Bilingual, source-visible engineering notes built around useful explanations, reproducible details, and real project context.
Knowledge, made useful
4 articles
Retrieval Engineering
What fixed-window, recursive, and semantic chunking really cost, how size and overlap move token-level recall and precision, and a reproducible protocol for picking parameters with evidence.
Retrieval Engineering
Only one HNSW knob is adjustable after the build. Version-checked defaults across pgvector, Qdrant, Lucene and Elasticsearch, where quantization really costs you, and when exhaustive search wins.
Retrieval Engineering
Build a versioned RAG retrieval benchmark, select Recall@k, MRR, and nDCG@k deliberately, and gate chunking, latency, context budget, LLM judges, shadow traffic, and rollout.
Retrieval Engineering
Combine $vectorSearch and $search in MongoDB with $rankFusion: version prerequisites, both index definitions, a working pipeline, input-pipeline constraints, and how to tune recall.
