What distance metrics does it cover?

Cosine, Euclidean (L2), dot product, and Manhattan (L1). Cosine suits normalized embeddings, Euclidean suits raw embeddings, dot product is for when magnitude matters, and Manhattan is for sparse vectors.

Which index types does it explain?

Flat (exact, O(n) search, 100% recall, small data), HNSW (graph-based, O(log n), roughly 95 to 99% recall, medium to large data), and IVF+PQ (quantized, O(sqrt n), roughly 90 to 95% recall, very large data).

How do I tune for better recall?

Adjust parameters such as ef_search and nprobe, and measure recall and latency before and after each change. The skill advises evaluating first rather than optimizing blindly.

Similarity Search Patterns

SkillSearch

A skill that guides an AI agent through implementing similarity search with vector databases. It covers distance metrics such as cosine and Euclidean, index types including flat, HNSW, and IVF+PQ, and the speed and accuracy tradeoffs between them. Use it when building semantic search, nearest neighbor queries, or retrieval performance tuning.

Connect your AI Use it in a pipeline

Then ask your AI: use the Similarity Search Patterns skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Account requirements not reviewed. Check the skill instructions before use; Ahel provides instructions and does not run this skill.

No other account needed.

Have a vector database and a set of embeddings available to search over.

What your AI can do with it

Choose a distance metric: cosine, Euclidean, dot product, or Manhattan
Compare flat, HNSW, and IVF+PQ index types and their tradeoffs
Tune parameters such as ef_search and nprobe for recall and speed
Combine keyword and semantic search in a hybrid setup
Monitor recall and latency to measure search quality

Getting started

Have a vector database and a set of embeddings available to search over.
Add the skill to your agent so it can read the patterns and reference templates.
Tell the agent what you are building, such as semantic search or a recommendation engine.
Have the agent pick a distance metric and index type for your data size and recall needs.
Measure recall and latency before tuning parameters or scaling the index.

What this skill tells your AI

The instructions your AI receives, as published by wshobson/agents in plugins/llm-application-dev/skills/similarity-search-patterns/SKILL.md and read by Ahel’s review.

Patterns for implementing efficient similarity search in production systems.

When to Use This Skill

Building semantic search systems
Implementing RAG retrieval
Creating recommendation engines
Optimizing search latency
Scaling to millions of vectors
Combining semantic and keyword search

Core Concepts

1. Distance Metrics

| Metric | Formula | Best For | | ------------------ | ------------------ | --------------------- | --- | -------------- | | Cosine | 1 - (A·B)/(‖A‖‖B‖) | Normalized embeddings | | Euclidean (L2) | √Σ(a-b)² | Raw embeddings | | Dot Product | A·B | Magnitude matters | | Manhattan (L1) | Σ | a-b | | Sparse vectors |

2. Index Types

┌─────────────────────────────────────────────────┐
│                 Index Types                      │
├─────────────┬───────────────┬───────────────────┤
│    Flat     │     HNSW      │    IVF+PQ         │
│ (Exact)     │ (Graph-based) │ (Quantized)       │
├─────────────┼───────────────┼───────────────────┤
│ O(n) search │ O(log n)      │ O(√n)             │
│ 100% recall │ ~95-99%       │ ~90-95%           │
│ Small data  │ Medium-Large  │ Very Large        │
└─────────────┴───────────────┴───────────────────┘

Templates and detailed worked examples

Full template library and detailed worked examples live in references/details.md. Read that file when you need the concrete templates.

Best Practices

Do's

Use appropriate index - HNSW for most cases
Tune parameters - ef_search, nprobe for recall/speed
Implement hybrid search - Combine with keyword search
Monitor recall - Measure search quality
Pre-filter when possible - Reduce search space

Don'ts

Don't skip evaluation - Measure before optimizing
Don't over-index - Start with flat, scale up
Don't ignore latency - P99 matters for UX
Don't forget costs - Vector storage adds up

Signals

GitHub stars: 40k
Forks: 4k
Last commit: Oct 2026

Questions

What distance metrics does it cover?: Cosine, Euclidean (L2), dot product, and Manhattan (L1). Cosine suits normalized embeddings, Euclidean suits raw embeddings, dot product is for when magnitude matters, and Manhattan is for sparse vectors.
Which index types does it explain?: Flat (exact, O(n) search, 100% recall, small data), HNSW (graph-based, O(log n), roughly 95 to 99% recall, medium to large data), and IVF+PQ (quantized, O(sqrt n), roughly 90 to 95% recall, very large data).
How do I tune for better recall?: Adjust parameters such as ef_search and nprobe, and measure recall and latency before and after each change. The skill advises evaluating first rather than optimizing blindly.

Advanced

Item type: skill
Key: similarity-search-patterns-wshobson
Source: github.com/wshobson/agents

github.com/wshobson/agents