Parametric memory
Knowledge encoded in the pretrained sequence-to-sequence model.
The architecture behind a widely used acronym
The Original RAG Research
The original RAG paper combined a pretrained generator with retrieved non-parametric memory for knowledge-intensive language tasks.
Model parameters are difficult to update and inspect
Large pretrained models can store factual patterns in their parameters, but accessing precise knowledge and updating it remain difficult. The RAG paper explored a general-purpose approach that retrieves explicit documents and conditions generation on them.
Parametric model plus non-parametric index
Non-parametric memory
A dense vector index of Wikipedia passages accessed through a neural retriever.
Retriever
Selects candidate documents conditioned on the input.
Generator
Produces the output using the input and retrieved evidence.
RAG-Sequence
Uses the same retrieved document assumptions across the generated sequence.
RAG-Token
Allows different retrieved-document influence across generated tokens.
Readable and updateable outside knowledge
Retrieved documents remain human-readable and can be changed by updating the index rather than retraining the entire language model. The architecture also opened a path toward provenance because a system could expose the documents used during generation.
That potential does not guarantee faithful citation. A product must still preserve the connection between the retrieved passage and the final claim.
The acronym now covers a family of systems
- Hybrid lexical and semantic retrieval.
- Query rewriting and multi-query retrieval.
- Metadata filters and access control.
- Reranking and context compression.
- Knowledge graphs and structured tools.
- Citation generation and faithfulness checks.
- Agent loops and repeated retrieval.
Evaluate the retriever and generator separately
The paper’s enduring lesson is that retrieval can extend and update model knowledge. Modern evaluation must still determine whether the source was correct, the best passage was retrieved, the context preserved its limits and the output remained faithful.
Primary material behind this analysis
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
The original RAG paper combining parametric generation with retrieved non-parametric memory.
Continue through the discovery stack
Retrieval-Augmented Generation Explained
A detailed explanation of RAG, the original architecture, modern pipelines, failure modes and the boundary between publishers and AI-product operators.
Information Retrieval Explained
A practical guide to lexical, semantic, hybrid, graph and structured retrieval, failure modes, passage design and evaluation.
AI Search Reliability and Source Quality
A framework for source quality, retrieval quality, generation fidelity, citations, freshness, conflicts and lifecycle risk management.
Search and AI Research Library
A research hub connecting NIST, Stanford HAI, OECD, Google, OpenAI, Perplexity, Microsoft, IETF, W3C, Schema.org and original GEO, RAG and REG sources.