Emissary的动态

We don't make decisions in words, we just explain them so. At Emissary, we're working on creating a new class of discriminative language models that are faster, more deterministic and incredibly easy to build. They're designed for decision-making, not just generating words. Excited to launch the first one - the Classification LMs. If you're working on classification steps in your workflows and agents using LLMs and looking for a deterministic way to improve, take a shot on finetuning your own classification LM!

If you're spending hours manually tweaking prompts to LLMs, iterating over and over again, changing words, examples, structure to get better results.. Don't you wonder if theres an easier way to do this? This isn't very scientific is it? Don't you ever wonder whats the most "optimal" prompt for my task? Enter automatic prompt engineering or prompt optimization. This is an absolutely fantastic article written by Heiko Hotz and I highly recommend reading this FIRST to understand how this works then begin exploring the various tools/methods out there. Automated Prompt Engineering: The Definitive Hands-On Guide: https://lnkd.in/ehqFfPa2 Here are some of my favorite tools out there for this, I think once you give them a try you will really have an eye-opening moment: ?? Prompt Poet: https://lnkd.in/euzGvw2S ?? Quality Prompts: https://lnkd.in/eJ8R2Gd7 ?? AutoPrompt: https://lnkd.in/eEWyxbjd ?? DSPy: https://lnkd.in/edN8XGsb ?? TextGrad: https://lnkd.in/eMqAYdYa ?? AdalFlow: https://lnkd.in/eeEauCYx Bonus --> Awesome LLM Prompt Optimization: https://lnkd.in/eWYTjtmc If you have an LLM application/pipeline with multiple prompts, prompt chains, all the minutiae and nuances that make up each individual prompt, how can we optimize the entire pipeline to increase the quality of our output..see where I'm going with this. Comment below if you've come across any other automatic prompt engineering tools.

?? The Open Source Tool That Solves Document Parsing for LLMs – MegaParse! ?? Tired of manual document parsing? Say hello to MegaParse, the revolutionary open-source tool designed to make feeding documents into LLMs seamless while preserving formatting, structure, and content integrity. ? Manual parsing is history! Parsing documents has been a bottleneck in LLM workflows—slow, error-prone, and tedious. MegaParse changes the game, handling everything from PDFs to Excel files effortlessly. ?? Benchmark Results Speak for Themselves: ? MegaParse Vision: ? 87% similarity ratio ? Unstructured (Check Table): ?? 77% ? Basic Unstructured: ?? 59% ? LlamaParser: ? 33% ? Powered by Cutting-Edge Multimodal Models! MegaParse’s Vision variant utilizes advanced models like Claude 3.5, Claude 4, and GPT-4V, delivering unparalleled accuracy in document parsing. ?? What challenges do you face when preparing data for LLMs? Share your thoughts below and join the conversation on how MegaParse can optimize your workflow! ?? Discover more in the paper: https://lnkd.in/dbmf9dj8

?? Struggling with Destructured Documents in Your Workflow? ?? Restructuring destructured documents, especially those containing both drawings and text, can be a nightmare. But it's a crucial task if you want to recover the original structure and make sense of the data. The good news? Advanced techniques are making this complex operation more efficient than ever. Here’s how we tackle it: ??Interpreting multiple file formats to handle various destructured documents. ??Applying stringology algorithms to uncover patterns and repetitions. ??Leveraging data mining and KDD (Knowledge Discovery in Databases) to recover lost structures. ??Using cutting-edge AI and deep-learning algorithms to automate and streamline the process. ??Employing Computer Human Interfaces (CHI) for a more intuitive experience. At 1A3i, we specialize in restructuring graphic documents, where text meets drawings, helping you bring back order from chaos. ?? Discover how COMPARE Software can help recover your destructured documents today! ?? Learn more at 1a3i.com/en. ?? [email protected] | ?? +91 9963713841 | ?? firstDCS.in #DocumentRecovery #AIInTech #DataMining #GraphicDocuments #DeepLearning #COMPARESoftware #RestructuringDocuments #WorkSmart #TechInnovation #DocumentManagement

When optimizing RAG, teams often obsess over the shiny thing— embedding, rerankers, or other complex techniques. Yet, the?? is somewhere else.. In our recent work, we saw that the most valuable uplift can come from the simplest part — chunking. Here's something that really surprised me: After testing different chunk sizes, simply switching to page-sized chunks (2048 tokens) alone improved our hit rate by 27.7 points. That's a massive jump—and it happened without changing anything in our embedding models or metadata. Just chunk size. But, as with most things, there's a catch. Larger chunks bring both improvements and challenges: - Improved context understanding - Better retrieval accuracy - Higher computational costs - Increased processing time So how do you find the optimal chunk size for your system? Well, it depends. We started with a simple line search, checking different variations of the param. We also went full “Bayesian Optimization” mode with Optuna, but this is typically overkill. The use case and complexity of your data are important. We worked with financial PDFs so it made a lot of sense to go for —- page-sized chunks. I know…Is it even considered “chunking” if you take the whole page? Why not simply use Long Context? Honestly, I didn’t care as long as the results were dramatically better. Long-context has shown in the past it’s power in other experiments (see first comment). I still believe that the future is in RAG. Key learning: Chunking is a powerful tool for improving retrieval accuracy, but it's not just about making everything bigger. You need to test and find the sweet spot that works for your specific needs. What chunk sizes have worked best for your use case?

