@prefix rdf: . @prefix sl: . @prefix skos: . @prefix rdfs: . @prefix tag: . @prefix foaf: . @prefix dc: . tag:vector_database a sl:Tag ; skos:prefLabel "Vector database" . tag:retriever a sl:Tag ; skos:prefLabel "Retriever" . dc:title "[2002.06504] Differentiable Top-k Operator with Optimal Transport" ; sl:comment "> if the top-k operation is implemented in an algorithmic way, e.g., using bubble algorithm, the resulting model cannot be trained in an end-to-end way using prevalent gradient descent algorithms. This is because these implementations typically involve swapping indices, whose gradient cannot be computed. Moreover, the corresponding mapping from the input scores to the indicator vector of whether this element belongs to the top-k set is essentially discontinuous. To address the issue, we propose a smoothed approximation, namely the SOFT (Scalable Optimal transport-based diFferenTiable) top-k operator\r\n> ...\r\n> We apply the proposed operator to the [k-nearest neighbors](tag:k_nearest_neighbors_algorithm) and [beam search](tag:beam_search) algorithms, and demonstrate improved performance" ; sl:creationDate "2020-06-29" ; sl:tag tag:top_k , tag:arxiv_doc . tag:llamaindex a sl:Tag ; skos:prefLabel "LlamaIndex" . tag:top_k a sl:Tag ; rdfs:isDefinedBy ; sl:comment "eg. finding the k largest or smallest elements from a collection of scores" ; skos:prefLabel "Top-k" ; skos:related tag:nearest_neighbor_search , tag:kd_mkb ; foaf:page tag:top_k.html . tag:nearest_neighbor_search a sl:Tag ; skos:prefLabel "Nearest neighbor search" . dc:title "Jerry Liu sur Twitter : \"Using cross-encoding as a reranking step can dramatically speed up LLM inference time AND improve accuracy!\"" ; sl:comment "(speedup inference, because you can pass less nodes to the context)\r\n\r\n> We use an [MSMarco SBERT cross-encoder from \r\n@huggingface](https://www.sbert.net/docs/pretrained-models/ce-msmarco.html)\r\n\r\n```\r\nfrom sentence_transformers import CrossEncoder\r\nmodel = CrossEncoder('model_name', max_length=512)\r\nscores = model.predict([('Query', 'Paragraph1'), ('Query', 'Paragraph2') , ('Query', 'Paragraph3')])\r\n```\r\n\r\n(cf. https://www.sbert.net/docs/pretrained-models/ce-msmarco.html" ; sl:creationDate "2023-07-20" ; sl:tag tag:tweet , tag:top_k , tag:retrieval_augmented_generation , tag:jerry_liu , tag:cross_encoders . tag:semantic_search a sl:Tag ; skos:prefLabel "Semantic Search" . dc:title "Jerry Liu sur Twitter : \"Tuning top-k for semantic search is challenging... Introducing the LlamaIndex AutoRetriever for vector databases\"" ; sl:comment "> Tuning top-k for semantic search is challenging ; **the value can change depending on the context**. We now allow you to *infer* this value + other params during retrieval-time, using an LLM! Introducing the **LlamaIndex AutoRetriever for vector databases**. " ; sl:creationDate "2023-05-13" ; sl:tag tag:semantic_search , tag:jerry_liu , tag:llamaindex , tag:retriever , tag:vector_database , tag:tweet , tag:top_k . tag:retrieval_augmented_generation a sl:Tag ; skos:prefLabel "RAG (Retrieval-Augmented Generation)" . tag:cross_encoders a sl:Tag ; skos:prefLabel "Cross-Encoders" . tag:tweet a sl:Tag ; skos:prefLabel "Tweet" . tag:jerry_liu a sl:Tag ; skos:prefLabel "Jerry Liu" . tag:kd_mkb a sl:Tag ; skos:prefLabel "KD-MKB" . tag:arxiv_doc a sl:Tag ; skos:prefLabel "Arxiv Doc" .