Dr. Ruth Janning

I still remember very clearly how I trained a semantic bilingual German-English embedding model almost four years ago. Back then for T-Systems on site services GmbH. Nowadays, with the hype of #RAG, it is probably one of the most popular German-language open source models, measured by the more than one million downloads per month. This success fills me with joy and pride and also with some doubts. But why the doubts? I think it is important to understand that there is a big difference between semantic embeddings and Q/A retrieval models. Semantic embeddings can be used to cluster texts. Or, for example, to search large texts on the basis of a few keywords. However, they are less suitable for finding answers to questions. The reason is that the semantic similarity between a question and a text with the answer does not necessarily have to be high. For this reason, Q/A retrieval models that are trained to embed questions and potential answers close to each other are primarily suitable for retrieval in RAG systems. I'm afraid many are using my semantic embedding as a replacement for a Q/A retrieval model in a RAG application. This should not be done. At Telekom, we use self-trained Q/A retrieval models for our RAG retrieval. We also have our own data sets for this. For other EU languages, by the way, we have had very good experiences with the intfloat/multilingual-e5-large model. Incidentally, this also works very well for German. - my semantic bilingual German-English embedding called "T-Systems-onsite/cross-en-de-roberta-sentence-transformer": https://lnkd.in/eSx6kc6m - the "intfloat/multilingual-e5-large model" model: https://lnkd.in/e7aetC3t Deutsche Telekom #AICC #iHub #PIX #weloveai #VTI #GenAI #LLM

178

21 Kommentare

The Essential GraphRAG Patterns

3

Speaking about my participation in social dataviz challenges like #30DayChartChallenge at the Dataviz Meetup Bremen yesterday was great fun. Thanks to the group for all of your questions. I’m sharing the links to some resources that were requested: 💻 Code: ▸ I share the code for most of my (code-based) charts on Github. The repository for my plots created for the #30DayChartChallenge: https://lnkd.in/eWdkGpTV ▸ And my contributions to the 2023 #30DayMapChallenge: https://lnkd.in/emPteyQQ (change the year in the URL for 2021 and 2022). 📂 Datasets: ▸ On the Github repo for the #30DayChartChallenge 2024, you will find a collection of data sources, including links to the data repos for other dataviz challenges like #TidyTuesday or #MakeoverMonday https://lnkd.in/emPteyQQ  ▸ Statistisches Bundesamt / Destatis (German Statistics Office): https://www.destatis.de/EN ▸ Spotify API, see https://lnkd.in/eVwaZqBx. I use the {spotifyR} package to access the data. ▸ FBref.com for football data such as player career statistics and match results, accessed via {worldfootballR} package ▸ The Office full script: https://lnkd.in/eZ4shabx  ✨ Design: ▸ Material Design: https://lnkd.in/e-J54Y4A You can find the slides of the talk on Github: https://lnkd.in/eJr8magj Thanks again for the invitation to Dr. Niels Ohlsen, Gabriela Molina León, and Moritz Stefaner. #dataviz #datavisualization #ggplot2 

54

2 Kommentare

Setup an Advanced Multimodal RAG Pipeline in Minutes

5

CharXiv: A Comprehensive Evaluation Suite Advancing Multimodal Large Language Models Through Realistic Chart Understanding Benchmarks https://buff.ly/4bpCNiv

1

GraphRAG Microsoft recently open-sourced GraphRAG, https://lnkd.in/en_Qa5hU, a tool to develop pipelines to extract knowledge graphs from raw text. A decade ago, I worked on a project with Vineel Gujjar to construct knowledge graphs from aircraft technician notes. A cascaded design involved a system extracting entities such as "issues," "parts," and "fixes" and encoding their relationship in a knowledge graph to ultimately figure out the best "fix" for an "issue" involving a "part." LLM-based NLP systems augmented with GraphRAG approach would have made it much easier to rapidly develop and deploy complex systems like the one we developed a decade ago. Ramasubramanian (Ramsu) Sundararajan Debasis Bal

34

2 Kommentare

Building a Multimodal RAG Pipeline with Claude 3 (Anthropic) and SingleStore A lot of data out in the wild is in the form of images, even within documents, and we see multimodal RAG playing a bigger and bigger role in the future.

3

Making Sense of Data Context: A Simple Guide We often see LinkedIn posts and read about "understanding data context" but what does this really mean in practice? For me, it boils down to asking these three questions in a respectful, safe environment where no one feels offended: • Inquire about data sources: • Would you mind sharing the sources of this data? • Once you know the sources, ask: • Who can tell me which systems touched this data along along the way? • Before delivering the dashboards or reports ask: • Why does it matter to our users? Pro tip: Ask these questions early in your data projects. They'll save you time, build trust with stakeholders, and ensure your analysis actually drives value. Remember: Better context leads to better decisions but make sure the environment is safe before you shoot your questions. What other things you do to learn the data context better? Please share. #businessIntelligence #dataContext #databestpractices #dataleadership #bi #dataengineering

6

Ever wondered why you studied all those complex linear algebra concepts like matrix multiplication, covariance or eigen vectors in college. They form the foundation for PCA and even face/image recognition models in machine learning… of course it is one of the many exciting areas to apply the conceptual mathematics in real world situations. Mathematics forms the foundation for AI and machine learning models. Where do YOU use your college math concepts in your real world profession?

19

2 Kommentare

In this notebook, we will first explore what is text embeddings and vector stores. Then we will cover how to create and store text Embeddings in vector stores with LangChain. We will conclude this notebook with some failure cases of this method.

