The all-in-one reasoning engine with the AEGIS knowledge shield.
NYXIS-PRO combines the NYXIS1.1B base model with a curated 5,000+ chunk knowledge base covering coding, math, science, history, and common knowledge. A dual‑stage retriever grounds every factual answer in verified data — zero hallucination.
🧠 What Makes It Different
| Feature | Description |
|---|---|
| AEGIS Knowledge Base | 5,426 curated Wikipedia, Trivia QA, and ArXiv chunks |
| Dual‑Stage Retriever | Dense recall (bge‑small) + Cross‑encoder rerank (ms‑marco‑MiniLM) |
| Emotion Engine | Detects user emotion on a 1‑10 scale and matches tone |
| 4‑Tier Knowledge Cascade | AEGIS → External Search → Model Brain → Honest Fallback |
| Zero Hallucination | If it doesn't know, it says so — no fabrication |
🚀 Quick Start
from huggingface_hub import snapshot_download
import sys
model_dir = snapshot_download("QuantaSparkLabs/NYXIS-Pro")
sys.path.insert(0, model_dir)
from pipeline import NYXISPro
nyxis = NYXISPro(model_dir)
result = nyxis.generate("What is a binary search tree?")
print(result["response"])
Requirements:
sentence-transformers,faiss-cpu,transformers,accelerate,bitsandbytes
VRAM: ~2.1 GB (4‑bit) | Hardware: T4 or better
📦 What's Inside
| Component | Model | Purpose |
|---|---|---|
| Base LLM | NYXIS1.1B) | Text generation |
| Dense Retriever | bge‑small‑en‑v1.5 | Initial candidate recall |
| Cross‑Encoder | ms‑marco‑MiniLM‑L‑6‑v2 | Precision re‑ranking |
| FAISS Index | 5,426 chunks | Knowledge storage |
| Emotion Engine | Custom keyword‑based | Tone matching |
🤖 Emotion Intelligence
NYXIS-PRO detects your emotional state from text and adjusts its personality:
| Level | Emotion | Response Style |
|---|---|---|
| 1‑3 | Sad / Angry | Gentle, supportive, empathetic |
| 4‑6 | Neutral | Balanced, warm, helpful |
| 7‑9 | Happy / Excited | Energetic, playful |
| 10 | Overjoyed | Celebratory |
🛡️ Knowledge Cascade
When you ask a factual question, NYXIS-PRO follows a strict 4‑tier protocol:
- AEGIS — Searches internal verified knowledge base
- External Search — Falls back to web search (user‑provided API)
- Model Brain — Uses training data with honesty guard
- Honest Fallback — Admits "I don't have enough information"
It never fabricates an answer.
Limitations
- Knowledge base covers ~5,400 chunks — not infinite
- Emotion detection is keyword‑based, not deep sentiment analysis
- External search requires user‑provided API function
- English‑only, 1.5B model size limits complex reasoning
License
Apache‑2.0
Built with 🛡️ by QuantaSparkLabs