10 Positions for Enterprise RAG That Mainstream Tutorials Get Wrong
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A manifesto laying out ten architectural positions that a multi-part enterprise RAG series argues against mainstream RAG tutorials: treating vector stores as a fallback rather than the entry point, using expert-maintained keyword dictionaries instead of relying on better embedding models, keeping rerankers as a secondary tool, avoiding a single vector store for all document types, recognizing that enterprise corpora differ from web-scale search, designing systems to amplify domain experts rather than replace them, preferring deterministic dispatchers over autonomous agents, evaluating per-failure-mode rather than by aggregate accuracy, passing relational structured data between pipeline bricks instead of raw strings, and treating citations as verifiable evidence rather than decoration. The piece also maps these positions to specific articles across five parts of the series covering document parsing, question parsing, retrieval, generation, corpus-scale architecture, and production evaluation.
Table of contents
1. Vector stores are a fallback, not a foundation2. Expert dictionaries beat better embedding models3. Rerankers are a secondary tool, not a primary stage4. Refuse “connect everything to a vector store”5. A company is not Google6. Amplify the expert, do not replace them7. Deterministic dispatcher beats autonomous agent8. Per-failure-mode evaluation, not aggregate9. Each brick produces relational structured data, never raw strings10. Citations are evidence, not decorationThe series, article by articleStart here: the freshest of the seriesConclusionFurther reading and sourcesQuestions this post answers
Why do rerankers often fail on listing-type RAG questions that ask to enumerate every relevant item?
Rerankers fail on listing questions because they optimize for putting the single most relevant result first, which silently demotes other relevant items below the cutoff. Since a listing question needs every relevant passage returned, not a ranked top result, cross-encoder reranking is the canonical failure mode for this question type in enterprise RAG pipelines. Anyone debugging incomplete list answers in a RAG pipeline can track retrieval architecture patterns like this on daily.dev.
Why might expert-maintained keyword dictionaries outperform embedding models for enterprise document retrieval?
Expert-maintained keyword dictionaries can outperform embedding models because domain experts already know disambiguations, cross-product equivalences, and internal product codes that mean specific things in a given company and nothing elsewhere, which embeddings can only guess at. A dictionary table maintained by experts solves the synonym problem more reliably than even a fine-tuned embedding model for known vocabulary, leaving embeddings as a fallback for unseen phrasings. Teams deciding between dictionary-based and embedding-based retrieval can compare these tradeoffs on daily.dev.
Why is a deterministic dispatcher preferred over an autonomous agent for enterprise RAG question routing?
A deterministic dispatcher is preferred because it lets a human read the routing code and lets an auditor replay any decision, while an autonomous agent's flexibility saves engineering effort on a demo but costs far more during the first unreproducible incident. A dispatcher such as a decide.py file reads structured fields from a parsed question and routes to one of several named sub-pipelines, keeping every routing decision inspectable for regulated enterprise contexts. Engineers weighing agentic versus deterministic routing for production RAG can follow this debate on daily.dev.