Rag Architect | Skills Pool
Rag Architect Use when the user asks to design RAG pipelines, optimize retrieval strategies, choose embedding models, implement vector search, or build knowledge retrieval systems.
alirezarezvani 11,734 estrellas 23 mar 2026 Ocupación Categorías Patrones de Arquitectura Contenido de la habilidad
RAG Architect - POWERFUL
Overview
The RAG (Retrieval-Augmented Generation) Architect skill provides comprehensive tools and knowledge for designing, implementing, and optimizing production-grade RAG pipelines. This skill covers the entire RAG ecosystem from document chunking strategies to evaluation frameworks, enabling you to build scalable, efficient, and accurate retrieval systems.
Core Competencies
1. Document Processing & Chunking Strategies
Fixed-Size Chunking
Character-based chunking : Simple splitting by character count (e.g., 512, 1024, 2048 chars)Token-based chunking : Splitting by token count to respect model limitsOverlap strategies : 10-20% overlap to maintain context continuityPros : Predictable chunk sizes, simple implementation, consistent processing time
Instalación rápida
Rag Architect npx skills add alirezarezvani/claude-skills
estrellas 11,734
Actualizado 23 mar 2026
Ocupación
Cons : May break semantic units, context boundaries ignored
Best for : Uniform documents, when consistent chunk sizes are critical
Sentence-Based Chunking
Sentence boundary detection : Using NLTK, spaCy, or regex patternsSentence grouping : Combining sentences until size threshold is reachedParagraph preservation : Avoiding mid-paragraph splits when possiblePros : Preserves natural language boundaries, better readabilityCons : Variable chunk sizes, potential for very short/long chunksBest for : Narrative text, articles, books
Paragraph-Based Chunking
Paragraph detection : Double newlines, HTML tags, markdown formattingHierarchical splitting : Respecting document structure (sections, subsections)Size balancing : Merging small paragraphs, splitting large onesPros : Preserves logical document structure, maintains topic coherenceCons : Highly variable sizes, may create very large chunksBest for : Structured documents, technical documentation
Semantic Chunking
Topic modeling : Using TF-IDF, embeddings similarity for topic detectionHeading-aware splitting : Respecting document hierarchy (H1, H2, H3)Content-based boundaries : Detecting topic shifts using semantic similarityPros : Maintains semantic coherence, respects document structureCons : Complex implementation, computationally expensiveBest for : Long-form content, technical manuals, research papers
Recursive Chunking
Hierarchical approach : Try larger chunks first, recursively split if neededMulti-level splitting : Different strategies at different levelsSize optimization : Minimize number of chunks while respecting size limitsPros : Optimal chunk utilization, preserves context when possibleCons : Complex logic, potential performance overheadBest for : Mixed content types, when chunk count optimization is important
Document-Aware Chunking
File type detection : PDF pages, Word sections, HTML elementsMetadata preservation : Headers, footers, page numbers, sectionsTable and image handling : Special processing for non-text elementsPros : Preserves document structure and metadataCons : Format-specific implementation requiredBest for : Multi-format document collections, when metadata is important
2. Embedding Model Selection
Dimension Considerations
128-256 dimensions : Fast retrieval, lower memory usage, suitable for simple domains512-768 dimensions : Balanced performance, good for most applications1024-1536 dimensions : High quality, better for complex domains, higher cost2048+ dimensions : Maximum quality, specialized use cases, significant resources
Speed vs Quality Tradeoffs
Fast models : sentence-transformers/all-MiniLM-L6-v2 (384 dim, ~14k tokens/sec)Balanced models : sentence-transformers/all-mpnet-base-v2 (768 dim, ~2.8k tokens/sec)Quality models : text-embedding-ada-002 (1536 dim, OpenAI API)Specialized models : Domain-specific fine-tuned models
Model Categories
General purpose : all-MiniLM, all-mpnet, Universal Sentence EncoderCode embeddings : CodeBERT, GraphCodeBERT, CodeT5Scientific text : SciBERT, BioBERT, ClinicalBERTMultilingual : LaBSE, multilingual-e5, paraphrase-multilingual
3. Vector Database Selection
Pinecone
Managed service : Fully hosted, auto-scalingFeatures : Metadata filtering, hybrid search, real-time updatesPricing : $70/month for 1M vectors (1536 dim), pay-per-use scalingBest for : Production applications, when managed service is preferredCons : Vendor lock-in, costs can scale quickly
Weaviate
Open source : Self-hosted or cloud options availableFeatures : GraphQL API, multi-modal search, automatic vectorizationScaling : Horizontal scaling, HNSW indexingBest for : Complex data types, when GraphQL API is preferredCons : Learning curve, requires infrastructure management
