KeyVec

keyvec

Notes on corpus deduplication, retrieval evaluation, and the evidence LLM systems leave behind.

About meGitHubRSS

Earlier writing

All posts
  1. A Retrieval Eval That Survives Index Changes

    Notes on building a small, fixed question set for a RAG system so it measures correctness instead of quietly agreeing with whatever the current index returns.

  2. Evidence-First LLM Systems

    A working note on the question this site keeps returning to: what evidence should technical work leave behind, and how to design pipelines so the evidence is a byproduct, not a chore.

Projects

All projects
  1. RAG Evaluation HarnessActive · 2026

    A repeatable loop for measuring retrieval quality, citation grounding, and refusal behavior in a small knowledge-base assistant — built so each run leaves behind debuggable evidence.