SaiAkhil066/CORTEX-AI-SUPER-RAG

What it solves

Cortex RAG is a local retrieval-augmented generation (RAG) system that allows users to upload documents and ask questions about them without relying on cloud services, API keys, or subscriptions. It addresses the limitations of basic RAG by implementing a multi-layered pipeline to improve retrieval accuracy, reduce noise, and provide transparent reasoning.

How it works

The system processes documents (PDF, DOCX, TXT, MD) through a 9-layer pipeline:

1. Indexing: It creates three separate indexes (BM25, FAISS, and a NetworkX knowledge graph) and can optionally enrich chunks with surrounding context using an LLM.
2. Query Processing: It uses a semantic cache to instantly return answers for repeat questions and RAG-Fusion to expand a single query into multiple variants for better recall.
3. Retrieval & Refinement: It merges results from vector and graph searches using Reciprocal Rank Fusion (RRF), applies a neural reranker (Cross-Encoder) to prioritize the most relevant passages, and uses Corrective RAG (CRAG) to drop irrelevant chunks.
4. Generation: The final answer is generated by a local LLM via Ollama, with a live reasoning panel that streams the model's chain-of-thought process.

Who it’s for

It is designed for users who need a private, local AI assistant for document analysis and organizations seeking production-grade RAG systems that run entirely on their own infrastructure.

Highlights

• Local-First: Runs entirely on the user's machine using Ollama, ensuring data privacy and zero cloud dependency.
• Advanced Retrieval: Combines vector search, sparse retrieval (BM25), and GraphRAG for comprehensive context.
• High Precision: Employs neural reranking and CRAG to filter out noise before generation.
• Reasoning Transparency: Includes a live reasoning panel to watch the model's internal thought process.
• Flexible Model Support: Compatible with any Ollama model, including Llama 3.1, Qwen 2.5, and Mistral.