| name | Vector Databases |
| skill_id | vector-databases |
| description | Guides retrieval-store design, indexing, and query behavior for embedding-backed systems without confusing storage with application truth. |
| version | 1.0.0 |
| source_kind | elephant-builtin |
| aliases | ["vector db","vector database","embedding store"] |
| trigger_phrases | ["set up a vector database","design the retrieval store","index embeddings for search"] |
| keywords | ["vector database","retrieval","embeddings","index","chunking","filtering"] |
| category | mlops |
Vector Databases
Use this built-in skill when the user needs an embedding-backed retrieval store, semantic search index, or hybrid filtering design.
Core rules
- Define the retrieval task before selecting an index or vendor.
- Make chunking, metadata, freshness, and filter behavior explicit.
- Keep canonical application truth outside the vector index unless the system is designed that way.
- Evaluate retrieval quality, not just index build success.
Default workflow
- Identify the query patterns, corpus shape, and update frequency.
- Choose the embedding and indexing strategy that matches that workload.
- Design metadata filters, ids, and refresh behavior intentionally.
- Test retrieval quality with representative queries before scaling up.
Guardrails
- Do not equate a working index build with a good retrieval system.
- Do not hide stale-data or reindex costs.
- Do not use the vector store as a vague substitute for product data modeling.