Qdrant
Rust-based : High performance, low memory footprintFeatures : Payload filtering, clustering, distributed deploymentAPI : REST and gRPC interfacesBest for : High-performance requirements, resource-constrained environmentsCons : Smaller community, fewer integrations
Chroma
Embedded database : SQLite-based, easy local developmentFeatures : Collections, metadata filtering, persistenceScaling : Limited, suitable for prototyping and small deploymentsBest for : Development, testing, small-scale applicationsCons : Not suitable for production scale
pgvector (PostgreSQL)
SQL integration : Leverage existing PostgreSQL infrastructureFeatures : ACID compliance, joins with relational data, mature ecosystemPerformance : ivfflat and HNSW indexing, parallel query processingBest for : When you already use PostgreSQL, need ACID complianceCons : Requires PostgreSQL expertise, less specialized than purpose-built DBs
4. Retrieval Strategies
Dense Retrieval
Semantic similarity : Using embedding cosine similarityAdvantages : Captures semantic meaning, handles paraphrasing wellLimitations : May miss exact keyword matches, requires good embeddingsImplementation : Vector similarity search with k-NN or ANN algorithms
Sparse Retrieval
Keyword-based : TF-IDF, BM25, ElasticsearchAdvantages : Exact keyword matching, interpretable resultsLimitations : Misses semantic similarity, vulnerable to vocabulary mismatchImplementation : Inverted indexes, term frequency analysis
Hybrid Retrieval
Combination approach : Dense + sparse retrieval with score fusionFusion strategies : Reciprocal Rank Fusion (RRF), weighted combinationBenefits : Combines semantic understanding with exact matchingComplexity : Requires tuning fusion weights, more complex infrastructure
Reranking
Two-stage approach : Initial retrieval followed by rerankingReranking models : Cross-encoders, specialized reranking transformersBenefits : Higher precision, can use more sophisticated models for final rankingTradeoff : Additional latency, computational cost
HyDE (Hypothetical Document Embeddings)
Approach : Generate hypothetical answer, embed answer instead of queryBenefits : Improves retrieval by matching document style rather than query styleImplementation : Use LLM to generate hypothetical document, embed thatUse cases : When queries and documents have different styles
Multi-Query Generation
Approach : Generate multiple query variations, retrieve for each, merge resultsBenefits : Increases recall, handles query ambiguityImplementation : LLM generates 3-5 query variations, deduplicate resultsConsiderations : Higher cost and latency due to multiple retrievals
Step-Back Prompting
Approach : Generate broader, more general version of specific queryBenefits : Retrieves more general context that helps answer specific questionsImplementation : Transform "What is the capital of France?" to "What are European capitals?"Use cases : When specific questions need general context
6. Context Window Optimization
Dynamic Context Assembly
Relevance-based ordering : Most relevant chunks firstDiversity optimization : Avoid redundant informationToken budget management : Fit within model context limitsHierarchical inclusion : Include summaries before detailed chunks
Context Compression
Summarization : Compress less relevant chunks while preserving key informationKey information extraction : Extract only relevant facts/entitiesTemplate-based compression : Use structured formats to reduce token usageSelective inclusion : Include only chunks above relevance threshold
7. Evaluation Frameworks
Faithfulness Metrics
Definition : How well generated answers are grounded in retrieved contextMeasurement : Fact verification against source documentsImplementation : NLI models to check entailment between answer and contextThreshold : >90% for production systems
Relevance Metrics
Context relevance : How relevant retrieved chunks are to the queryAnswer relevance : How well the answer addresses the original questionMeasurement : Embedding similarity, human evaluation, LLM-as-judgeTargets : Context relevance >0.8, Answer relevance >0.85
Context Precision & Recall
Precision@K : Percentage of top-K results that are relevantRecall@K : Percentage of relevant documents found in top-K resultsMean Reciprocal Rank (MRR) : Average of reciprocal ranks of first relevant resultNDCG@K : Normalized Discounted Cumulative Gain at K
End-to-End Metrics
RAGAS : Comprehensive RAG evaluation frameworkCorrectness : Factual accuracy of generated answersCompleteness : Coverage of all relevant aspectsConsistency : Consistency across multiple runs with same query
8. Production Patterns
Caching Strategies
Query-level caching : Cache results for identical queriesSemantic caching : Cache for semantically similar queriesChunk-level caching : Cache embedding computationsMulti-level caching : Redis for hot queries, disk for warm queries
