Topic hub
RAG & Vector Search
Master retrieval-augmented generation, vector databases, embeddings, and semantic search systems.
All posts · 45
RAG & VECTOR SEARCH01
pgvector 2000 Dimension Limit: Why CREATE INDEX Failspgvector indexes 2000 dimensions while the column type stores 16000. A vector(3072) column loads ten million rows, then refuses an index. Four ways out.
RAG & VECTOR SEARCH02
Qdrant Snapshot Restore Returns an Empty CollectionQdrant's docs say a restore left at the default priority returns an empty collection. The v1.19.0 source shows when priority is consulted at all.
RAG & VECTOR SEARCH03
Text Embeddings Inference vs vLLM: Pick an Embedding ServerTEI's token budget is 16,384 per forward pass and v1.9.0 flipped auto-truncate to true. Four embedding servers, compared by failure mode, not by speed.
RAG & VECTOR SEARCH04
Multilingual RAG Pipeline: One Index or One Per LanguageElasticsearch ships 36 stopword language lists and Polish is not one of them. Three index layouts for a corpus that is multilingual by statute.
RAG & VECTOR SEARCH05
Contextual Retrieval vs Late Chunking: Which One to UseOne controlled head-to-head puts contextual retrieval at 72.4 nDCG@10 against late chunking's 61.0, at 120x the indexing time. Where each one earns its cost.
RAG & VECTOR SEARCH06
RAG Returns Outdated Documents: Effective Dates and RankingNaive RAG answers 58% of version-sensitive questions correctly; a version-aware pipeline hits 90%. Here is the effective-dating schema that closes the gap.
RAG & VECTOR SEARCH07
RAG Update Triggers: Webhook vs CDC vs Polling vs TTLDelta indexing runs in 90 seconds and your answers are still a day old, because a nightly cron fires it. Six triggers, staleness math, and the reconciler.
RAG & VECTOR SEARCH08
Vector Database GDPR Erasure: What DELETE Actually Doespgvector zeroes the deleted vector inside its HNSW index. Yet stock Qdrant and pgvector defaults never vacuum a single erasure. What DELETE clears, per engine.
RAG & VECTOR SEARCH09
Why Metadata Filters Collapse Your Vector Search RecallOn a vendor benchmark, a filter passing 1% of rows dropped a plain HNSW graph to 0.1% recall with no error raised. The curve, the mechanisms, the test.