Circular Dependency Fixes when Bootstrapping a Golden Set
summary
Workflow for bootstrapping a golden set that actually works: start from a 62K MovieLens sample enriched via TMDB, index into Vespa + navigator + Python, generate synthetic queries from documents and facets, retrieve diverse top-N candidates via BM25, fastText (trained on the corpus) and Arctic embeddings, then use a dataset-generator to send batched (8–10) query-document groups to an LLM judge, save ratings plus explanations, iterate on queries and cleaning until coverage and candidate diversity are acceptable.
AI-written summary · about
chapters (11)
- 0:00Introduction and setup
- 1:51Tools and data overview
- 3:37Vespa schema and deployment
- 5:43Data enrichment process
- 8:12FastText model training
- 10:41Semantic search and diversity
- 12:26Dataset generator workflow
- 14:52Judgments and dataset refinement
- 17:01Visualization and dashboards
- 17:43Resources and closing
- 18:24Q&A session
related talks
33mDeep research at scale
46mBeyond Bag of Words A Practitioner’s Guide to Advanced NLP
01h 48mRecommendation engines explained
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures
32m"You may also like"
01h 14mBuilding a Recommendation System with a Hybrid Search From Scratch
19mHow to look at your data
41mHow to train your general purpose document retriever model