Why my thesis evaluation pipeline scored every run with both a deterministic benchmark and an LLM judge. The disagreements between them turned out to be the useful part.
In my thesis work the model only routes questions to the right data partition and pushes predicates down to PostGIS and Neo4j. Inference token cost dropped 3x and routing-index quality ended up deciding answer accuracy.
Papyrus runs the embedding model, the vector search and the reranker entirely in your browser. Here's how I made a million-document index fit on consumer hardware.
Memvid started as a joke about hiding AI memory inside MP4 video files. The serious version that replaced it drops the video entirely. It's worth explaining why.
A field guide to the failure modes that don't show up in tutorials and how to systematically debug them.
MTEB scores won't help you here. How I choose embeddings when the data has structure.
Everyone asks this question. Most of the answers online are wrong or at least incomplete.
How to go from messy documents to structured, queryable knowledge without a PhD in NLP.