Blog
Engineering & Product Updates
Deep-dives on embedding infrastructure, vector migration, and building reliable AI retrieval systems.
RAG Is Not Vector Search
Vector search is step 3 of 8. Here's the full checklist for production RAG -- and what breaks when you skip steps.
우리 엔진이 Qdrant보다 8.7배 빠른 이유
Schift 벡터 엔진의 실측 벤치마크를 공개한다. 이기는 곳, 지는 곳 모두 투명하게. Apple M5 Pro, 1M vectors, 1024d 기준으로 Qdrant, FAISS, pgvector와 비교했다.
自作ベクトルエンジンをQdrant・FAISSと比較してみた（1M vectors, 1024d）
Rustで書いたベクトル検索エンジンSchiftのベンチマークを取ってみました。Qdrantに8.7倍勝ち、FAISS Flatには負けます。勝つところも負けるところも全部出します。
pgvectorの限界を1M vectorsで実測してみた話
pgvectorは小規模ベクトル検索に便利だが、どこから専用エンジンが必要になるのか。Rust製エンジン・FAISS・Qdrantと比較しながら、境界線を数字で示します。
TypeScriptでAI Agentフレームワークを作っている理由
AI agentツールはPythonばかり。でもproductionアプリはTypeScript。このギャップがつらいので、自分たちで作ることにした話。
pgvector Is Not a Vector Database (And That's Fine)
pgvector is a solid choice for adding vector search to Postgres at low scale. But when does it stop being enough? We ran the numbers.
Why We're Building an AI Agent Framework in TypeScript
The AI agent tooling ecosystem is dominated by Python. But production applications are TypeScript. We think that mismatch has a real cost, and we built something to close it.
Making SQ8 the Default for New Collections
Why the engine moved to SQ8 as the default storage format — what we measured, what failed, and what we are not doing yet.
FAISS에서 SQ8까지
벡터 검색 엔진의 기본 저장 포맷을 찾기까지. FAISS를 기준선으로 두고 F32, SQ8, SQ4, SQ1, TQ4를 비교한 개발 기록.
HyperbolicRAG를 바로 도입하지 않고, hierarchy-aware retrieval부터 검증하기
HyperbolicRAG의 문제의식은 유효하지만, 먼저 검증할 것은 구조 신호 기반 rerank. dense baseline에서 relation-aware rerank까지의 실험 기록.
법률 데이터를 Vector DB로 만들면 얼마나 작아지고 얼마나 빨라질까
한국 법률 코퍼스 기반 벡터 DB 벤치마크. SQ8 압축, 계층 탐색, 본문 조회까지 포함한 전체 파이프라인 성능 기록.
2026 AI바우처 수요기업 신청 가이드
AI 도입 비용의 최대 80%를 정부가 지원합니다. 신청 자격, 절차, 일정, 비용 구조를 정리했습니다.
The Embedding Failover Pattern: Zero Downtime Across Providers
When your embedding provider goes down, your search breaks. Learn the failover pattern that keeps retrieval alive across provider outages using projection matrices.
Case Study: Cutting Embedding Costs to $0 with Gemini
How a startup paying $1,500/month in OpenAI embedding costs migrated to Gemini Embedding in one afternoon — without re-embedding a single document.
Why Vector Migration Matters More Than You Think
Embedding model upgrades silently break production retrieval. Here is why vendor lock-in is a hidden technical debt — and what you can do about it today.