المهارات similarity-search-patterns
🔍

similarity-search-patterns

آمن

Build production similarity search with vector databases

متاح أيضًا من: wshobson

Vector similarity search is essential for semantic search and RAG systems but requires careful index selection and tuning. This skill provides ready-to-use patterns for Pinecone, Qdrant, pgvector, and Weaviate with hybrid search and reranking.

يدعم: Claude Codex Code(CC)
🥉 75 برونزي
1

تنزيل ZIP المهارة

2

رفع في Claude

اذهب إلى Settings → Capabilities → Skills → Upload skill

3

فعّل وابدأ الاستخدام

اختبرها

استخدام "similarity-search-patterns". Set up Qdrant for 1 million product embeddings with 95% recall target

النتيجة المتوقعة:

  • Collection created with HNSW index (m=16, ef_construct=64)
  • Scalar quantization enabled (INT8) reducing memory by 4x
  • Recommended search params: ef=128 for 95% recall at ~50ms P99

استخدام "similarity-search-patterns". Hybrid search for documents about 'machine learning optimization'

النتيجة المتوقعة:

  • Vector search retrieves 20 semantically similar documents
  • BM25 filters for exact term matches on 'optimization'
  • Reranked results combine both signals with 0.6/0.4 weighting

التدقيق الأمني

آمن
v1 • 2/25/2026

All 31 static analysis findings are false positives. The scanner incorrectly flagged Markdown documentation syntax as executable code. Backticks are code fence markers, URLs are documentation links, and API key parameters represent secure credential handling. The skill contains Python code examples for vector database implementations with no security concerns.

2
الملفات التي تم فحصها
594
الأسطر التي تم تحليلها
0
النتائج
1
إجمالي عمليات التدقيق
لا توجد مشكلات أمنية
تم تدقيقه بواسطة: claude

درجة الجودة

38
الهندسة المعمارية
100
قابلية الصيانة
87
المحتوى
50
المجتمع
100
الأمان
100
الامتثال للمواصفات

ماذا يمكنك بناءه

RAG Application Builder

Implement retrieval-augmented generation with efficient vector search over document embeddings. Supports hybrid search combining semantic and keyword matching for accurate context retrieval.

Recommendation Engine Developer

Build nearest-neighbor recommendation systems using vector embeddings. Leverage HNSW indexing for sub-second queries over millions of items with configurable recall thresholds.

Search Platform Engineer

Deploy production similarity search infrastructure with PostgreSQL pgvector or managed services. Implement pre-filtering and reranking pipelines for enterprise search quality.

جرّب هذه الموجهات

Basic Vector Store Setup 
I need to store and search document embeddings. Help me set up a Pinecone vector index with cosine similarity for 1536-dimensional embeddings. Include upsert and search functions with metadata filtering.
Hybrid Search Implementation 
Create a hybrid search function that combines vector similarity with keyword matching using pgvector. Weight vector search at 0.7 and full-text search at 0.3. Include the SQL schema with HNSW index.
Reranking Pipeline 
I have 50 vector search results but need better ranking. Show me how to rerank with a cross-encoder model. Include the sentence-transformers code and explain the tradeoff between latency and accuracy.
Index Selection and Tuning 
I have 10 million vectors and need P99 latency under 100ms with 95% recall. Recommend an index type and parameters for Qdrant. Explain ef_construct, nprobe, and quantization settings for my use case.

أفضل الممارسات

  • Start with flat index for small datasets (under 100K vectors) and migrate to HNSW as data grows
  • Always measure recall at your target latency before deploying approximate indexes
  • Use metadata pre-filtering to reduce search space before vector comparison

تجنب

  • Using approximate nearest neighbor indexes without measuring recall degradation
  • Storing raw vectors without quantization when memory is constrained
  • Querying without pre-filters causing full index scans on large collections

الأسئلة المتكررة

Which vector database should I choose for my project?
For quick prototypes, use Pinecone (managed, no infra). For PostgreSQL shops, pgvector adds minimal complexity. For high-performance needs, Qdrant or Weaviate offer advanced features. Consider data size, query latency requirements, and existing infrastructure.
What is the difference between cosine similarity and dot product?
Cosine similarity measures angle only (ignores magnitude), making it ideal for normalized embeddings. Dot product includes magnitude effects. For most embedding models (OpenAI, Sentence Transformers), cosine similarity is the standard choice.
How do I improve search recall without sacrificing latency?
Increase index parameters (ef_search for HNSW, nprobe for IVF) gradually while measuring P99 latency. Use reranking to recover precision after high-recall retrieval. Consider two-stage retrieval: fast approximate search followed by exact reranking on top-K results.
Can I update vectors after insertion?
Yes, all supported vector databases support upsert operations. Re-inserting a vector with the same ID updates it in place. Note that frequent updates may require index rebuilding for optimal performance.
What embedding dimension should I use?
Use the native dimension of your embedding model (e.g., 1536 for text-embedding-ada-002, 384 for all-MiniLM-L6-v2). Higher dimensions capture more nuance but increase storage and compute. Match your index metric to the embedding model's training objective.
How do I handle multilingual search?
Use multilingual embedding models like multilingual-e5 or LaBSE that map different languages to the same vector space. Query and documents in different languages will still retrieve relevant results if the embedding model supports them.

تفاصيل المطور

المؤلف

sickn33

الترخيص

MIT

المستودع

https://github.com/sickn33/antigravity-awesome-skills/tree/main/skills/similarity-search-patterns

مرجع

main

بنية الملفات

📁 resources/

📄 implementation-playbook.md

📄 SKILL.md