Ankit Bhardwaj’s Post

Struggling with prompt engineering? Here's the ONLY right way to do it👇 1️⃣ Compile an initial dataset: Start with inputs and, if possible, expected outputs. Ideally, the dataset should be large and diversified. If you do not have access to any data, you can look at creating it synthetically or even manually. 10 samples are enough to begin with! 2️⃣ Define evaluation metrics: For open-ended text generations, you would typically use an LLM-as-judge to review the quality of the response, for example on a scale from 1 to 5. If available, compare the response with the expected outputs. 3️⃣ Draft a first version of your prompt, keep it simple! 4️⃣ Run evaluations: Use tools like PromptLayer, which is a user-friendly prompt management and evaluation tool and requires no coding. Another popular option is LangSmith. Or create your own evaluation framework customised to your use case. 5️⃣ Review and iterate: Analyze the evaluation results, add data samples to increase variety, and test for edge cases. Every time you edit your prompt, you should re-run evaluation to test for regressions and ensure the scores are improving. ... and keep iterating 😁 If you need help improving and evaluating the quality of your prompts, feel free to book a consultation via my profile, I'd love to help you! #AI #PrompEngineering #GenAI #LLMs --- https://lnkd.in/gHTCKtsd https://lnkd.in/gsZtmK-Z

🚀 Anthropic’s Model Context Protocol (MCP) is a Game-Changer for Enterprise AI. Why? LLMs produce the best results when they have access to relevant data.  But connecting your enterprise data and the models has been difficult. Engineers have resorted to costly, custom solutions to bridge the gap – until now. The Model Context Protocol establishes a standardized method for LLMs to access and query enterprise data directly. Imagine a seamless integration with your databases, file systems, content and code repositories, without additional coding.

The practical application of LLMs for Enterprise is apparent when you start to consider their deployment in replacing call centers.

Excited to have explored the power of prebuilt Document Intelligence models from Microsoft Azure in the recent challenge! 📄🤖 These models streamline document processing and enable quick extraction of key insights, allowing for more efficient workflows. Looking forward to leveraging these capabilities in future projects! #AzureAI #DocumentIntelligence #MicrosoftAzure #AI #Automation #TechInnovation #LearningJourney

Why You Should Use Continuous Integration and Continuous Deployment in Your Machine Learning Projects Beautifully presented here https://lnkd.in/gnQvuCH3

𝗪𝗵𝗮𝘁 𝗮𝗿𝗲 𝘁𝗼𝗽 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝗽𝗮𝘁𝘁𝗲𝗿𝗻𝘀 𝗳𝗼𝗿 𝗶𝗻𝘁𝗲𝗴𝗿𝗮𝘁𝗶𝗻𝗴 𝗟𝗟𝗠𝘀 𝗶𝗻𝘁𝗼 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 & 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝘀? Here are top 9 that are most relevant in FY24-25 that will help you build production level AI powered solutions: 1. 𝙀𝙫𝙖𝙡𝙨: To measure LLM performance 2. 𝙍𝘼𝙂: To add recent, external knowledge 3. 𝙁𝙞𝙣𝙚-𝙩𝙪𝙣𝙞𝙣𝙜: To get better at specific tasks/domain 4. 𝘾𝙖𝙘𝙝𝙞𝙣𝙜: To reduce latency & cost 5. 𝙏𝙧𝙪𝙨𝙩 & 𝙂𝙪𝙖𝙧𝙙𝙧𝙖𝙞𝙡𝙨: To ensure output quality and trust 6. 𝙃𝙪𝙢𝙖𝙣 + 𝘼𝙄 𝙐𝙓: To provide explanation, handle errors 7. 𝘾𝙤𝙡𝙡𝙚𝙘𝙩 𝙪𝙨𝙚𝙧 𝙛𝙚𝙚𝙙𝙗𝙖𝙘𝙠: To build our data flywheel 8. 𝙈𝙤𝙣𝙞𝙩𝙤𝙧𝙞𝙣𝙜 & 𝙊𝙗𝙨𝙚𝙧𝙫𝙖𝙗𝙞𝙡𝙞𝙩𝙮: To monitor quality, perf, cost in prod 9. 𝘼𝙜𝙚𝙣𝙩𝙨: To break down a goal in to set of tasks #llm #pattern #architecture

From Copilot to Pilots: Introducing AFlow and the Evolution from Chat to Agents In the dynamic world of AI, we are witnessing a significant shift from traditional chat experiences to more sophisticated agentic workflows and autonomous agents. An AI agent is a system capable of autonomous action in an environment to meet its designed objectives. Unlike chatbots, which are primarily reactive and follow predefined scripts, AI agents can make decisions, learn from interactions, and adapt to new information. 🤖 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐖𝐨𝐫𝐤𝐟𝐥𝐨𝐰𝐬 𝐯𝐬. 𝐀𝐮𝐭𝐨𝐧𝐨𝐦𝐨𝐮𝐬 𝐀𝐠𝐞𝐧𝐭𝐬 Agentic workflows completes tasks statically through predefined processes with multiple LLM invocations. Autonomous agents solves problems dynamically through flexible autonomous decision-making. Recent work aims to automate the design of agentic workflows by automated prompt optimization, hyperparameter optimization, and automated workflow optimization. 🆕 𝐈𝐧𝐭𝐫𝐨𝐝𝐮𝐜𝐢𝐧𝐠 𝐀𝐅𝐥𝐨𝐰 AFlow is an innovative framework introduced by Jiayi Zhang in an October 2024 paper, “AFlow: Automating Agentic Workflow Generation.” This framework automates the design of agentic workflows using large language models (LLMs), optimizing workflows through iterative refinement. AFlow represents a significant advancement in creating more efficient and adaptable AI systems. Empirical evaluations across six benchmark datasets (HumanEval, MBPP, MATH, GSM8K, HotPotQA, and DROP) demonstrate AFlow’s efficacy, yielding a 5.7% average improvement over state-of-the-art baselines and a 19.5% improvement over existing automated approaches. Additionally, AFlow enables smaller models to outperform GPT-4 on specific tasks at 4.55% of its inference cost. For more details, check out the AFlow paper on arXiv: AFlow: Automating Agentic Workflow Generation The AFlow paper on arXiv: https://lnkd.in/edc7iDmE For more details, watch the explanatory video on AFlow: https://lnkd.in/eDWFPgZZ Philippe Cordier Etienne Grass

Can you explain the difference between some of the logic-based or #rule-based or #heuristics-based or expert systems and #ML today? Want to understand this more? Here's a link to my related blogpost on the topic: https://lnkd.in/gn6uFKHs

In 2024 we will see more specialized LLMs which will lead to more accuracy and unlock many business use cases...

Scaling and more data alone won't achieve human-level intelligence in LLMs. Yann LeCun sees JEPA as a key step. Our new article covers: - How JEPA works - How it differs from transformers - Models based on JEPA: I-JEPA, MC-JEPA, V-JEPA https://lnkd.in/ehqXW6T7