Vectors.run is a lab notebook about the engineering reality of RAG and Information Retrieval. It focuses on practical failure modes and measurable trade-offs: embedding brittleness, evaluation methodology, and the cost of noise. Everything is driven by experiments, with the aim of producing actionable signal for teams past the proof-of-concept phase.

Query Processing Systems

Engineering Text-to-Elasticsearch DSL with LLMs: Constraints, Safety, and Trust

Text-to-DSL works when you stop treating it like prompting and start treating it like systems engineering: typed contracts, validation layers, policy budgets, and a UI that exposes constraints.

Evaluation Metrics

Redefining RAG Evaluation: A Set-Based Approach (Beyond nDCG)

Moving from ranked lists to prompt sets: a better way to benchmark RAG.

Embeddings Retrieval

How Proper Names Behave in Text Embedding Space

Quantifying the fragile "name signal" in modern vector spaces.

Retrieval Query Processing

When Small Talk Hurts Retrieval: Quantifying Conversational Noise in RAG

Do we really need query rewriting? Fixing the "Politeness Penalty" without the latency.

Evaluation Datasets

Turning Noise into Signal: Building RAG "Silver Sets" with Plackett-Luce and Lock-DAGs

Treating the LLM as a flawed sensor to build stable ground truth.