The original and multilingual E5 embedding models were trained with a "query:" or "passage:" prefix on every input. Leave the prefixes off, or store documents with the query prefix, and retrieval quietly gets worse with no error anywhere.
A Berkeley-led vector index keeps a pruned graph, not the vectors, so a local RAG index can be a few percent of the data's size. Costs and how to try it.
Five open NVIDIA retrieval models search page images and text, rerank results and parse PDFs, under the permissive OpenMDW license. What to check first.
A token embedding table can also score output tokens. The shared dimensions save one vocabulary-sized matrix, while tying changes how every row learns.
Perplexity's model cards describe pplx-embed-v2-late 0.6B and 9B as MIT-licensed ColBERT-style embedders, one 128-wide vector per token, sharing one embedding space.
Google published EmbeddingGemma 2, an open model that maps text, code, images, video, and audio into one 768-wide vector. Developers can load fewer encoders or store a shorter vector.
Neon's 2 October 2026 post adds POST /v1/embeddings on AI Gateway, with qwen3-embedding-0-6b at $0.02 and gte-large-en at $0.13 per million input tokens, for vectors stored in Lakebase Postgres.
Chunk size, overlap and boundaries decide what a retriever can return. This post covers defaults documented by Azure AI Search and OpenAI, plus a way to test a setting on your own questions.
turbopuffer made native embedding generally available (Sep 29, 2026): embed on write, on query, or keep BYOE. Systems overlap with index metadata fetch; Linear/Readwise beta color is vendor-stated.
Same-day Sep 30, 2026: Voyage ships rerank-3 and rerank-3-lite (“available today”, with no GA claim). MongoDB Atlas Native Reranking via $rerank is Preview only (docs). The Voyage and MongoDB lift claims are theirs. Separate from Atlas Agent Engine, and neither is a Vector Search substitute.
Qdrant’s Constella research preview (Sep 29, 2026): index docs once with Stella (400M English), then query with Zero, Nano, or Stella against the same collection—no re-embedding. Research preview, not GA. Encode-speed and nDCG figures are vendor-reported.
Retrieval quality is decided by the embedding model, the dimensions and the chunking, and it is only known once you measure recall. How to choose, how to test, and where vector search earns its place.
Perplexity published open-weight pplx-embed-v2-context-9b-preview on Hugging Face (MIT): contextual chunk embeddings for RAG. Preview may break; not on Perplexity API yet. Use encode_queries vs encode; int8 + MRL 1024/2048; transformers≥5.4 + trust_remote_code.
A vector database answers “what is closest in meaning to this?” instead of “which rows match?”. Here is the machinery: embeddings, distance, and why you never design the fields.