"Emotional RAG: Enhancing Role-Playing Agents through Emotional Retrieval"

Emotional RAG: AI now recall memories based on emotions, just like humans do.

📚 https://arxiv.org/abs/2410.23041

Original Problem 🤔:

Role-playing agents powered by LLMs struggle to maintain consistent personality traits and generate human-like responses due to limited emotional context in memory retrieval.

-----

Solution in this Paper 💡:

• Introduces Emotional RAG framework for role-playing agents

• Encodes both semantic and emotional vectors for queries and memory

• Implements two retrieval strategies:

- Combination: Fuses semantic and emotional similarity scores

- Sequential: Retrieves based on one factor, then reranks using the other

• Designs emotion-aware prompt templates for LLMs

-----

Key Insights from this Paper:

→ Incorporating emotional states in memory retrieval enhances personality consistency

→ Mood-Dependent Memory theory from psychology applies to AI agents

→ Different retrieval strategies work best for different personality evaluation metrics

→ Emotional congruence improves the human-likeness of generated responses

-----

Results 📊:

• Outperforms traditional RAG methods across multiple datasets

• Significant improvements in full personality evaluations (MBTI, BFI)

• Better performance on open-source models (ChatGLM-6B, Qwen-72B) compared to GPT-3.5

• Achieves higher accuracy in overall personality trait predictions