Library
I don't collect papers. I collect ideas that change how I build systems.
Why it's in my library
This paper changed how I think about retrieval: the generator and the retriever are separate concerns, which means the corpus can be updated without retraining anything. It's the reason Compliance Agent and Graph-Augmented Agentic RAG both treat retrieval quality as a first-class problem, not a preprocessing step.
Applied in
Key ideas
Why it's in my library
This is the paper that convinced me retrieval hit-rate isn't enough to evaluate a RAG system. A model can retrieve the right passage and still fail to use it, just because of where it landed in the context window. That's why eval suites need a faithfulness metric separate from retrieval accuracy.
Applied in
Key ideas
Why it's in my library
This paper is the clearest argument I've seen for writing rules down instead of hoping a model infers them from examples. The same instinct shows up in how I build deploy gates and validation layers: an explicit, auditable rule beats an implicit one, even when the implicit one technically works more often.
Applied in
Key ideas
Why it's in my library
Everything else in this library assumes this paper. Self-attention is the reason retrieval, agents, and long-context evaluation are even the right things to be arguing about — it's the substrate, not a technique applied on top of one.
Key ideas
Why it's in my library
The chapters on consistency and replication are the reason DiffDDL exists in its current form: string-level schema diffing is the implicit-tradeoff version of a problem this book treats as a first-class design decision.