Bm25

All Posts architecture (19)llm (17)develop (16)paper-reading (14)javascript (11)ai-agent (11)react (7)transformer (7)mcp (7)post (6)jekyll (5)ai (5)question (4)typescript (4)agent (4)github (3)react-native (3)nextjs (3)browser-automation (3)claude-code (3)rag (3)interview (2)frontend (2)rails (2)html (2)css (2)git (2)bash (2)script (2)leetcode (2)openclaw (2)python (2)lightpanda (2)cdp (2)go (2)agent-browser (2)playwright (2)arxiv (2)math (2)bert (2)redis (2)backend (2)skills (2)coding-agent (2)tree-sitter (2)book (1)front-end (1)job (1)daily (1)algorithm (1)conference (1)angularjs (1)reactjs (1)django (1)gcp (1)ios (1)ml (1)keras (1)뉴스-머신러닝 (1)뉴스-크롤링 (1)빅카인즈 (1)angular (1)지킬 (1)에러 (1)bundle (1)form (1)hook (1)test (1)product (1)toy-project (1)gatsby (1)github-blog (1)migration (1)infra (1)domain (1)aws (1)route53 (1)아임웹 (1)postgresql (1)orm (1)error (1)civic-hacking (1)firebase (1)problemsolving (1)bit-manipulation (1)ruby (1)til (1)langchain (1)zig (1)browser (1)headless (1)v8 (1)nanoclaw (1)claude (1)ollama (1)llamacpp (1)local-ai (1)rust (1)vercel (1)mermaid (1)contentlayer (1)rehype (1)blog (1)beads (1)issue-tracker (1)distributed-system (1)dolt (1)orchestration (1)paperclip (1)open-source (1)comparison (1)loss (1)cross-entropy (1)perplexity (1)softmax (1)probability (1)vector (1)dot-product (1)moc (1)research (1)e2e (1)testing (1)attention (1)qkv (1)residual (1)layernorm (1)ffn (1)bloom-filter (1)database (1)scaling (1)game-dev (1)encoder-only (1)decoder-only (1)rnn (1)sequence-modeling (1)masked-language-model (1)mlm (1)pre-training (1)fine-tuning (1)training (1)redisbloom (1)signup (1)superpowers (1)codex (1)encoder (1)decoder (1)agentmemory (1)memory (1)iii (1)hermes-agent (1)nousresearch (1)tools (1)gateway (1)dify (1)llm-app (1)workflow (1)plugin (1)flask (1)gpt-2 (1)zero-shot (1)openai (1)openhands (1)sandbox (1)fastapi (1)software-agent-sdk (1)ruflo (1)claude-flow (1)agentdb (1)swarm (1)weknora (1)tencent (1)react-agent (1)wiki (1)knowledge-base (1)golang (1)codegraph (1)code-intelligence (1)knowledge-graph (1)static-analysis (1)sqlite (1)semble (1)code-search (1)hybrid-search (1)semantic-search (1)model2vec (1)static-embeddings (1)bm25 (1)

2026년 6월 3일

Semble 프로젝트 분석: 코드 검색을 정적 임베딩으로 푼 에이전트용 RAG는 어떻게 만들어졌나요?

semble code-search hybrid-search semantic-search model2vec static-embeddings bm25 rag tree-sitter mcp architecture

Semble은 tree-sitter로 코드를 청크로 자르고, Model2Vec 정적 임베딩과 BM25를 RRF로 융합한 뒤 코드 인지 리랭킹을 얹어, CPU에서 밀리초 만에 코드 검색을 돌리는 Python 라이브러리입니다. CodeGraph가 AST 지식 그래프로 같은 문제를 풀었다면, Semble은 검색(retrieval)으로 풉니다. 두 접근을 대조하며 구조를 분석합니다.