Skills embedding-strategies
📦

embedding-strategies

Low Risk

Optimize Embedding Models for Semantic Search

Choosing the right embedding model and chunking strategy is critical for retrieval quality. This skill provides templates and best practices for implementing high-quality vector search pipelines.

Supports: Claude Codex Code(CC)
📊 69 Adequate
1

Download the skill ZIP

2

Upload in Claude

Go to Settings → Capabilities → Skills → Upload skill

3

Toggle on and start using

Test it

Using "embedding-strategies". Recommend an embedding model for a legal document search system. I need high accuracy and can use API services.

Expected outcome:

  • Recommended: text-embedding-3-large (3072 dimensions) or voyage-2 (1024 dimensions)
  • text-embedding-3-large: Best accuracy, handles 8191 tokens, ideal for long legal clauses
  • voyage-2: Specialized for legal/code, 1024 dimensions, 4000 token limit
  • Consider chunking legal documents by section headers to preserve clause context

Using "embedding-strategies". How should I chunk my technical documentation for a RAG system?

Expected outcome:

  • Strategy: Use semantic chunking by headers combined with recursive character splitting
  • Recommended chunk size: 512 tokens with 50 token overlap
  • Preserve code examples as complete chunks
  • Add context metadata linking chunks to original sections

Security Audit

Low Risk
v5 • 1/21/2026

All static findings are false positives. C2 keyword alerts triggered by hash hex strings. Weak crypto alerts from hash substrings. External command alerts from ASCII flow diagrams using arrows. Hardcoded URL alerts are legitimate documentation links. No malicious code, command execution, or data exfiltration patterns found.

2
Files scanned
818
Lines analyzed
0
findings
5
Total audits
No security issues found
Audited by: claude View Audit History →

Quality Score

38
Architecture
100
Maintainability
87
Content
29
Community
90
Security
91
Spec Compliance

What You Can Build

Build RAG Systems

Implement retrieval-augmented generation by selecting appropriate embedding models and chunking strategies for your document corpus.

Optimize Semantic Search

Improve search relevance by choosing embedding models matched to your content type and implementing proper chunking and preprocessing.

Create Embedding Pipelines

Build scalable pipelines that process documents, chunk content, generate embeddings, and prepare records for vector databases.

Try These Prompts

Select Embedding Model 
I need to choose an embedding model for my [use case: code search / multilingual documents / legal contracts]. My priorities are [priority: accuracy / cost / speed]. I have [constraints: limit on dimensions / need open source / need API access]. Recommend 3 models with rationale.
Implement Chunking Strategy 
Help me implement chunking for my [data type: technical documentation / conversational data / code]. I need to handle [requirement: preserve context / maintain semantic boundaries / limit chunk size]. Provide Python code for [strategy: token-based / sentence-based / recursive character] chunking.
Build Embedding Pipeline 
Create a Python pipeline that [input: processes documents from source / generates embeddings / stores in vector database]. Include [feature: batching / progress tracking / metadata handling]. Use [model: OpenAI embeddings / sentence-transformers].
Evaluate Retrieval Quality 
My embedding-based retrieval has [problem: low recall / inconsistent results / poor precision]. My setup uses [model details]. Analyze potential causes and suggest improvements for [metric: precision at k / recall / ndcg].

Best Practices

  • Match embedding model to content type: code, prose, or multilingual
  • Normalize embeddings for reliable cosine similarity comparisons
  • Use token overlap when chunking to preserve context across boundaries

Avoid

  • Mixing different embedding models in the same index
  • Ignoring token limits and truncating content mid-thought
  • Skipping preprocessing, allowing noise to degrade embedding quality

Frequently Asked Questions

What embedding model should I start with?
Start with text-embedding-3-small for general use. It balances cost and quality. Switch to text-embedding-3-large if you need higher accuracy, or voyage-2 for code and legal content.
How do I choose chunk size?
512 tokens is a good starting point for most use cases. Adjust based on your content complexity and model token limits. Overlap by 50 tokens to maintain context across chunks.
Can I use local embedding models?
Yes. Sentence-transformers supports models like BAAI/bge-large-en-v1.5 and intfloat/multilingual-e5-large. These run locally and work well for open-source or offline scenarios.
How do I evaluate my embedding quality?
Use precision@k, recall@k, MRR, and NDCG@k metrics. Test with known relevant documents and compare retrieved results against ground truth.
Should I normalize embeddings?
Yes. Normalize embeddings before using cosine similarity. Most modern embedding models produce normalized vectors by default, but local models may require explicit normalization.
What preprocessing should I apply?
Remove excessive whitespace, normalize unicode characters, and filter special characters. Keep content semantically meaningful. Domain-specific cleaning may be needed for code or structured data.

Developer Details

Author

wshobson

License

MIT

Repository

https://github.com/wshobson/agents/tree/main/plugins/llm-application-dev/skills/embedding-strategies

Ref

main

File structure

📄 SKILL.md