3

Speeding up index creation Let's talk about speeding up index creation in PostgreSQL. If you've ever had the pleasure of waiting for hours while your database churns away at creating indexes on a large dataset, you know the feeling. Fortunately, Hans-Jürgen Schönig has some insights that may help you cut that time significantly. Indexes are crucial for optimal query performance. But creating them can feel like watching paint dry. Thankfully, there are techniques floating around that can make the process a bit less torturous. Speaking of insights, PGConf.dev 2024 had a lineup of talks worth tuning into if you missed it. Topics ranged from autovacuum mishaps to scaling databases at petabyte levels. Definitely a treasure trove for any PostgreSQL enthusiast. Also, if you fancy learning some magic tricks for psql, Elizabeth Christensen shares some nifty settings and presets that can make your life easier. So, whether you're indexing or attending conferences, the PostgreSQL community is here to keep you entertained and informed. Cheers to faster queries and smarter database management! Source: https://lnkd.in/e8SSrvJd

2

𝗦𝗼𝗹𝘃𝗶𝗻𝗴 𝘁𝗵𝗲 𝗨𝗻𝘀𝗼𝘃𝗮𝗯𝗹𝗲: Leveraging LLMs for Real-World Problems 💡 Before LLMs, automation mainly dealt with structured digital data, unless you invested heavily in specialized models, which still left many areas untouched. Structured (tabular) digital data was essentially the cornerstone of IT. Then came LLMs. They impressed me with their ability to understand and contextualize virtually any data within their training scope. This capability allows us to extract structure from unstructured data, interpret scenarios, and translate artefacts like never before. To harness this potential for maximum value, consider these points: 🤔 What is unstructured data? Any data request not yet translated into code is unstructured as much as more typically identified irregular invoicing formats. The mind-stretch of what is unstructured is key to ideation with LLMs. 🔎 Identify human-driven, sequential, and frequent processes that involve interpretation and restructuring. Rather than asking for “structured digital” (like we did with RPA), we should focus on “sequential and frequent” processes. 🧐 Recognize data and artefacts we currently lack insights into and determine which might be covered by LLM training data (and thus workable by them). At Basico we tackle real-world problems using these points of enquiry. The best part? We solve the problems swiftly, because we work with pretrained models that are easy to configure. By combining these models with our automation experience and best practices, we expand the realm of digital and IT. This expansion marks a genuine, immediate paradigm shift in IT and automation💡 #ai #llm #genai #technology #rpa #automation 🤝 #Connect or #Follow for a blend of serious insights and lighthearted humor about everything IT and business 🎧 #Listen to The Only Constant podcast

32

2 Kommentare

A few years ago, I have seen many organizations invest millions of dollars in products that delivered similar features! Today, as AI evolves, many of those products are becoming obsolete. #AI #MachineLearning #DigitalTransformation #ThoughtLeadership #AITransformation

3

1 Kommentar

Kolmogorov–Arnold Networks seem like a paradigm shift in ML architectures. Curious to see how they evolve. "While MLPs have fixed activation functions on nodes ("neurons"), KANs have learnable activation functions on edges ("weights"). KANs have no linear weights at all -- every weight parameter is replaced by a univariate function parametrized as a spline. " https://lnkd.in/gdPk_qCQ

29

"LLMs are but RecSys in Disguise" 🥸 - Chang She, Pandas, LanceDB ⚡This blog post explores Retrieval-Augmented Generation (RAG), a method that combines information retrieval with language generation for Recommendation Systems. It examines RAG's potential promising approach to improve recommendation quality and discusses its advantages and limitations. 🗝️ Key Points: 🔸 LLMs for Recommendations: Large language models (LLMs) offer a new approach to recommendations by understanding and generating human-like text. However, directly prompting LLMs for recommendations can be ineffective due to limited context memory, hallucinations, bias, and slow inference speed. 🔸 RAG to the Rescue: RAG addresses these limitations by retrieving relevant information before generating recommendations. This allows LLMs to leverage a broader knowledge base and produce more accurate and personalized suggestions. 🔹 Building a RAG System: Here's a breakdown of building a RAG system: 🔸Data Ingestion and Retrieval: Store candidate items and search for the most relevant ones using vector search techniques. 🔹 Encoding User Queries: Encode user queries into semantic vectors to find similar items in the database. 🔸 LLM for Reranking: Use an LLM like GPT-3.5 to refine the retrieved items and generate the final recommendations. 🔸 RAG's Benefits: 🔹 Improved Performance: Studies show that RAG can enhance the performance of various recommendation models. 🔸 Long-Term Memory: RAG enables LLMs to effectively utilize a vast amount of data, overcoming context length limitations. 🔹 Multimodal Data Support: RAG can potentially handle various data formats like text, images, and videos. 🔸 Dynamic Updates: The system can be continuously updated with new user interactions and data. 🔹 Challenges and Considerations: 🔸 Training and Maintenance: Careful training and fine-tuning are crucial for optimal RAG performance. Frequent retraining might be necessary for large datasets. 🔹 Long Context Models: While beneficial, using long context models for re-ranking can lead to slow output times, requiring careful engineering to minimize this trade-off. 🔸Hallucination: LLMs can still generate inaccurate recommendations. Techniques like temperature control and fuzzy string matching can help mitigate this issue. 🔹 Data Structure: The effectiveness of RAG relies heavily on the quality of your data embeddings. Poorly structured data can hinder RAG's ability to find the best matches. 🔸 Multimodal Data Research: The effectiveness of RAG with multimodal data requires further investigation. 👉Conclusion: RAG presents a promising approach for enhancing LLM-based RecSys. While challenges exist, RAG's ability to ✅improve recommendation accuracy, ✅ leverage long-term memory, and ✅potentially handle multimodal data makes it a valuable technique to explore. #RecSys #LLMs #GenAI https://lnkd.in/eCaiWAKD

3