RAG - just a poor engineering workaround?
RAG - just a poor engineering workaround?
My week kicked off nicely with some inspiring talks on an internal conference. In one of the talks Johannes Langer dived deep on how to build production-ready RAG systems. I answered his opening questions to the audience - โWhat is RAG?โ - with โ๐๐ผ๐ณ๐ณ๐ฒ๐ฟ๐ธ๐น๐ฎ๐๐๐๐ฟโ, which translates to ๐ผ๐ฝ๐ฒ๐ป ๐ฏ๐ผ๐ผ๐ธ ๐๐ฒ๐๐ in my head.
Thinking more about this analogy, I find it is helpful to approach the question if RAG is just a workaround to overcome limitations of our current foundation models or is here to stay, one a more conceptual level. The German wikipedia article on โ๐๐ผ๐ณ๐ณ๐ฒ๐ฟ๐ธ๐น๐ฎ๐๐๐๐ฟโ talks about some of the motivations for this kind of test: huge efforts for students on memorising independent facts are eliminated, the test scope can be wider and the test is focussing more on the ability to creatively think and find new solutions approaches. In other words this approach is frugal with students resources and incentives creation of new solutions.
โ Are you into building Generative AI applications an looking for a good way to start?
โ Are you into building Generative AI applications an looking for a good way to start?
๐ฆ Oliver Mรถller and I run a webinar on how to build Generative AI Applications. From “just” using a LLM to more complex workflows - we guide you through building your first GenAI Application on top of Amazon Bedrock, which brings a lot of flexibility and takes away a lot of undifferentiated heavy lifting from you as an application developer.
๐ On my way to IBC - International Broadcasting Convention 2024. Reflecting on what happened since l
๐ On my way to IBC - International Broadcasting Convention 2024. Reflecting on what happened since last yearโs IBC. Leaving inflated expectations on AGI aside, itโs amazing to see how much progress we see in the adoption of (Gen)AI in M&E industry to improve end-user experience and drive successful monetisation of content, which eventually can translate in more and more-interesting content for end-users.
๐ก Having that said, one of the key impact of the strong GenAI focus in the society was thatย - finally - โtraditional AI got more and more adopted. Not every use case requires a frontier foundation model ;)
๐ฏ Wrapping up my visit in Armenia and the DataFest Yerevan over a very good cup
๐ฏ Wrapping up my visit in Armenia and the DataFest Yerevan over a very good cup of coffee.
โค๏ธ Letโs start with the conference. A lovely crafted event, right sized so that you have a chance to interact with people and a very nice selection of speakers. Soon it felt like family.
๐ธ Naturally picking out just a few talks isnโt fair to the others. I listend to many very interesting talks, though I like always indexed on getting in touch with people & exchange ideas and view - the real value of conference IMHO - and hence only participated in a very real subset of talks.
๐ Just back from vacation - my heart is full of amazing impressions of the alps - I'm already up to
๐ Just back from vacation - my heart is full of amazing impressions of the alps - I’m already up to new exciting adventures. I arrived at the DataFestival in Yerevan(Armenia, DataFest Yerevan). A beautiful crafted conference. We met some of the speakers already last night - I can tell you it is a super exciting agenda waiting for you.
๐ก Nensi Hakobjanyan and I finished the final prep for out talk โUnlocking the Power of LLMs: Next generation recommender systemsโ which will be on this afternoon (4:30). If you are around make sure that you drop by - it will be AWSome ๐
Awesome event, indeed Lajos Lange. Extending your kudos to the team - I had a fantastic time yesterd
Awesome event, indeed Lajos Lange. Extending your kudos to the team - I had a fantastic time yesterday. #GenerativeAI is a very powerful technology applicable across different industries. What really stands out for me is that is a powerful tool not just for tech folks but also for non-technical people, who can apply it without a steep learning curve.
Yesterday we talked about how Prompt Engineering can improve results with little effort. Can’t wait to see what participants are going to build from there ๐.
Last week I had the pleasure to join and present at the AWS GenAI Roundtable: Transforming the Marke
Last week I had the pleasure to join and present at the AWS GenAI Roundtable: Transforming the Marketing, Advertising, and Publishing Industry (https://lnkd.in/epV8Wa26) event.
