菜单
Version the vector index and reindex or migrate embeddings without downtime or quality regressions.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use this skill when building, debugging, or improving Retrieval-Augmented Generation systems, including chunking, vector database selection, hybrid search, reranking, multimodal RAG, code documentation RAG, retrieval latency, and production RAG architecture.
Chunk nested documents into parent-child levels so retrieval can move from broad sections to fine-grained passages.
Use semantic boundaries and embedding similarity to chunk text for higher-relevance retrieval.
Route RAG chunking decisions across semantic, hierarchical, sliding-window, contextual-header, and framework-selection strategies.
Use overlapping windows to preserve context across chunk boundaries while controlling retrieval size.
Reduce retrieval latency with caching, batching, and index-level optimization.
| name | index-versioning-and-reindexing |
| title | Index Versioning and Reindexing |
| description | Version the vector index and reindex or migrate embeddings without downtime or quality regressions. |
| allowed-tools | ["Read","Grep","Glob","Bash"] |
| category | deployment |
| tags | ["reindexing","versioning","embedding-migration","blue-green"] |
This skill covers evolving the index behind a live RAG system: rebuilding embeddings after a model or chunking change, migrating between embedding models, and versioning indexes so changes are reversible and queries never go dark.
The index is not static. Embedding models improve, chunking strategies change, and documents update. Reindexing in place takes search offline and, if the new build is worse, there is no way back. Mixing vectors from two embedding models silently corrupts similarity scores.
Create the new version as a separate collection or index while the current one keeps serving traffic.
Why: Building beside the live index means queries never hit a half-built or empty collection.
Run the evaluation golden set against the new index and compare scores to the current baseline before switching.
Why: Comparing against a baseline catches a regression before it reaches users instead of after.
Cut over by repointing an alias to the new version and retain the previous version until the new one is proven.
Why: An alias swap makes both cutover and rollback near-instant and low-risk.
Never serve queries from a mix of two embedding models: re-embed the whole corpus when the model changes. See the Qdrant aliases guide and Qdrant snapshots.
