From demo to accuracy infrastructure
A series of eight articles on building RAG systems that work reliably on real enterprise data. Read the preface first — it explains what the series is for and how to use it as a diagnostic tool rather than a checklist.
Why a RAG demo succeeds and a production system does not, what this series covers, who it is for, and how to read it as a diagnostic tool rather than a checklist.
4 min · Updated June 2026
Read the preface→What exactly breaks, and why, when you scale a naive RAG system to real enterprise data?
The structural problem with LLM hallucination, how first-generation RAG responds, and the five specific ways it fails at scale: dirty documents, single-channel retrieval, context-stripped chunks, pipeline blindness, and multi-tenancy.
6 min · Updated June 2026
Read this part→What do those four words actually mean, and which problem does each one solve?
Four independent concepts, each solving a distinct failure mode: multimodal ingestion for heterogeneous corpora, hybrid retrieval with RRF, agentic self-correction before and after generation, and grounded generation with citations.
8 min · Updated June 2026
Read this part→What are the specific, named failure modes across ingestion, retrieval, and generation that you need to design against?
A precise taxonomy of failure modes across ingestion (format diversity, OCR, context destruction, embedding limits, scale), the eight retrieval failures F1-F8, generation challenges, and operational realities of multi-tenancy and cost.
7 min · Updated June 2026
Read this part→How do you organise the system so that each concern has a clear owner, a defined boundary, and a deployable footprint?
The seven-plane architecture: control, observability, ingestion, tooling, indexing, retrieval, and agentic planes. What each plane owns, where its boundary sits, and why observability is structural plumbing rather than optional instrumentation.
7 min · Updated June 2026
Read this part→How do you parse every document format faithfully, and how do you know when parsing has failed?
Modality-specific parsing strategies, the ten ingestion accuracy patterns, how Docling, OpenAI embeddings, and vision APIs collaborate without overlapping, confidence-gated human review, and the ingestion observability model.
14 min · Updated June 2026
Read this part→What are the eight ways retrieval silently returns the wrong answer, and what is the specific pattern that kills each one?
F1 through F8: semantic gap, lexical miss, top-k cliff, multi-hop failure, lost-in-the-middle, distractor poisoning, no-answer, and wrong-tool routing. Nine retrieval accuracy patterns with mechanisms, implementations, and key metrics.
10 min · Updated June 2026
Read this part→How does an agent self-correct after retrieval and generation, and how does the system get more accurate the longer it runs in production?
Why agents earn their cost, seven agentic accuracy patterns (Adaptive Router, CRAG, Self-RAG, Query Decomposition, Citation, Bounded Budget, Tool Fallback), the full LangGraph graph structure, and the compound accuracy flywheel.
12 min · Updated June 2026
Read this part→What does the confirmed production stack look like, what was deliberately traded away, and what separates a production system from a demo?
The full confirmed stack with every tradeoff stated explicitly, the most consequential choices explained (embeddings, LLM generation, chunking strategy), the open-source tool reference, key empirical claims with attribution, and the five things that actually separate a production system from a demo.
12 min · Updated June 2026
Read this part→