llm instructions with examples

Thanks so much for your question! LiveKit voice apps mainly use zero-shot prompting, which means the LLM follows instructions that describe the behavior without needing multiple input-output examples. For example, agents like the Meditation Coach or Customer Support Agent work based on clear, descriptive instructions, but they don’t include structured few-shot examples in their prompts. That said, LiveKit does offer some great features like dynamic context injection and multi-agent workflows. These can be really helpful if you want to try out few-shot prompting. For instance, you could dynamically add input-output pairs into the LLM’s context during a session to help guide its responses more precisely. If you’d like, I’d be more than happy to walk you through how to set this up or explore these options together. Just let me know, and we can take it from there! Sources: web/src/data/presets.ts | examples/src/gemini_realtime_agent.ts | pipeline-tts/playai_tts.py | examples/voice_agents/multi_agent.py

Sure

I'd be happy to help you implement few-shot prompting in your LiveKit voice apps! A great way to do this is by structuring your LLM instructions with input-output examples. Here’s a friendly guide to get you started: 1. Define the Prompt: Begin with a clear description of the agent’s role and behavior. Then, include some input-output examples to gently guide the model. For instance:

Role: You are a helpful customer support agent.

Example 1:
User: I want to book an appointment.
Assistant: Sure! When would you like to come in? We have openings tomorrow morning or Thursday evening.

Example 2:
User: Can I get a callback later?
Assistant: Of course. What time works best for you?

2. Inject Context Dynamically: You can use LiveKit’s dynamic context injection features to add these examples into the LLM’s context during a session. This might involve using the

llm_node

function or similar tools to update instructions on the fly based on what the user says. 3. Test and Iterate: It’s always a good idea to run some tests to see how the LLM responds. Feel free to tweak the examples or instructions to make the agent’s behavior just right. If you’d like, I’m here to help with any specific steps, like using

llm_node

for context injection or crafting prompts tailored to your needs. Just let me know, and I’ll gladly guide you through it!