Rag Architect

What This Skill Does

The RAG Architect skill provides tools and knowledge for designing, implementing, and optimizing production-grade RAG pipelines. It covers document chunking strategies and evaluation frameworks. This skill is designed to build scalable, efficient, and accurate retrieval systems.

When to Use

• Implement fixed-size chunking.
• Implement sentence-based chunking.
• Implement paragraph-based chunking.
• Implement semantic chunking.

Key Features

Document processing & chunking strategies. Fixed-Size ChunkingSentence-Based ChunkingParagraph-Based ChunkingSemantic Chunking

Manual Installation

Manual installationView Full Skill ContentThe complete markdown content that gets installedRAG 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 limits Overlap strategies: 10-20% overlap to maintain context continuity Pros: Predictable chunk sizes, simple implementation, consistent processing time 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 patterns Sentence grouping: Combining sentences until size threshold is reached Paragraph preservation: Avoiding mid-paragraph splits when possible Pros: Preserves natural language boundaries, better readability Cons: Variable chunk sizes, potential for very short/long chunks Best for: Narrative text, articles, books

Paragraph-Based Chunking

Paragraph detection: Double newlines, HTML tags, markdown formatting Hierarchical splitting: Respecting document structure (sections, subsections) Size balancing: Merging small paragraphs, splitting large ones Pros: Preserves logical document structure, maintains topic coherence Cons: Highly variable sizes, may create very large chunks Best for: Structured documents, technical documentation

Semantic Chunking

Topic modeling: Using TF-IDF, embeddings similarity for topic detection Heading-aware splitting: Respecting document hierarchy (H1, H2, H3) Content-based boundaries: Detecting topic shifts using semantic similarity Pros: Maintains semantic coherence, respects document structure Cons: Complex implementation, computationally expensive Best for: Long-form content, technical manuals, research papers

Recursive Chunking

Hierarchical approach: Try larger chunks first, recursively split if needed Multi-level splitting: Different strategies at different levels Size optimization: Minimize number of chunks while respecting size limits Pros: Optimal chunk utilization, preserves context when possible Cons: Complex logic, potential performance overhead Best for: Mixed content types, when chunk count optimization is important

Document-Aware Chunking

File type detection: PDF pages, Word sections, HTML elements Metadata preservation: Headers, footers, page numbers, sections Table and image handling: Special processing for non-text elements Pros: Preserves document structure and metadata Cons: Format-specific implementation required Best 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 domains 512-768 dimensions: Balanced performance, good for most applications 1024-1536 dimensions: High quality, better for complex domains, higher cost 2048+ 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 Encoder Code embeddings: CodeBERT, GraphCodeBERT, CodeT5 Scientific text: SciBERT, BioBERT, ClinicalBERT Multilingual: LaBSE, multilingual-e5, paraphrase-multilingual

3. Vector Database Selection

Pinecone

Managed service: Fully hosted, auto-scaling Features: Metadata filtering, hybrid search, real-time updates Pricing: $70/month for 1M vectors (1536 dim), pay-per-use scaling Best for: Production applications, when managed service is preferred Cons: Vendor lock-in, costs can scale quickly

Weaviate

Open source: Self-hosted or cloud options available Features: GraphQL API, multi-modal search, automatic vectorization Scaling: Horizontal scaling, HNSW indexing Best for: Complex data types, when GraphQL API is preferred Cons: Learning curve, requires infrastructure management

Qdrant

Rust-based: High performance, low memory footprint Features: Payload filtering, clustering, distributed deployment API: REST and gRPC interfaces Best for: High-performance requirements, resource-constrained environments Cons: Smaller community, fewer integrations

Chroma

Embedded database: SQLite-based, easy local development Features: Collections, metadata filtering, persistence Scaling: Limited, suitable for prototyping and small deployments Best for: Development, testing, small-scale applications Cons: Not suitable for production scale

pgvector (PostgreSQL)

SQL integration: Leverage existing PostgreSQL infrastructure Features: ACID compliance, joins with relational data, mature ecosystem Performance: ivfflat and HNSW indexing, parallel query processing Best for: When you already use PostgreSQL, need ACID compliance Cons: Requires PostgreSQL expertise, less specialized than purpose-built DBs

4. Retrieval Strategies

Dense Retrieval

Semantic similarity: Using embedding cosine similarity Advantages: Captures semantic meaning, handles paraphrasing well Limitations: May miss exact keyword matches, requires good embeddings Implementation: Vector similarity search with k-NN or ANN algorithms

Sparse Retrieval

Keyword-based: TF-IDF, BM25, Elasticsearch Advantages: Exact keyword matching, interpretable results Limitations: Misses semantic similarity, vulnerable to vocabulary mismatch Implementation: Inverted indexes, term frequency analysis

Hybrid Retrieval

Combination approach: Dense + sparse retrieval with score fusion Fusion strategies: Reciprocal Rank Fusion (RRF), weighted combination Benefits: Combines semantic understanding with exact matching Complexity: Requires tuning fusion weights, more complex infrastructure

Reranking

Two-stage approach: Initial retrieval followed by reranking Reranking models: Cross-encoders, specialized reranking transformers Benefits: Higher precision, can use more sophisticated models for final ranking Tradeoff: Additional latency, computational cost

5. Query Transformation Techniques

HyDE (Hypothetical Document Embeddings)

Approach: Generate hypothetical answer, embed answer instead of query Benefits: Improves retrieval by matching document style rather than query style Implementation: Use LLM to generate hypothetical document, embed that Use cases: When queries and documents have different styles

Multi-Query Generation

Approach: Generate multiple query variations, retrieve for each, merge results Benefits: Increases recall, handles query ambiguity Implementation: LLM generates 3-5 query variations, deduplicate results Considerations: Higher cost and latency due to multiple retrieva