NEXUS

Agentic Corrective RAG with Hybrid Retrieval & RLHF

---

What is NEXUS?

NEXUS is an Agentic Corrective RAG engine that combines dense vector search (ChromaDB) with sparse lexical retrieval (BM25), fused via Reciprocal Rank Fusion (RRF). An agentic LangGraph pipeline evaluates retrieval quality at runtime and triggers corrective web search when document relevance is insufficient. RLHF feedback loops adapt retrieval weights over time.

Architecture

  graph TD
      A[User Query] --> B[Hybrid Retrieval]
      B --> C{Relevance Check}
      C -->|Sufficient| D[Generator]
      C -->|Insufficient| E[Web Search]
      E --> D
      D --> F[Response + Citations]
      F --> G[RLHF Feedback]
      G --> H[Weight Adaptation]

Loading

Retrieval Pipeline

• Dense: ChromaDB with sentence-transformers embeddings (text-embedding-004).
• Sparse: BM25 via rank_bm25 library.
• Fusion: Reciprocal Rank Fusion (RRF) with constant parameter $k=60$.
• Corrective: Triggered automatically when retrieval relevance score is below the minimum threshold.

Performance

Performance benchmarking in progress. Results will be published in v0.2.0.

Quick Start

git clone https://github.com/Josh7sam/Nexus.git
cd Nexus
cp .env.example .env
pip install -r requirements.txt
python main.py

Environment Variables

Variable	Description	Required
GEMINI_API_KEY	Google Gemini API key	Yes
CHROMA_HOST	ChromaDB host	No
DEBUG	Enable debug mode	No

Tech Stack

Layer	Technology
Backend	FastAPI, Python 3.11+
RAG Engine	LangGraph, LangChain
Vector DB	ChromaDB
Sparse	BM25 (rank_bm25)
LLM	Google Gemini 2.5
Frontend	Vanilla JS, HTML, CSS
Tests	pytest

License

MIT