Entha is a proof-of-concept (PoC) implementation of multi-hop document retrieval that enhances traditional Retrieval-Augmented Generation (RAG) pipelines. It improves contextual understanding by not only retrieving the top-k most relevant documents for a query, but also expanding context by retrieving top-n semantically similar documents based on those initial results.

Traditional RAG systems retrieve a fixed number of documents (top-k) based solely on query relevance. However, this can miss related but indirectly connected context. Entha addresses this by:

1. Retrieving the top-k documents relevant to the user's query.
2. For each of those documents, retrieving top-n additional documents based on embedding similarity.
3. Combining all results to create a richer and more connected context for the LLM to generate responses.

• ✅ Multi-hop context retrieval (top-k + neighbors of top-k)
• ✅ Google Generative AI Embeddings for semantic similarity
• ✅ ChromaDB for vector storage and retrieval
• ✅ Streamlit-based interface for easy interaction
• ✅ Modular design for plug-and-play LLMs

• LangChain
• ChromaDB
• Google Generative AI
• Streamlit

User Query
   │
   ├──▶ Top-K Document Retrieval (based on embeddings)
   │         │
   │         └──▶ Top-N Neighbors (per document)
   │                     │
   └────────────────────▶ Combined Context
                          ↓
              Sent to LLM (Gemini)