In this article, Sameer Ahmed describes the development of Vektr, a retrieval-augmented generation (RAG) engine that includes a custom-built vector search system instead of relying on established solutions such as Pinecone, Weaviate, or FAISS.The main goal was educational: to gain a deeper understanding of how modern vector databases work by implementing their core components from scratch.The project centers on HNSW (Hierarchical Navigable Small World) graphs, a widely used approximate nearest neighbor search algorithm.

The author explains how HNSW organizes vectors into multiple graph layers, enabling efficient navigation through sparse upper levels before refining results in denser lower layers.Java code examples illustrate index construction, insertion, graph connectivity management, and search procedures.Beyond vector similarity search, the system incorporates hybrid retrieval by combining dense vector search with BM25 keyword retrieval.

Results from both methods are merged using Reciprocal Rank Fusion (RRF), improving overall relevance by balancing semantic understanding and exact term matching.

The article also explores HyDE (Hypothetical Document Embeddings), a technique that first generates a hypothetical answer to a query and then embeds that generated text rather than the original question.According to the author's testing, this approach increased recall@10 by approximately 8% for factual queries.Another engineering focus is reliability.The author implemented atomic index persistence using a temporary file and atomic rename operations to prevent index corruption during writes.Performance testing on 1,000 sentence-embedding vectors achieved a recall@10 of 0.

984, cold-query latency of about 35 milliseconds, cached-query latency below 1 millisecond, index loading under 15 milliseconds, and index construction in roughly 200 milliseconds.

The article concludes with lessons about HNSW parameter tuning, the value of hybrid retrieval, the importance of atomic writes, and the effectiveness of HyDE for information retrieval systems.