Compétences rag-engineer
📦

rag-engineer

Sûr

Build Production-Ready RAG Systems

RAG systems often fail due to poor chunking strategies and naive retrieval. This skill provides expert patterns for semantic chunking, hybrid search, and context optimization to build production-grade retrieval systems.

Prend en charge: Claude Codex Code(CC)
⚠️ 67 Médiocre
1

Télécharger le ZIP du skill

2

Importer dans Claude

Allez dans Paramètres → Capacités → Skills → Importer un skill

3

Activez et commencez Ă  utiliser

Tester

Utilisation de "rag-engineer". How should I chunk PDF documents with headers and tables for a RAG system?

RĂ©sultat attendu:

Use semantic chunking that respects document structure: (1) Parse PDF to extract headers, paragraphs, and table boundaries, (2) Chunk at paragraph level while preserving header metadata, (3) Include 10-20% overlap between chunks for context continuity, (4) Store structural metadata for filtering during retrieval. This approach maintains meaning better than fixed-token chunking.

Utilisation de "rag-engineer". My RAG system returns relevant documents but the answers are still poor. How do I debug this?

RĂ©sultat attendu:

Separate retrieval evaluation from generation: (1) Check retrieval precision by inspecting top-k results manually, (2) Measure hit rate and MRR on a test set, (3) If retrieval looks good, the issue is in your prompt or LLM selection. Common fix: add a reranking step between initial retrieval and final generation.

Audit de sécurité

Sûr
v1 • 2/24/2026

All 16 static findings evaluated as false positives. The skill is a legitimate RAG (Retrieval-Augmented Generation) engineering documentation skill. External commands flagged are markdown code fence syntax (backticks) showing JavaScript examples. Cryptographic and reconnaissance warnings match keywords in context of semantic search and data pipelines, not security-relevant code.

1
Fichiers analysés
95
Lignes analysées
3
résultats
1
Total des audits

Problèmes critiques (3)

SKILL.md:38-91
External Commands - False Positive
Markdown code fence syntax (backticks) containing JavaScript code examples. Lines 38-91 contain documentation examples for semantic chunking, hierarchical retrieval, and hybrid search implementations. No actual shell or Ruby execution.
SKILL.md:3SKILL.md:21SKILL.md:94
Weak Cryptographic Algorithm - False Positive
Scanner incorrectly matches keywords: 'generation' (in Retrieval-Augmented Generation), 'pipeline' (data pipeline), 'execute' (workflow). No cryptographic code present.
SKILL.md:24SKILL.md:57SKILL.md:81SKILL.md:87
System Reconnaissance - False Positive
Scanner flags 'search' in context of semantic search and hybrid search. This is vector search in RAG systems, not network reconnaissance.
Audité par: claude

Score de qualité

38
Architecture
100
Maintenabilité
87
Contenu
50
Communauté
55
Sécurité
100
Conformité aux spécifications

Ce que vous pouvez construire

Build a document Q&A system

Create a RAG-powered Q&A system that accurately retrieves relevant context from large document collections

Improve existing RAG accuracy

Diagnose and fix retrieval quality issues in an existing RAG implementation using advanced chunking and hybrid search

Design multi-hop reasoning system

Architect a system that can answer complex questions requiring information from multiple source documents

Essayez ces prompts

Basic RAG Pattern Query 
How should I chunk my [document type] for a RAG system? What strategies preserve meaning and context?
Hybrid Search Implementation 
Design a hybrid search system that combines BM25 keyword matching with vector similarity. How do I fuse the scores?
Retrieval Evaluation 
How do I evaluate my retrieval system separately from the LLM generation? What metrics should I track?
Context Window Optimization 
My retrieved context exceeds the LLM context window. How do I select the most relevant chunks while preserving coherence?

Bonnes pratiques

  • Always evaluate retrieval quality separately from generation using metrics like hit rate and MRR
  • Use semantic chunking that respects document structure rather than fixed token counts
  • Implement hybrid search to combine semantic understanding with exact keyword matching

Éviter

  • Using fixed-size chunking regardless of content structure - this breaks sentences and loses meaning
  • Embedding everything without filtering - increases cost and reduces relevance
  • Skipping retrieval evaluation - assuming generation issues are always LLM problems

Foire aux questions

What is semantic chunking?
Semantic chunking divides documents by meaning rather than arbitrary token counts. It uses embedding similarity to detect topic shifts and preserves document structure like headers and paragraphs.
When should I use hybrid search?
Use hybrid search when queries contain specific terms (names, codes, exact phrases) alongside semantic concepts. It combines BM25 for keyword matching with vector similarity for semantic understanding.
How do I evaluate RAG retrieval quality?
Use metrics like hit rate (relevant docs in top-k), mean reciprocal rank (ranking quality), and recall. Create a test set of queries with ground-truth relevant documents.
What is reciprocal rank fusion?
Reciprocal Rank Fusion (RRF) combines rankings from multiple retrieval methods. It converts ranks to scores using 1/(rank + k) and sums them, giving a unified ranking without score normalization.
How do I handle long documents that exceed context windows?
Use hierarchical retrieval: index at multiple granularities (paragraph, section, document), retrieve coarse results first, then fetch fine-grained chunks from relevant sections.
What embedding model should I use?
Choose based on your use case. General-purpose models work for most cases. Specialized models exist for code, scientific text, or multilingual content. Always test on your specific data.

Détails du développeur

Auteur

sickn33

Licence

MIT

DĂ©pĂ´t

https://github.com/sickn33/antigravity-awesome-skills/tree/main/skills/rag-engineer

RĂ©f

main

Structure de fichiers

đź“„ SKILL.md