Great opportunity to share my holiday pics to a wider audience. Kidding and not kidding. Used those to visualise the inflated expectations of LLM capabilities we experience and at the same time talked about the fantastic opportunities LLMs bring to the various industries. Check out Mark Watkinsโs post on the event(https://lnkd.in/ey8Q8WiF), where he is referencing some of the demos the team showcased in the event and available online for your own consumption.ย Highly recommended to have a look. Source code available on GitHub!
โ Successful Building a GenAI use cases just requires the latest and greatest fr
โ Successful Building a GenAI use cases just requires the latest and greatest frontier model, right?!
๐ In my conversations with customers I often realize that the choice of the best & shiniest model is highly occupying their resources, while thinking and focusing on architecting and building the use cases which ultimately should fulfil users needs gets very little attention. This naturally fuelled by the fast-pace and loud announcements of new frontier models, which come withย superior benchmark results and new capabilities. What we often forget is that the majority of capabilities become commodity among different models very fast. Hence it often makes much more sense to focus on the use cases and building the generative application for the use case in a way it really addresses the userโs need and is able to evolve to utilize new release models where this makes sense. I wrote about this inย my blog post (https://lnkd.in/e_YhCinM.)
๐ Bagrat Ter-Akopyan, Carmen Heger and team. Congrats and thanks for the nice technical write-up.
๐ Bagrat Ter-Akopyan, Carmen Heger and team. Congrats and thanks for the nice technical write-up. It’s amazing to see how you continue to innovate on behalf of your customers. I love the outcome.
