Building Telephony Agents using LLMs
Transformer based Large Language Models have massive potential in understanding lossy audio conversations. LLM-agent is an open source best practice repo and hosted online playground to allow exploration with this emerging technology for free in order to evaluate its usefulness.
No code playground
The playground allows simple experimentation, often generating breathtaking results from very basic prompt engineering. The Example Interaction on this page is an amusing demonstration which was generated by connecting two playground agents to each other with prompts written in a few minutes which asked them to negotiate a contract for doughnut supply as purchaser and sales person respectively.
Building production services
The playground is a great way to quickly explore the technology without writing code. To deploy production agents interacting with real humans, a little more work is required to apply guardrails and shape the conversation to dialogue goals by interacting with external services.
Developing to the llm-agent API is the way to do this. It will allow you to:
- Create an agent and link it to a phone number
- Monitor interactions with the caller and feed data into your business logic
- Modify the prompt as the call proceeds to change contexts and narrow the conversation to align with goals for the interactions
- Interact with external services within your code in order to:
- Further target the prompt being used by the agent
- Inject spoken results directly into the conversation