vector-index-tuning Skill for Claude: What It Does and How to Install It

Makes Claude tune vector indexes for speed, recall and memory: HNSW settings, quantization and Qdrant config.

Installs: 10.9K; on skills.sh
Stars: 42.1K; on GitHub
Files: 1; besides SKILL.md
Updated: Oct 2026; last repo commit

This skill is an index tuning guide for Claude, from wshobson's agents collection. SKILL.md maps data size to an index type, from flat search under 10K vectors to IVF with PQ or DiskANN above 100M. It explains the HNSW settings M, efConstruction and efSearch (defaults 16, 100 and 50) and compares quantization levels from FP32 to binary.

references/details.md has four Python templates: HNSW parameter tuning, quantization strategies, Qdrant index configuration, and performance monitoring. It loads when you tune HNSW, cut memory, reduce search latency or balance recall against speed.

Good for

Engineers running vector search in production
Teams cutting memory use with quantization
Developers balancing recall against latency
Anyone planning for very large vector sets

Install vector-index-tuning

Claude Code

Copy this prompt.

Install the vector-index-tuning skill in Claude Code. Run this command:

/plugin marketplace add wshobson/agents
/plugin install llm-application-dev@claude-code-workflows

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Pick one route. The plugin routes install the whole llm-application-dev plugin: 3 agents, 3 commands and 8 skills (counted from the plugin folder in the repo tree, which has no hooks or MCP config files). The skills-only routes install just this skill, with no agents or commands. The skill needs Python and a vector index of your own. The templates use numpy and Qdrant.

From the project README. Source.

What's in the skill

When it loads

Optimize vector index performance for latency, recall, and memory. Use when tuning HNSW parameters, selecting quantization strategies, or scaling vector search infrastructure.

Sections in SKILL.md

When to Use This Skill
Core Concepts
1. Index Type Selection
2. HNSW Parameters
3. Quantization Types
Templates and detailed worked examples
Best Practices
Do's
Don'ts

Key facts

Author: wshobson (repo owner)
License: MIT
Installs on skills.sh: 10,871 (Nov 2, 2026)
GitHub stars: 42,081 for the whole wshobson/agents repo (Nov 2, 2026)
Last repo commit: Oct 30, 2026
What it includes: 2 files including SKILL.md: references/details.md (no scripts)
Needs: Python and a vector index of your own. The templates import numpy, and one uses the Qdrant client
Part of: The llm-application-dev plugin (version 2.0.7) in wshobson/agents. Counted from the plugin folder: 3 agents, 3 commands and 8 skills
Works in: Claude Code, Codex CLI, Cursor, OpenCode, Antigravity CLI, GitHub Copilot and Pi, per the README. Skills-only installers ask which tool you use

What it includes

Flat, HNSW, HNSW plus quantization, IVF+PQ or DiskANN by data size
A size ladder: under 10K vectors use flat, 10K to 1M use HNSW, 1M to 100M add quantization, above 100M use IVF+PQ or DiskANN.
Try: “Which index fits 20 million vectors?”
Read · Index type selection
What M, efConstruction and efSearch change
A table of the three settings with their defaults and the trade-off each makes between recall, memory, build time and search speed.
Read · HNSW parameters
FP32, FP16, INT8, product and binary compared
Shows the bytes per vector for each level so you can estimate memory savings.
Read · Quantization types
Benchmark HNSW settings and recommend values (details.md)
Python functions that benchmark parameters, compute recall and suggest HNSW settings.
Write · Template 1: HNSW tuning
Scalar INT8, product and binary quantizers (details.md)
A VectorQuantizer class plus a memory estimate function.
Write · Template 2: Quantization
Create an optimized collection and tune search (details.md)
Functions that create a Qdrant collection with tuned settings and adjust search parameters.
Write · Template 3: Qdrant config
Search metrics, a monitor class and index build profiling (details.md)
Measures p50, p95 and p99 latency, queries per second and recall, and profiles how long an index takes to build.
Write · Template 4: Monitoring

From the official docs

Permissions

Low risk

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

The Qdrant template creates a collection
details.md calls client.create_collection inside create_optimized_collection. Run it against a test Qdrant instance before you point it at production. github.com
The plugin route installs more than this skill
The llm-application-dev plugin has 3 agents, 3 commands and 8 skills. We found no hooks or MCP config files in its folder. The skills-only routes install only this skill. github.com

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details

Turning on read-only	No read-only mode. Use version control.
Talks to	Your Qdrant server, if you run the Qdrant template
Sign-in and keys	SKILL.md and details.md do not mention any account or key.
Scopes	Claude can write benchmarking and tuning code in your project when you ask The Qdrant template creates a collection with create_collection

Questions

What does the vector-index-tuning skill do?

It gives Claude a tuning guide for vector indexes. Claude uses it to choose an index type, set HNSW parameters, pick a quantization level and measure recall and latency.

How do I install vector-index-tuning?

Use the install box at the top of this page. Pick a skills-only route to get just this skill, or the plugin route to get the whole llm-application-dev plugin.

Is this skill from Anthropic?

No. It is in wshobson/agents, a community collection. The skill sits in the llm-application-dev plugin there.

Should I install the whole llm-application-dev plugin?

Only if you want the rest. The plugin folder holds 3 agents, 3 commands and 8 skills, and the repo tree shows no hooks or MCP config files in it.

How is it different from similarity-search-patterns?

This skill tunes an index you already have: HNSW settings, quantization and monitoring. similarity-search-patterns, a sibling in the same plugin, covers building the search layer on Pinecone, Qdrant, pgvector or Weaviate. Neither SKILL.md names the other.

Does it work in the Claude app?

The install box has no Claude app upload for it, because the skill is built around Python code that runs in your own project. Use Claude Code or Codex. The README also lists Cursor, OpenCode, Antigravity CLI, GitHub Copilot and Pi.

Is it safe?

It bundles no scripts: just SKILL.md and one reference file. When you ask for code, Claude writes it in your project. One template creates a Qdrant collection, so point it at a test instance first (our advice).

Information

updated Nov 2, 2026

Repository: wshobson/agents
Folder: plugins/llm-application-dev/skills/vector-index-tuning

wshobson/agents skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 4

vector-index-tuning SKILL.md · github.com
wshobson/agents README · github.com
llm-application-dev plugin.json · github.com
vector-index-tuning references/details.md · github.com