vector-index-tuning
Optimisez les performances des index vectoriels pour la latence, le rappel et la mémoire. Utilisez lors du tuning des paramètres HNSW, de la sélection de stratégies de quantification ou de la mise à l'échelle de l'infrastructure de recherche vectorielle.
Le contenu de ce skill est dans sa langue d’origine (souvent l’anglais).
Vector Index Tuning
Guide to optimizing vector indexes for production performance.
Use this skill when
- Tuning HNSW parameters
- Implementing quantization
- Optimizing memory usage
- Reducing search latency
- Balancing recall vs speed
- Scaling to billions of vectors
Do not use this skill when
- You only need exact search on small datasets (use a flat index)
- You lack workload metrics or ground truth to validate recall
- You need end-to-end retrieval system design beyond index tuning
Instructions
- Gather workload targets (latency, recall, QPS), data size, and memory budget.
- Choose an index type and establish a baseline with default parameters.
- Benchmark parameter sweeps using real queries and track recall, latency, and memory.
- Validate changes on a staging dataset before rolling out to production.
Refer to resources/implementation-playbook.md for detailed patterns, checklists, and templates.
Safety
- Avoid reindexing in production without a rollback plan.
- Validate changes under realistic load before applying globally.
- Track recall regressions and revert if quality drops.
Resources
resources/implementation-playbook.mdfor detailed patterns, checklists, and templates.
Limitations
- Use this skill only when the task clearly matches the scope described above.
- Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
- Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.