https://www.medhaa.au/blog daily 0.75 2025-08-18 https://www.medhaa.au/blog/blog-post-one-g357p monthly 0.5 2025-08-17 https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/7292adca-666f-41c8-beee-f54e36633df0/3.png Blog - An Architectural Deep Dive of Production Grade RAG Systems - HNSW provides excellent query performance but requires substantial memory. IVF (Inverted File Index) offers better memory efficiency with slightly higher latency. Production systems often implement a tiered approach: Hot tier: HNSW for frequently accessed, recent documents Warm tier: IVF with compression for moderate access patterns Cold tier: Flat indices on object storage for archival content https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/bf8a7761-5280-4acc-aafb-197312524862/6.png Blog - An Architectural Deep Dive of Production Grade RAG Systems - Multi Tenant Architecture Enterprise deployments require sophisticated access control that extends beyond document-level permissions. Embeddings themselves can leak information through similarity searches. Consider a scenario where searching for "layoff plans" returns high similarity to documents a user cannot access—even without seeing the documents, the similarity score reveals sensitive information. This architecture ensures complete isolation between tenants while allowing for resource sharing at the infrastructure level. Each tenant's embeddings remain separate, preventing information leakage through vector similarity. https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/3d455a01-f45a-4e95-be0b-cf4026589e5b/10.png Blog - An Architectural Deep Dive of Production Grade RAG Systems - Blue-Green Deployment for Index Updates This architecture enables zero-downtime updates while maintaining the ability to instantly rollback if quality metrics degrade. https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/28d363d2-67fb-4c08-9acd-7d5ea3523027/1.png Blog - An Architectural Deep Dive of Production Grade RAG Systems The document processing pipeline represents the first collision between theoretical elegance and practical chaos. Production systems inherit decades of enterprise documents: scanned PDFs at various angles, password-protected files, Excel sheets with data buried in cell comments, and Word documents with tracked changes spanning years. Document processing becomes a multi-stage pipeline where each document type requires specialized handling. OCR services must handle rotated scans, detect table structures, and preserve layout information. The challenge intensifies when dealing with multi-column layouts, where reading order becomes ambiguous, or technical diagrams where spatial relationships carry semantic meaning. https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/b837cbee-c8b8-4323-bf65-09fe00a8b974/2.png Blog - An Architectural Deep Dive of Production Grade RAG Systems - Pure semantic search fails spectacularly on technical content. The query "Q4 2023 EBITDA" has minimal semantic overlap with the phrase "fourth quarter earnings before interest taxes depreciation and amortization for twenty twenty-three" despite identical meaning. This necessitates a sophisticated hybrid approach The fusion layer implements Reciprocal Rank Fusion (RRF) to combine results from different retrieval methods. Each retrieval method excels in different scenarios: dense retrieval captures semantic similarity, sparse retrieval handles exact matches and rare terms, while metadata filtering ensures temporal and document-type relevance. https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/67050ddb-b0d5-4c5a-8874-6f65d5579994/8.png Blog - An Architectural Deep Dive of Production Grade RAG Systems - Observability Stack Key metrics include not just system performance but semantic quality measures. Normalized Discounted Cumulative Gain (NDCG) and Mean Reciprocal Rank (MRR) provide insights into retrieval quality. User feedback loops enable continuous improvement of ranking algorithms. https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/46f2fb5c-7f37-47ec-a3d4-3fec38bfa213/5.png Blog - An Architectural Deep Dive of Production Grade RAG Systems - Every claim must be traceable to source documents. This requires maintaining provenance throughout the pipeline https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/b0f77d39-16dc-4ec2-a34e-6733a86c3217/7.png Blog - An Architectural Deep Dive of Production Grade RAG Systems - Request Flow Timing Breakdown Caching strategies must be sophisticated enough to handle semantic similarity rather than just exact matches. A query for "quarterly revenue" should potentially return cached results for "Q4 earnings" if the semantic similarity exceeds threshold and temporal relevance remains valid. https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/c0bff251-f2a0-4828-9889-65f6a723ada7/9.png Blog - An Architectural Deep Dive of Production Grade RAG Systems - Intelligent caching, index compression, and query routing to appropriately sized models become essential for cost control. Production systems implement sophisticated cost allocation to track usage by team, project, or query type. https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/d9425bbc-48ec-4ad1-8b9f-0a9984dca63b/4.png Blog - An Architectural Deep Dive of Production Grade RAG Systems - When confidence falls below threshold, the system must gracefully decline to answer rather than generate plausible-sounding fabrications. This requires careful prompt engineering that explicitly instructs the model about uncertainty handling. https://www.medhaa.au/blog/blog-post-two-38nhb monthly 0.5 2025-08-17 https://www.medhaa.au/blog/blog-post-three-8bka7 monthly 0.5 2025-08-17 https://www.medhaa.au/blog/blog-post-four-rlal8 monthly 0.5 2025-08-17 https://www.medhaa.au/areas-of-practice daily 0.75 2025-08-18 https://www.medhaa.au/home daily 1.0 2025-08-18 https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/8ca91a48-af3c-4a3d-a839-39bd4843338d/4.png https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/c868c4d3-42a2-4266-9812-359fc4d57a66/1.png https://images.squarespace-cdn.com/content/v1/688e93c174067f3578e2dd62/fbc1bccf-9cd6-405c-bc1b-4e9269c9e253/3.png https://www.medhaa.au/contact daily 0.75 2025-08-17