one thing your agent can do with flexvec is search for information that is similar to "architecture" but suppress "changelogs" before selection by applying arithmetic to scores.
ai agents are becoming the primary consumers of retrieval and search tools. we used to type into google what we wanted to find. now we ask ChatGPT, Claude or Gemini to find it for us — they run searches and return the results.
in building tools for claude code to search session history, memories, specs, plans, documents and personal knowledge bases — i created a vector retrieval library called flexvec that (1) works great for structured data (2) gives more control over semantic retrieval.
most retrieval systems out there are built on the constraints of this past era.
i wrote a paper on the approach: arxiv.org/abs/2603.22587
