35 lines
2.2 KiB
Markdown
35 lines
2.2 KiB
Markdown
---
|
||
tags:
|
||
- project/iso27DIY
|
||
- dev
|
||
- llm
|
||
- agent
|
||
---
|
||
# Create an interview agent
|
||
|
||
We need to create a conversational agent that proactively drives the dialogue, also called a 'mixed-initiative' or 'proactive' conversational agent.
|
||
|
||
We need to design prompts/instructions that encourage initiative. Here's an example prompt:
|
||
```
|
||
You are a proactive assistant. Start the conversation by greeting the user and asking them about their needs. If their response is unclear, ask follow-up questions to clarify. Guide them step by step through the process, making suggestions and confirming understanding at each stage.
|
||
```
|
||
|
||
|
||
Implement dialogue management (manage conversation state):
|
||
|
||
- In a **stateless** setup, your application manages the conversation history and sends it with each request.
|
||
- In a **stateful** setup, the platform manages context, and you only send new messages.
|
||
- For advanced initiative, consider reinforcement learning or preference optimization techniques (such as Action-Based Contrastive Self-Training) to train the model to recognize ambiguity and take action to clarify or drive the conversation forward.
|
||
|
||
Incorporate prompts and follow-ups:
|
||
|
||
- Use follow-up prompts and context tracking to manage multi-turn flows, refining the conversation at each step and ensuring the bot leads the user toward a goal.
|
||
|
||
Use LLMs with strong context retention and multi-turn support.
|
||
|
||
Recommended Chat Models:
|
||
- Qwen: Frequently praised by users for its conversational quality and reliability in chat-based applications. It is considered a strong choice for structured interviews due to its ability to maintain context and handle follow-up questions effectively.
|
||
- Meta Llama (Llama 3 and variants): Popular for general-purpose chat and conversational AI tasks. Llama models are known for their robust performance and can be fine-tuned or prompted to follow structured interview formats.
|
||
- Cohere Command R: Cohere’s latest conversational model, optimized for dialogue and tool use, is highlighted for its 2024 update and is suitable for building structured, interactive interview agents.
|
||
- Mistral (Dolphin and others): Noted by some users as a reliable alternative for chat applications, offering strong conversational abilities and context retention.
|
||
|