My Lookup Framework for Machine Learning Systems Over the years, this framework has been proven to simplify my work and makes it easier to identify where I am in the process and what exactly needs to be done during model development and deployment. 1?? Understanding the End Goal I have come to understand, especially within a business context, that an ML system is simply a means to an end. Therefore, I first strive to understand the ultimate goal the model is intended to serve. For example, is it revenue optimization, targeted campaigns, or something else? This goes beyond understanding the problem itself; it includes understanding how the solution aligns with the overall objective. 2?? Defining Success Here, I try to determine whether there is a performance benchmark already set for the solution. What does success look like for the model we are about to build? How will I measure this success? 3?? Identifying Requirements and Resources From the "what" and the "how," I begin to identify the necessary components—the dataset, where to source it, its origin, latency requirements, metrics to collect, and tools for visualization. 4?? System Design Flowchart/Workflow I create a sketch of the complete flow: the workflow, the end-to-end system design. This includes the model pipeline, data pipeline, and all other relevant components. 5?? Building the Model/Training This involves selecting the model architecture, choosing the appropriate model, and computing its performance. 6?? Model Serving and Deployment. I focus on creating an endpoint, using Docker for containerization, and employing tools like Swarm or Kubernetes for orchestration. 7?? Logging and Monitoring I ensure continuous monitoring using tools like Prometheus to track the system’s performance and maintain reliability. #machinelearning #MLSystemDesign

Real-Time Machine Learning Service Deployment Explained ! 1. Feature Retrieval and Validation: Initiates with feature retrieval linked to real-time validation to ensure data quality. 2. Experiment Tracking System: Central hub for managing and tracking machine learning experiments. 3. Model Training and Validation: Involves training ML models and subsequent validation to confirm their accuracy and reliability. 4. Model Registry and Staging: Models are registered and staged, allowing for version control and systematic deployment to production. 5. Deployment Pipeline: A streamlined process that integrates with APIs and manages the deployment of models to the production environment. 6. Load Balancing: Ensures efficient distribution of requests to the deployed models, optimizing resource use and response time. 7. Product Application Integration: Integrates the ML models into product applications, enhancing functionality with smart features. 8. Ranking Engine: Analyzes and ranks data inputs using deployed models to drive decision-making in applications. [Explore More In The Post] Don’t Forget to save this post for later and follow @theaiprofessor for more such information.

Model Validation Techniques, Explained: A Visual Guide with Code Examples 12 must-know methods to validate your machine learning ??? by Samy Baladram

Introducing new pre-trained models in our TotalAgility document library, including the latest addition : the International Safety Data Sheet model! Our latest blog explores how intelligent document libraries with pre-trained extraction models are transforming manual, time-consuming tasks into swift, automated processes. Read the full blog to learn more about our latest document models and recent enhancements: https://ow.ly/2ZvA50StZyn #TungstenAutomation #DocumentProcessing #GenAI #TotalAgility8 #DocumentLibrary

?? Mastering RAG (Retrieval-Augmented Generation): Best Practices You Need to Know! ??? Are you exploring ways to harness the true power of Retrieval-Augmented Generation? Here's a practical guide to make the most out of your RAG workflows! ?? ?? Key Components for Success: 1?? Evaluation ?? Assess General Performance ??? Fine-tune for Specific Domains ?? Optimize Retrieval Capability ?? 2?? Fine-Tuning ?? Experiment with: Disturb ?? Random ?? Normal ? 3?? Chunking ?? Play with Chunk Sizes ?? Utilize Sliding Windows ?? Add Metadata for Context ?? 4?? Embedding ?? Use top-notch models like: LLM-Embedder ?? BGE, Jina, or all-mpnet-base-v2 ?? 5?? Vector Database ??? Go for robust solutions: Milvus, Weaviate, or Qdrant ?? 6?? Query Classification ?? Ensure precise classification for better retrieval. ??? 7?? Retrieval Techniques ???♂? Leverage: BM25 ?? HyDE + Hybrid Search ???? 8?? Repacking for Summarization ?? Options include: Forward, Reverse, or Selective Context ?? 9?? Reranking ?? Enhance with: monoT5, RankLLmAMA, or TILDE ?? ?? Pro Tip: Always keep iterating to find what works best for your use case! Small tweaks can lead to BIG results. This cheat sheet is your roadmap to deploying high-performing RAG systems! ??? What are your go-to practices or tools for RAG? Let’s discuss in the comments! ???? Credit: Piyush Ranjan #RAG #AI #MachineLearning #LLM #BestPractices #AIInnovation #RetrievalAugmentedGeneration #TechLeadership