Simon Willison’s Post

LLMs can be brittle at generating Cypher, even if you use GPT-4. In my latest blog post, I have written about ensuring robustness of LLMs interacting with a knowledge graph through the implementation of semantic layers with #LangChainAI and @#neo4j https://t.co/LHwVvHEZKW #LLM #LLMs #largelanguagemodels #largelanguagemodel #gpt4 #gpt #graph #robusness #langchain #knowledgegraphs #knowledgegraph #predefinedparameters #parameters #retrieveinformation #query #queryoptimization .

Another paper 🗞 🗞 🗞 Learn about the structured data extraction, emphasizing the growing importance of LLMs like GPT-3.5 and GPT-4 for efficient data extraction from unstructured text. Read more here: https://lnkd.in/g_pCQKK4

Graph data hybrid with vector data store can potentially bring hierarchical concepts and relationships to documents, allowing RAG system to tranverse through document context, generating more accurate response. What we have experimented is only at 1-dimentional hierarchical tree, but what LlamaIndex has launched here is a full scale Graph x Vector framework! Cool stuff! Great jobs to Jerry Liu and the team at Llama Index!

📢 LlamaIndex launches knowledge graph framework for LLMs - 𝐏𝐫𝐨𝐩𝐞𝐫𝐭𝐲 𝐆𝐫𝐚𝐩𝐡 𝐈𝐧𝐝𝐞𝐱🚀 𝐖𝐡𝐲 𝐏𝐫𝐨𝐩𝐞𝐫𝐭𝐲 𝐆𝐫𝐚𝐩𝐡𝐬📈? 🔶Property graphs offer a more expressive and flexible way to model and query knowledge graphs compared to traditional knowledge triples. Here are the key advantages: • 𝐋𝐚𝐛𝐞𝐥 𝐚𝐧𝐝 𝐏𝐫𝐨𝐩𝐞𝐫𝐭𝐲 𝐒𝐮𝐩𝐩𝐨𝐫𝐭: Assign labels and properties to nodes and relationships, providing richer metadata. • 𝐕𝐞𝐜𝐭𝐨𝐫 𝐄𝐦𝐛𝐞𝐝𝐝𝐢𝐧𝐠𝐬: Represent text nodes as vector embeddings for hybrid search capabilities. • 𝐇𝐲𝐛𝐫𝐢𝐝 𝐑𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥: Perform both vector and symbolic retrieval for more comprehensive querying. • 𝐂𝐲𝐩𝐡𝐞𝐫 𝐐𝐮𝐞𝐫𝐲 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞: Express complex queries using the Cypher graph query language. • 𝐓𝐲𝐩𝐞 𝐂𝐚𝐭𝐞𝐠𝐨𝐫𝐢𝐳𝐚𝐭𝐢𝐨𝐧: Categorize nodes and relationships into types with associated metadata. • 𝐒𝐮𝐩𝐞𝐫𝐬𝐞𝐭 𝐨𝐟 𝐕𝐞𝐜𝐭𝐨𝐫 𝐃𝐚𝐭𝐚𝐛𝐚𝐬𝐞: Treat your graph as a superset of a vector database for hybrid search. 🔶These features make property graphs a powerful and flexible choice for building knowledge graphs with large language models (LLMs). ✒️ Read a full blog here: https://lnkd.in/gh9Veyyb #GenAI #LLM #KnowledgeGraph

Fine Tune BERT for Text Classification with TensorFlow 

This week's #KeyAlgorithms article explains #LogisticRegression, a simple classifier that is also the mathematical link between #Bayesian algorithms and #NeuralNetworks https://lnkd.in/eautPPjS #DataScience #ArtificialIntelligence #MachineLearning

Complex Document RAG with GPT-4o 📑 GPT-4o’s multimodal capabilities means that it’s great at parsing complex PDFs/slide decks with background images, irregular layouts, charts into semi-structured markdown with text and tables. We’ve natively integrated GPT-4o with LlamaParse letting you use it as a key part in the RAG ETL step. Our cookbook below shows you how to parse the 2019 Tesla impact report deck into text representations, and ask questions over chart data in the original deck 🔥 Bonus 💫: GPT-4o on LlamaParse is now $0.03 per page! (20x reduction from before). Notebook: https://lnkd.in/grwUVr-G Signup for LlamaParse: https://lnkd.in/gi8dxGnt LlamaParse repo: https://lnkd.in/g3UmUkcD

New Blog Post from Taylor Richardson and Michael S. Hoffmann! Querying Complex Tabular Datasets with LLMs: Learnings from the Lab At Kensho, we're tackling a major challenge for #GenAI. While querying text data with #LLMs is feasible using vector databases, querying tabular/relational data remains difficult. LLMs struggle with the complex SQL needed to join multiple interrelated tables. Our solution? Creating specialized "LLM-Ready" APIs that abstract away the underlying data complexity. We've had initial success with this approach for S&P Global's Transactions dataset. The LLM-Ready API allows users to ask natural language questions like "What were the biggest M&A deals in German tech in 2021?" and get precise answers from the tabular data. Read this blog to learn more! The link is in the comments. #blog #data

Document RAG with GPT-4o explained the chart as tables.

Want to boost the performance and accuracy of your large language models (#LLMs)? This blog post dives into how LlamaIndex, a vector search library, can be leveraged with Supabase, a popular backend platform, to store and retrieve #vectors efficiently. Learn how to extract document headings and build an index within your Supabase database for faster information retrieval. This approach is ideal for managing and retrieving data from research papers, articles, or any structured text content with headings. Check out the article to learn more about how #LlamaIndex and #Supabase can help you improve your LLM applications! - https://lnkd.in/dDsHSwaB #goML