Fuli & deltaEGO – Emotion-aware RAG 1.0 Agent

Experimental emotion-aware memory orchestrator for character AI.
Built by a college student for learning and feedback.

(Figure: Sequence Diagram of Fuli's interaction with LLM and deltaEGO)

Project Links

• RAG Core (Fuli): GitHub Repository
• Emotion Module (deltaEGO): GitHub Repository

Looking for code review

• architecture (C++ deltaEGO core + Python wrapper),
• RAG 1.0 memory design (FAISS + JSON),
• concurrency / I/O patterns (async + threads),
• any bad practices or better patterns.

• open an issue on GitHub, or
• start a Discussion on this Hugging Face repo.

Overview

• FAISS + SentenceTransformer for retrieval (RAG 1.0).
• deltaEGO (C++ + Python) as a VAD-based emotion engine.
• Fuli as the orchestration layer that:
  • receives VAD JSON from an LLM,
  • runs deltaEGO search + analysis,
  • updates memories and logs,
  • and returns text context for the LLM.

Architecture (high-level)

1. User input → Fuli
2. Fuli → FAISS: retrieve relevant memories
3. Fuli → LLM: ask for VAD (Valence, Arousal, Dominance) as JSON
4. VAD → deltaEGO:
   • VAD vector DB search (top-k emotion labels)
   • emotion analysis (stress / reward / whiplash)
5. Fuli:
   • decides impressiveness of the turn
   • updates memories (recent / impressive)
   • logs everything as Fuli_LOG
6. Fuli → LLM: builds an emotion-aware prompt for the final reply