rag-architect（RAG 架构师）

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 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 retrievals

Step-Back Prompting

• Approach: Generate broader, more general version of specific query
• Benefits: Retrieves more general context that helps answer specific questions
• Implementation: 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 first
• Diversity optimization: Avoid redundant information
• Token budget management: Fit within model context limits
• Hierarchical inclusion: Include summaries before detailed chunks

Context Compression

• Summarization: Compress less relevant chunks while preserving key information
• Key information extraction: Extract only relevant facts/entities
• Template-based compression: Use structured formats to reduce token usage
• Selective inclusion: Include only chunks above relevance threshold

7. Evaluation Frameworks

Faithfulness Metrics

• Definition: How well generated answers are grounded in retrieved context
• Measurement: Fact verification against source documents
• Implementation: NLI models to check entailment between answer and context
• Threshold: >90% for production systems

Relevance Metrics

• Context relevance: How relevant retrieved chunks are to the query
• Answer relevance: How well the answer addresses the original question
• Measurement: Embedding similarity, human evaluation, LLM-as-judge
• Targets: Context relevance >0.8, Answer relevance >0.85

Context Precision & Recall

• Precision@K: Percentage of top-K results that are relevant
• Recall@K: Percentage of relevant documents found in top-K results
• Mean Reciprocal Rank (MRR): Average of reciprocal ranks of first relevant result
• NDCG@K: Normalized Discounted Cumulative Gain at K

End-to-End Metrics

• RAGAS: Comprehensive RAG evaluation framework
• Correctness: Factual accuracy of generated answers
• Completeness: Coverage of all relevant aspects
• Consistency: Consistency across multiple runs with same query

8. Production Patterns

Caching Strategies

• Query-level caching: Cache results for identical queries
• Semantic caching: Cache for semantically similar queries
• Chunk-level caching: Cache embedding computations
• Multi-level caching: Redis for hot queries, disk for warm queries

Streaming Retrieval

• Progressive loading: Stream results as they become available
• Incremental generation: Generate answers while still retrieving
• Real-time updates: Handle document updates without full reprocessing
• Connection management: Handle client disconnections gracefully

Fallback Mechanisms

• Graceful degradation: Fallback to simpler retrieval if primary fails
• Cache fallbacks: Serve stale results when retrieval is unavailable
• Alternative sources: Multiple vector databases for redundancy
• Error handling: Comprehensive error recovery and user communication

9. Cost Optimization

Embedding Cost Management

• Batch processing: Batch documents for embedding to reduce API costs
• Caching strategies: Cache embeddings to avoid recomputation
• Model selection: Balance cost vs quality for embedding models
• Update optimization: Only re-embed changed documents

Vector Database Optimization

• Index optimization: Choose appropriate index types for use case
• Compression: Use quantization to reduce storage costs
• Tiered storage: Hot/warm/cold data strategies
• Resource scaling: Auto-scaling based on query patterns

Query Optimization

• Query routing: Route simple queries to cheaper methods
• Result caching: Avoid repeated expensive retrievals
• Batch querying: Process multiple queries together when possible
• Smart filtering: Use metadata filters to reduce search space

10. Guardrails & Safety

Content Filtering

• Toxicity detection: Filter harmful or inappropriate content
• PII detection: Identify and handle personally identifiable information
• Content validation: Ensure retrieved content meets quality standards
• Source verification: Validate document authenticity and reliability

Query Safety

• Injection prevention: Prevent malicious query injection attacks
• Rate limiting: Prevent abuse and ensure fair usage
• Query validation: Sanitize and validate user inputs
• Access controls: Ensure users can only access authorized content

Response Safety

• Hallucination detection: Identify when model generates unsupported claims
• Confidence scoring: Provide confidence levels for generated responses
• Source attribution: Always provide sources for factual claims
• Uncertainty handling: Gracefully handle cases where answer is uncertain

Implementation Best Practices

Development Workflow

1. Requirements gathering: Understand use case, scale, and quality requirements
2. Data analysis: Analyze document corpus characteristics
3. Prototype development: Build minimal viable RAG pipeline
4. Chunking optimization: Test different chunking strategies
5. Retrieval tuning: Optimize retrieval parameters and thresholds
6. Evaluation setup: Implement comprehensive evaluation metrics
7. Production deployment: Scale-ready implementation with monitoring

Monitoring & Observability

• Query analytics: Track query patterns and performance
• Retrieval metrics: Monitor precision, recall, and latency
• Generation quality: Track faithfulness and relevance scores
• System health: Monitor database performance and availability
• Cost tracking: Monitor embedding and vector database costs

Maintenance & Updates

• Document refresh: Handle new documents and updates
• Index maintenance: Regular vector database optimization
• Model updates: Evaluate and migrate to improved models
• Performance tuning: Continuous optimization based on usage patterns
• Security updates: Regular security assessments and updates

Common Pitfalls & Solutions

Poor Chunking Strategy

• Problem: Chunks break mid-sentence or lose context
• Solution: Use boundary-aware chunking with overlap

Low Retrieval Precision

• Problem: Retrieved chunks are not relevant to query
• Solution: Improve embedding model, add reranking, tune similarity threshold

High Latency

• Problem: Slow retrieval and generation
• Solution: Optimize vector indexing, implement caching, use faster embedding models

Inconsistent Quality

• Problem: Variable answer quality across different queries
• Solution: Implement comprehensive evaluation, add quality scoring, improve fallbacks

Scalability Issues

• Problem: System doesn't scale with increased load
• Solution: 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.