Streaming Retrieval
Progressive loading : Stream results as they become availableIncremental generation : Generate answers while still retrievingReal-time updates : Handle document updates without full reprocessingConnection management : Handle client disconnections gracefully
Fallback Mechanisms
Graceful degradation : Fallback to simpler retrieval if primary failsCache fallbacks : Serve stale results when retrieval is unavailableAlternative sources : Multiple vector databases for redundancyError handling : Comprehensive error recovery and user communication
9. Cost Optimization
Embedding Cost Management
Batch processing : Batch documents for embedding to reduce API costsCaching strategies : Cache embeddings to avoid recomputationModel selection : Balance cost vs quality for embedding modelsUpdate optimization : Only re-embed changed documents
Vector Database Optimization
Index optimization : Choose appropriate index types for use caseCompression : Use quantization to reduce storage costsTiered storage : Hot/warm/cold data strategiesResource scaling : Auto-scaling based on query patterns
Query Optimization
Query routing : Route simple queries to cheaper methodsResult caching : Avoid repeated expensive retrievalsBatch querying : Process multiple queries together when possibleSmart filtering : Use metadata filters to reduce search space
10. Guardrails & Safety
Content Filtering
Toxicity detection : Filter harmful or inappropriate contentPII detection : Identify and handle personally identifiable informationContent validation : Ensure retrieved content meets quality standardsSource verification : Validate document authenticity and reliability
Query Safety
Injection prevention : Prevent malicious query injection attacksRate limiting : Prevent abuse and ensure fair usageQuery validation : Sanitize and validate user inputsAccess controls : Ensure users can only access authorized content
Response Safety
Hallucination detection : Identify when model generates unsupported claimsConfidence scoring : Provide confidence levels for generated responsesSource attribution : Always provide sources for factual claimsUncertainty handling : Gracefully handle cases where answer is uncertain
Implementation Best Practices
Development Workflow
Requirements gathering : Understand use case, scale, and quality requirementsData analysis : Analyze document corpus characteristicsPrototype development : Build minimal viable RAG pipelineChunking optimization : Test different chunking strategiesRetrieval tuning : Optimize retrieval parameters and thresholdsEvaluation setup : Implement comprehensive evaluation metricsProduction deployment : Scale-ready implementation with monitoring
Monitoring & Observability
Query analytics : Track query patterns and performanceRetrieval metrics : Monitor precision, recall, and latencyGeneration quality : Track faithfulness and relevance scoresSystem health : Monitor database performance and availabilityCost tracking : Monitor embedding and vector database costs
Maintenance & Updates
Document refresh : Handle new documents and updatesIndex maintenance : Regular vector database optimizationModel updates : Evaluate and migrate to improved modelsPerformance tuning : Continuous optimization based on usage patternsSecurity updates : Regular security assessments and updates
Common Pitfalls & Solutions
Poor Chunking Strategy
Problem : Chunks break mid-sentence or lose contextSolution : Use boundary-aware chunking with overlap
Low Retrieval Precision
Problem : Retrieved chunks are not relevant to querySolution : Improve embedding model, add reranking, tune similarity threshold
High Latency
Problem : Slow retrieval and generationSolution : Optimize vector indexing, implement caching, use faster embedding models
Inconsistent Quality
Problem : Variable answer quality across different queriesSolution : Implement comprehensive evaluation, add quality scoring, improve fallbacks
Scalability Issues
Problem : System doesn't scale with increased loadSolution : Implement proper caching, database sharding, and auto-scaling
Conclusion Building effective RAG systems requires careful consideration of each component in the pipeline. The key to success is understanding the tradeoffs between different approaches and choosing the right combination of techniques for your specific use case. Start with simple approaches and gradually add sophistication based on evaluation results and production requirements.
This skill provides the foundation for making informed decisions throughout the RAG development lifecycle, from initial design to production deployment and ongoing maintenance.38:["$","$L40",null,{"content":"$41","frontMatter":{"name":"rag-architect","description":"Use when the user asks to design RAG pipelines, optimize retrieval strategies, choose embedding models, implement vector search, or build knowledge retrieval systems."}}]
02
Overview