๐ What stands out for me from your technical report is your description of the why and how you moved away from your own, already very good, initial RAG implementation. Citing from the report:
“๐๐ฏ๐ฆ ๐ฐ๐ง ๐ฐ๐ถ๐ณ ๐ญ๐ฆ๐ข๐ณ๐ฏ๐ช๐ฏ๐จ๐ด ๐ง๐ณ๐ฐ๐ฎ ๐ต๐ฉ๐ฆ ๐ญ๐ข๐ด๐ต ๐ฃ๐ช๐จ ๐๐๐ ๐ฑ๐ณ๐ฐ๐ซ๐ฆ๐ค๐ต ๐ธ๐ข๐ด ๐ต๐ฉ๐ข๐ต ๐ธ๐ฆ ๐ฏ๐ฆ๐ฆ๐ฅ๐ฆ๐ฅ ๐ต๐ฐ ๐ด๐ฑ๐ฆ๐ฆ๐ฅ ๐ถ๐ฑ ๐ฐ๐ถ๐ณ ๐ฆ๐น๐ฑ๐ฆ๐ณ๐ช๐ฎ๐ฆ๐ฏ๐ต๐ข๐ต๐ช๐ฐ๐ฏ ๐ค๐บ๐ค๐ญ๐ฆ๐ด. ๐๐ฏ ๐ต๐ฉ๐ฆ ๐ต๐ฆ๐ค๐ฉ๐ฏ๐ช๐ค๐ข๐ญ ๐ด๐ช๐ฅ๐ฆ, ๐ต๐ฉ๐ข๐ต ๐ฎ๐ฆ๐ข๐ฏ๐ต ๐ถ๐ด๐ช๐ฏ๐จ ๐ฎ๐ฐ๐ณ๐ฆ ๐ฐ๐ง๐ง-๐ต๐ฉ๐ฆ-๐ด๐ฉ๐ฆ๐ญ๐ง ๐ด๐ฐ๐ญ๐ถ๐ต๐ช๐ฐ๐ฏ๐ด ๐ง๐ฐ๐ณ ๐๐๐ ๐ข๐ณ๐ค๐ฉ๐ช๐ต๐ฆ๐ค๐ต๐ถ๐ณ๐ฆ๐ด ๐ต๐ฉ๐ข๐ต ๐ฉ๐ข๐ท๐ฆ ๐ฃ๐ฆ๐ค๐ฐ๐ฎ๐ฆ ๐ข๐ท๐ข๐ช๐ญ๐ข๐ฃ๐ญ๐ฆ ๐ข๐ฏ๐ฅ ๐ณ๐ฆ-๐ถ๐ด๐ช๐ฏ๐จ ๐ฐ๐ถ๐ณ ๐ฐ๐ธ๐ฏ ๐ฆ๐น๐ช๐ด๐ต๐ช๐ฏ๐จ ๐ด๐ฐ๐ญ๐ถ๐ต๐ช๐ฐ๐ฏ๐ด.” and “๐๐ฐ ๐ฎ๐ข๐ฌ๐ฆ ๐ต๐ฉ๐ฆ ๐ฅ๐ฐ๐ค๐ถ๐ฎ๐ฆ๐ฏ๐ต๐ด ๐ด๐ฆ๐ข๐ณ๐ค๐ฉ๐ข๐ฃ๐ญ๐ฆ ๐ง๐ฐ๐ณ ๐ฐ๐ถ๐ณ ๐๐๐ ๐ด๐บ๐ด๐ต๐ฆ๐ฎ, ๐ธ๐ฆ ๐ถ๐ด๐ฆ๐ฅ ๐๐๐ ๐ฌ๐ฏ๐ฐ๐ธ๐ญ๐ฆ๐ฅ๐จ๐ฆ ๐ฃ๐ข๐ด๐ฆ๐ด. ๐๐ฉ๐ช๐ด ๐จ๐ข๐ท๐ฆ ๐ถ๐ด ๐ด๐ญ๐ช๐จ๐ฉ๐ต๐ญ๐บ ๐ง๐ฆ๐ธ๐ฆ๐ณ ๐ค๐ฐ๐ฏ๐ง๐ช๐จ๐ถ๐ณ๐ข๐ต๐ช๐ฐ๐ฏ ๐ฐ๐ฑ๐ต๐ช๐ฐ๐ฏ๐ด ๐ต๐ฉ๐ข๐ฏ ๐ข ๐ค๐ถ๐ด๐ต๐ฐ๐ฎ ๐ฃ๐ถ๐ช๐ญ๐ฅ ๐ฆ๐ข๐ณ๐ญ๐ช๐ฆ๐ณ ๐ด๐ฐ๐ญ๐ถ๐ต๐ช๐ฐ๐ฏ ๐ฃ๐ถ๐ต ๐ด๐ฑ๐ฆ๐ฅ ๐ถ๐ฑ ๐ช๐ฏ๐จ๐ฆ๐ด๐ต๐ช๐ฐ๐ฏ ๐ฃ๐บ ๐ด๐ฆ๐ท๐ฆ๐ณ๐ข๐ญ ๐ฎ๐ข๐จ๐ฏ๐ช๐ต๐ถ๐ฅ๐ฆ๐ด ๐ธ๐ฉ๐ช๐ค๐ฉ ๐ข๐ญ๐ญ๐ฐ๐ธ๐ฆ๐ฅ ๐ถ๐ด ๐ต๐ฐ ๐ช๐ต๐ฆ๐ณ๐ข๐ต๐ฆ ๐ฑ๐ณ๐ฆ-๐ฑ๐ณ๐ฐ๐ค๐ฆ๐ด๐ด๐ช๐ฏ๐จ ๐ฎ๐ถ๐ค๐ฉ ๐ง๐ข๐ด๐ต๐ฆ๐ณ.”
๐ข If you are on the hunt for an image ๐๐ก๐ ๐ฉ๐๐๐๐ข segmentation model, which is open and you can deploy
๐ข If you are on the hunt for an image ๐๐ก๐ ๐ฉ๐๐๐๐ข segmentation model, which is open and you can deploy on your own, have a look at the just released ๐ฆ๐ฒ๐ด๐บ๐ฒ๐ป๐ ๐๐ป๐๐๐ต๐ถ๐ป๐ด ๐ ๐ผ๐ฑ๐ฒ๐น ๐ฎ (๐ฆ๐๐ ๐ฎ). The model capabilities can be nicely experienced in Metaโs Demo). Read more about the announcement at their announcement page.
๐ทโโ๏ธIf you are looking into deploying the model to build your own application on AWS, ๐๐บ๐ฎ๐๐ผ๐ป ๐ฆ๐ฎ๐ด๐ฒ๐ ๐ฎ๐ธ๐ฒ๐ฟ is a very good alternative for you. Quoting from Metaโs announcement website: