Satyadeep Behera’s Post

Building a search engine is confusing: retrieval, reranking, embeddings, indexes, etc. Here is a brief introduction to the building blocks required for a simple search application.

When conducting an advanced SEO analysis, I frequently utilise vector embeddings for text feature extraction, similarity searches, clustering, retrieval, ranking and so on. One of the main burdens on top of compute is storage space, as these files tends go into terabytes for very large websites. Today I did a deep analysis and realised I've been wasting time, money and hard drive space this whole time. I started with a SOTA embedding model and tested the quality of vector embeddings after applying: a. Matryoshka Representation Learning (MRL) b. Binary Embeddings c. Combined Both Y = cosine spearman on MTEB/STS12 dataset. X = embedding dimensionality reduction via MRL 🟠= MRL🟡= Binary Embeddings + MRL

For a production RAG pipeline, you need to go beyond the initial vector search to ensure your embedding results are relevant to your users’ questions. “”” A reranker, functioning as the second-pass document filter in information retrieval(IR)systems, focuses on reordering documents retrieved by the initial retriever (semanticsearch, keyword search, etc.). This reordering is based on the query-document relevance,emphasizing the quality of document ranking. Balancing efficiency with effectiveness,rerankers prioritize the enhancement of search result quality, often involving more complexmatching methods than traditional vector inner products “”” #softwaredesign #softwarearchitecture #softwareengineering #genai

How to improve RAG implementations? Anthropic's Contextual Retrieval seems to be an interesting approach. Check out the attached document for details. My view is while text retrieval gets better here, I think we need to solve for image/table based content retrievals so that the chunks with these data types also work well. What are your views on Contextual Retrieval? Keen to know if you have seen solutions which helps with the image or table handling?

This seems quite interesting! Especially combining keyword search with similarity with query rewrite.

gold advice here for RAG performance. Certainly recommended

Cohere releases Rerank-3 model focused on semantic reranking The report very substantial increase in NDCG metrics vs BM25 (yes, it's very old baseline, it's not used in companies who work on their search engines since there are very substantial improvements of the search when you modify ranking, but still it's a default rank in several open source search engine ) The new reranking is multilingual. I see this development by Cohere as very important. Last few months, I talked to multiple companies about their search engines, and I see a persistent pattern that it's harder to work on search in German, French, Italian etc vs English. There are libraries, there is a lot of work done within companies, but still it's very hard problem to develop search engine for companies whose customers need search in those languages Latency still needs to be improved. A typical hard requirement for any search engine is to have not more than 1 sec end-to-end search experience (query is typed -> the page is rendered ) (of course, it can be different and even below , but I consider the whole landscape, including travel, ecommerce, fashion and other companies, which customer do not compare their search experience to Google directly and do not expect 200ms). 1 sec end-to-end, leaves approximately 200-400ms (depending on system) for search backend including retrieval, ranking, or typically less than 200 ms for reranking (I take upper limits across various search engines in various sectors, in many cases, limits are significantly smaller ). Typical *reranking* is hundreds-thousand of documents. Cohere model is around 3 sec for 50 seconds (for 4096 context window, ), Still a lot to improve to make it useful for many companies

Dive deeper into these strategies and enhance your retrieval systems beyond the basics!

How To Choose The Right Content Clustering Algorithm For Your Project https://buff.ly/3YEpqrQ @[Penfriend](urn:li:organization:100267386)

Blank nodes, deduction, and explanation: a comparative analysis of Semantic Web systems (fulltext on https://lnkd.in/dvc-ThdF)