wtf is distributed-vector-memory-routing?

This repository holds the code and data behind a research paper about making distributed vector search cheaper to run. Vector search is how systems like semantic search or AI memory find the most relevant pieces of information for a query, and in a distributed setup that data is split across many separate partitions, or shards, that each need to be checked. Checking every partition for every query is accurate but expensive, so this project tests smarter ways to decide which partitions actually need to be contacted. Four different routing strategies are compared. One expands the search to more partitions only when the system is not confident enough in what it already found. Another always checks a fixed number of the closest partitions. A third keeps expanding partitions gradually until a target confidence level is reached. The last one sets a hard limit on how many partitions can be contacted at all, regardless of confidence. These are tested against two simple baselines: checking every single partition, which is accurate but costly, and checking only the closest one, which is cheap but less accurate. Using a dataset called DBpedia 14, which contains 100,000 documents across categories like companies, athletes, films, and animals, the experiments show that the adaptive strategies can cut communication between partitions by 42 percent while keeping the same retrieval quality as checking everything. The paper also finds that grouping data with a clustering method called KMeans works noticeably better than grouping it randomly. The project is delivered as a Jupyter notebook that can be run on Kaggle or locally, along with the paper itself in LaTeX and PDF form and the figures used in it. It requires a few Python packages including a library called TurboVec, along with common tools for machine learning and text data. The code is MIT licensed while the paper and figures are released under Creative Commons Attribution 4.0, and it has 0 stars listed on GitHub.