RAG research, evaluation, and AI tokenomics

Retrieval-augmented generation (RAG) pairs a model with a selected evidence corpus. This hub covers how to test that pairing, choose the right amount of context, and reason about model cost and error—not generic claims that every knowledge problem needs a vector database.

A reading path

1. Retrieval-Augmented Generation as a Capability Multiplier for Research Tasks — Begin with controlled evidence for when retrieval improves research quality.
2. The Dose-Response Curve of RAG — Learn why more retrieved context is not always better context.
3. Domain-Specific RAG with Gemini 3 Flash Beats PRO with Web Search Grounding — See how a targeted corpus can outperform a larger model with web-search grounding.

All RAG research articles

1. AI Tokenomics for Software Engineering: A Practical Economic Model for Iterative Model Routing

   A canonical HTML landing page for Dominik Gorecki's AI Tokenomics for Software Engineering PDF, with citation, scope, provenance, and download details.

2. Domain-Specific RAG with Gemini 3 Flash Beats Pro with Web Search Grounding

   A controlled comparison finds that Gemini 3 Flash with a domain academic corpus outperforms Gemini 3 Pro with web search grounding on overall score and four of five judged dimensions.

3. Retrieval-Augmented Generation as a Capability Multiplier for Research Tasks

   Two controlled experiments show that domain retrieval improves research outputs on four of five quality dimensions and can let a smaller model outperform a larger model without retrieval.

4. The Dose-Response Curve of RAG: More Context Yields Diminishing Returns

   An empirical study of retrieval volume and response quality showing that balanced context improves RAG results while additional context eventually produces diminishing returns.

Explore agentic engineering for delivery systems, or AI, cognition, and society for broader consequences. Return to all articles.