51 lines
1.8 KiB
Markdown
51 lines
1.8 KiB
Markdown
---
|
|
name: vector-index-tuning
|
|
description: "Optimize vector index performance for latency, recall, and memory. Use when tuning HNSW parameters, selecting quantization strategies, or scaling vector search infrastructure."
|
|
risk: safe
|
|
source: community
|
|
date_added: "2026-02-27"
|
|
---
|
|
|
|
# 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
|
|
|
|
1. Gather workload targets (latency, recall, QPS), data size, and memory budget.
|
|
2. Choose an index type and establish a baseline with default parameters.
|
|
3. Benchmark parameter sweeps using real queries and track recall, latency, and memory.
|
|
4. 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.md` for 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.
|