serhiykucherenko.dev

ship log

everything that shipped or changed — newest first

start here · three of 13

how the measuring startedWhen 60% recall meant 90% accuracy: a RAG measurement story ↗6/10 → 9/10 the eval that keeps it honestJudge strictly to avoid hallucinations ↗refusal 0 · fabrication 0 the hand-built one against five othersComparing RAGs, Part 2: the benchmark ↗4th on accuracy, 1st on cost and speed

wroteshippedmilestone

oct 02

wrote

same answer page: rank 2, 46 and 10 depending on the wording

The same question, asked three ways

measuring a rag · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

sep 25

wrote

the passage was at rank 7; the model only sees the top 5

The answer was at rank 7

measuring a rag · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

sep 18

wrote

hybrid search lost 2 to 3 answers in 20, in every language

Asking a RAG in the wrong language

measuring a rag · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

sep 14

wrote

66/80 outcomes identical across 6 model stacks

Swapping every model in a RAG

judging answers · read on dev.to ↗

also on hashnode ↗

sep 12

wrote

recall@5 0.60 → 0.80

Improving the scores of a RAG

measuring a rag · read on dev.to ↗

also on hashnode ↗

sep 09

shipped

3 estates → 1 brain

oncall-triage went live

triage.serhiykucherenko.dev · three bank estates feeding one ADK triage agent · aws · azure · kubernetes + kafka

sep 07

shipped

6 systems · 10 questions

Six-system RAG comparison harness

payments-rag vs file_search, NotebookLM, Haystack, LlamaIndex, LangChain · one adapter, one golden set

sep 07

wrote

4th on accuracy, 1st on cost and speed

Comparing RAGs, Part 2: the benchmark

comparing rags · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

sep 04

wrote

6 systems, 1 golden set

Comparing RAGs, Part 1: the framework

comparing rags · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

aug 21

wrote

65,930 → 1,948 words (34x)

How to hurt yourself with AI

working with ai · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

aug 14

wrote

refusal 0 · fabrication 0

Judge strictly to avoid hallucinations

judging answers · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

aug 09

wrote

495 → 484 chunks, 0.60 → 0.60

The illusion of improvement

measuring a rag · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

jul 31

wrote

4.9 ms of 2,823 ms

Three small lessons from building a RAG by hand

measuring a rag · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

jul 24

wrote

When 60% recall meant 90% accuracy: a RAG measurement story

measuring a rag · read on linkedin ↗

also on dev.to ↗ · hashnode ↗

jul 20

shipped

payments-rag demo went live

rag.serhiykucherenko.dev · fly.io · pgvector + claude

jul 19

milestone

100/100

Site launched, built static

cloudflare · lighthouse 100 · 0 trackers

jul 17

wrote

10,137 ms → 27 ms

The localhost trap: a 10-second database connection on Windows

working with ai · read on linkedin ↗

also on dev.to ↗ · hashnode ↗