Vector Desk - vector-desk.skillsafe.ai

This app's agent prompt is a derivative work of one agent skill. It paraphrases and adapts its guidance;
it does not reproduce the skill or its reference files verbatim.

1. vector-index-tuning
   Source:  github.com/wshobson/agents
   Path:    plugins/llm-application-dev/skills/vector-index-tuning/SKILL.md (and references/details.md)
   Author:  Seth Hobson (wshobson)
   Licence: MIT License
   Modifications: the skill's index-type-by-scale table (flat, HNSW, HNSW with quantization, IVF with PQ or
   DiskANN), its HNSW parameters (M, efConstruction, efSearch), its quantization sizes (FP32, FP16, INT8,
   PQ, binary), its memory estimate (vectors plus about 2 x M graph links per node), its best practices
   (benchmark with real queries, monitor recall, start with defaults, use quantization, do not over-optimize
   early, do not ignore build time, plan reindexing) and its Qdrant and benchmarking templates were restated
   as deterministic browser checks (vectorkit.js) and as the audit and fix lanes of this app's prompt, made
   concrete for pgvector (opclass and operator pairs, ef_search against the LIMIT, index dimension limits,
   iterative scans, lists and probes guidance), Qdrant (rescoring, always_ram beside on-disk originals, exact
   search), FAISS (PQ divisibility, training points per list, nprobe, refinement, cosine through normalized
   inner product) and hnswlib. The sweep reading (the configured row, the cheapest row inside every budget,
   recall that falls as ef rises, recall at another k), the operating-point contract, the exact before/after
   edit contract, the re-check of every edit and returned file, the name preservation check and the rule that
   the workload and the sweep are never edited are additions of this app.

The browser-side checks (vectorkit.js), the reconciliation (recon.js) and the rest of the page are original
to this app; the edit-apply routine and the page shell are shared with other SkillSafe desks. pgvector, Qdrant,
FAISS and hnswlib are projects of their own authors; this app is not affiliated with or endorsed by any of them.