VectorDB/Embedding deep research 결과 정리 — KURE-v1 + Neo4j hybrid

위치

• Raw result: /home/tlswk/projects/maester-agent/context/deep_research_prompts/deep_research-gpt.md
• Decision memo: /home/tlswk/projects/maester-agent/context/deep_research/2026-06-19-vectordb-embedding-decision.md
• Handoff updated: /home/tlswk/projects/maester-agent/context/handoff/2026-06-18-mainai-hermes-vectorize.md

결정 요약

• Embedding first pick: nlpai-lab/KURE-v1
• Challenger: Qwen3-Embedding-0.6B
• Vector layer: keep Neo4j native vector index + fulltext hybrid retrieval
• Migration: do not overwrite Titan vectors; add embedding_kure_1024 shadow index, pilot 1k-2k chunks, evaluate with real Maester queries, then switch.

Caveat

Raw research citations are ChatGPT citation tokens, not direct URLs. Primary-source verification is still required before treating benchmark claims as final.