Infrastructure
-
On the read axis alone, self-hosting a vector DB loses at 6.76 million vectors even at zero labor cost
The split between self-hosting a vector DB and buying dedicated nodes closes on one number, the allowable labor budget. L = P k r + W + A − (self-hosted infrastructure − s S), where P is the monthly rate per dedicated node, k the shard count, r the replica factor, W the managed write bill, A managed-only add-ons, and s S the managed storage bill. On this post's baseline, which sets W = 0 and A = 0, L goes negative past 6,764,154 vectors, so self-hosting loses even if you book zero monthly labor. On-demand and dedicated nodes split where corpus GB × QPS = 8.0 × k × r, which at $16 per RU is 0.40% × k × r of the on-demand default rate limit.
-
The 16.22% Reindex Break-Even Is Not Your Number
In RAG reindexing, the break-even between incremental and full reindexing is a rewritten-chunk share, and it closes into one formula, q* = (π + c + ρ) ÷ (π + 1 + ρ). π is parsing cost, c is full-reindex write cost, and ρ is embedding cost, each divided by incremental write cost, and all three come off vendor price lists. Depending on how the parse cache, the load path, and the embedding cache combine, that value lands anywhere from the 3% range to the 90% range.