André Mourão

@weaviate_io

[Weaviate AI Database @weaviate_io]

Have you wondered how keyword search works at Weaviate?

This blog post shows how it works from the inside out and the recent 1.29 improvements with BlockMax WAND that led to 90%+ performance gains.

Check out the blog post and demo by André Mourão and @_jphwang: https://t.co/IOzaqHQpIV

[Weaviate AI Database @weaviate_io]

BM25 is sometimes all you need. But how does it work?

Compared to semantic search, BM25, or Best Match 25, ranks document relevance to a query based on 𝘦𝘹𝘢𝘤𝘵 𝘬𝘦𝘺𝘸𝘰𝘳𝘥 𝘮𝘢𝘵𝘤𝘩𝘦𝘴.

Here's a simplified breakdown of how BM25 works:

1. 𝗧𝗲𝗿𝗺 𝗙𝗿𝗲𝗾𝘂𝗲𝗻𝗰𝘆 (𝗧𝗙): BM25 considers how often a query term appears in a document. The more frequently a term appears, the more relevant that document is considered.
2. 𝗜𝗻𝘃𝗲𝗿𝘀𝗲 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁 𝗙𝗿𝗲𝗾𝘂𝗲𝗻𝗰𝘆 (𝗜𝗗𝗙): BM25 also takes into account how common or rare a term is across all documents. Rare terms are given more weight, as they're more likely to be relevant.
3. 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁 𝗟𝗲𝗻𝗴𝘁𝗵 𝗡𝗼𝗿𝗺𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻: Longer documents are penalized compared to shorter ones to prevent them from having an unfair advantage just because they contain more words.

Whilst semantic search is good for identifying the meaning behind a query, BM25 is best for exact matches. Although, you can always try hybrid search, which combines semantic search with BM25, offering the best of both worlds!

BM25 and hybrid search can easily be used in Weaviate, here’s how: https://t.co/dyI7tUr4Kd

[Bob van Luijt @bobvanluijt]

What I really like about @ahmed_besbes_’s article is that he exactly describes what we aim to achieve with @weaviate_io — building end-to-end AI-native applications! https://t.co/gyFWFW5BS5

[André Mourão @andmou]

RT @bobvanluijt: 🥳 The Google Cloud Next video is live! It was awesome recording this one with Brian Kaufman!
💡 The video below is the sect…

[ajit @ajit_thinks]

Binary quantization and product quantization are two strategies for reducing memory usage in vector databases - while keeping retrieval accuracy surprisingly high. In this @kaggle notebook, I implemented both with @weaviate_io and compared them.

-> Binary quantization reduces each vector element to 2 bits
-> Product quantization reduces each vector element to 8 bits and also reduces the number of vector dimensions

https://t.co/PU2iolALnD

[Bob van Luijt @bobvanluijt]

🔥 Today at Google Cloud Next, I presented a good ol' Weaviate demo!
🔮 I quickly touched on vector search, set the scene with RAG, but doubled down on Generative Feedback Loops!
💡 Did you know you can use Weaviate OOTB with Gemini?
🗓️ The video and code will be posted soon!

[Bob van Luijt @bobvanluijt]

🤩 the Weaviate team has access to @googledevs's Gemini Pro 1.5!
📦 I'm super excited about the multimodal possibilities that can be used for RAG and for Generative Feedback Loops (i.e., on import time) into the database.
💡 We'll keep you posted in the coming weeks and ensure it's fully integrated with @weaviate_io when the APIs become available!
🖼️ the image below shows a bit of the GFL thinking in the GUI... Soon (very much) more...

[André Mourão @andmou]

@AirEuropa já seguiu reclamação oficial no "Livro de Reclamações"

[André Mourão @andmou]

@AirEuropa Estou há duas semanas a tentar entrar em contacto convosco porque estou com problemas a pedir uma factura para um voo comprado pelo vosso site. Não estou nada satisfeito com a vossa assistência ao cliente.

[André Mourão @andmou]

@AirEuropa @AirEuropa mais uma semana mais uma semana sem resposta...

[Searchable.ai @searchableai]

What are you looking for in 2022? Whatever it is, we're excited to help. Here's wishing you a safe and happy new year!

[Product Hunt 😸 @ProductHunt]

Find your files across your local, cloud, and app storage with a single search.

Simplify your search: https://t.co/pMYArPKZr4

[RTP Antena 3 @antena3rtp]

Se têm ideias para impulsionar a informação histórica preservada pelo https://t.co/tXtwHtEfan, submetam a vossa candidatura até 4 de maio. Os prémios vão até aos 10.000 euros!

https://t.co/Ghu9lSKoYp

[ricardobasilio @ricardobasilio_]

Millions of images are now searchable at @PT_WebArchive (https://t.co/oRwPvMMqQ8). To know more about, see the presentation by @andmou "https://t.co/F1gZuLoDZR image search 2020-2021": https://t.co/5o8VJoo13Z

#WebArchiveWednesday

[ricardobasilio @ricardobasilio_]

@ArquivoWeb_PT Image search 2020-2021, explained by @andmou on YT: https://t.co/5o8VJoo13Z

[Interruptor @InterruptorPT]

Como estamos a trabalhar num artigo sobre arquivos digitais, gostávamos de saber: qual a vossa recordação favorita da web portuguesa?

[Luis Bravo Martins @o_bravomartins]

An open source AR Cloud project that you can start working on today. :) Big up, @OlenkaPolak @andmou for this community-driven push, that can also be used as an educational tool. Let's make those hackathons happen. :)

https://t.co/mtawf0tFGf

[ricardobasilio @ricardobasilio_]

8 de maio, às 15 horas, «API do https://t.co/F1gZuLoDZR. Como processar informação a grande escala?» por @andmou André Mourão.
Inscreva-se aqui: https://t.co/Rr6nHZRPTF. Até sexta-feira!👍

[Arquivo.pt @ArquivoWeb_PT]

https://t.co/al71F7VZe7 da autoria de @andmou e @flaviomartins será o tema desta sessão. Uma oportunidade saber como funciona esta ferramenta no contexto da Imprensa e para colocar as suas questões.
Inscreva-se aqui: https://t.co/XEFN6b6xdn

[Arquivo.pt @ArquivoWeb_PT]

Recordamos as sensações dos vencendores do ano passado. 2ºs classificados (@flaviomartins e @andmou): https://t.co/ge8lqVUifX - Des-cobre as notícias.

Prepare um trabalho ou recandidate-se ao Prémio 2020. Esta pode ser a sua vez. ⏳https://t.co/6Wicm2dgzw