Self-hosted alternatives to vector-index-tuning

V

vector-index-tuning

Optimize vector index performance for latency, recall, and memory

vector-index-tuning is an agent skill from the wshobson/agents collection for optimizing vector indexes in production for latency, recall and memory. A vector index lets a search find nearest neighbors among many embeddings quickly, and tuning it means trading speed against accuracy and memory. The skill's core concepts begin with index type selection, and it covers tuning HNSW parameters, the settings of a common graph-based index that control how many connections are kept and how far the search explores, along with implementing quantization, which stores vectors in fewer bits to save memory at some cost in accuracy. It addresses balancing recall against speed, meaning the share of true nearest neighbors the search returns versus how fast it returns them, reducing search latency, and scaling to very large collections, up to billions of vectors. It applies when tuning HNSW settings, choosing quantization, optimizing memory, reducing search latency, balancing recall with speed, and scaling up. The right settings depend on the data and the queries, so the agent should measure recall against an exact search on a sample of real queries before a change is adopted. The skill is installed with the skills CLI from the wshobson/agents repository. Its skills.sh page showed about 10.4K installs and passing results from three security audits at the time of the scan.