Topic · 7 stories
7 stories tagged retrieval on AI Tamer, newest first.
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.
HyDE embeds a passage a language model writes in answer to the query, then searches with it. How to build it, and where it stops helping.
Cut Matryoshka-trained embedding vectors to fewer dimensions, renormalize them, and measure the retrieval loss on your own queries before you commit.
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.
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.
Cohere GA’d Embed 5 (2026-09-30): embed-v5.0-pro / embed-v5.0-fast share one space (index Pro → query Fast). Blog pricing $0.12 / $0.08 per 1M text; image $0.40 both. 128K, Matryoshka dims, multimodal—ViDoRe/finance benches are Cohere-reported.
Compare hosted APIs with models you run yourself, then choose dimensions, language and context length against your own retrieval workload.
