Llamacon2025: Open-sourced tools for LLM developers

Issue #8 AI is no longer just a layer on top - it’s becoming part of the stack itself. This week’s lineup highlights tools that embed intelligence right where developers work: in databases, crawlers, and runtime environments. 🔹 txtai (David Mezzetti) - All-in-one semantic search framework with rerankers and pipelines so you can build hybrid RAG without the glue. 🔹 pgai (Tiger Data (creators of TimescaleDB), John Pruitt) - Postgres extension and Python library that lets you run AI workflows (embedding, vectorization, and RAG functions) directly in SQL inside your database. 🔹 Promptfoo - CLI-driven LLM evaluation and security testing toolkit with red-team plugins and CI gates. 🔹 OpenLLMetry (Traceloop) - OpenTelemetry-based instrumentation that auto-traces LLM, provider, and vector-DB calls. 🔹 Crawl4AI (Unclecode (Hossein)) - Python crawler that drives Playwright with built-in anti-bot tactics and LLM-aware extractors. 🔹 DSPy 3.0 - Programmatic LLM optimization with self-improving, auto-tuned prompts and modules. 🔹 E2B Desktop Sandbox (Tomas Valenta) - Local-first, containerized runtime that safely executes agent code with filesystem and network controls. #devtools #AIengineering

RAG that dazzles in demos but cracks in production? If you’re tired of flaky answers and silent failures, this is for you. I just published LlamaIndex Review 2025: Best Practices — a practical, no-hype guide to building reliable RAG and agent systems for real apps. What you’ll get: 📚 Chunking that preserves meaning (semantic + structural) for stable recall 🧭 Retriever configs that balance precision/recall without blowing up latency 🛠️ Workflow patterns for tools/agents with guardrails that actually catch loops 📏 Evaluation + telemetry to spot regressions before users do Why this matters: I’ve shipped RAG that looked great until traffic hit—these are the patterns that kept our answers consistent at scale. 👉 Read the full guide: https://lnkd.in/ezsxPuY8 Your turn: if you had to fix one first, which is it—A) chunking/indexing or B) evaluation & telemetry? Drop A or B and tell me why. #llamaindex #rag #llmops #aiengineering #productionai

𝐅𝐢𝐧𝐝 𝐌𝐞𝐝𝐢𝐚𝐧 𝐟𝐫𝐨𝐦 𝐃𝐚𝐭𝐚 𝐒𝐭𝐫𝐞𝐚𝐦 Problem in statement: We need to design a data structure that supports two operations efficiently: 1️⃣ addNum(int num) – Insert a number from a stream. 2️⃣ findMedian() – Return the median of all numbers so far. The median is the middle element in a sorted list (or average of two middle elements if even length). ⚡ Brute Force Approach: • Store all numbers in an array and keep it sorted after every insertion. • Insertion: O(n log n) (due to sorting each time) • Median query: O(1) • Inefficient for large data streams. ->𝐎𝐩𝐭𝐢𝐦𝐚𝐥 𝐀𝐩𝐩𝐫𝐨𝐚𝐜𝐡 (𝐇𝐞𝐚𝐩𝐬 / 𝐏𝐫𝐢𝐨𝐫𝐢𝐭𝐲 𝐐𝐮𝐞𝐮𝐞𝐬): Use two heaps to maintain balance: • Max Heap (left) → stores the smaller half of numbers. • Min Heap (right) → stores the larger half of numbers. Maintain these conditions: ✔️ Size difference ≤ 1. ✔️ Max of left ≤ Min of right. • Insert: O(log n) (heap push/pop) • Median Query: O(1) (peek heap tops) 💭 Intuition – Why it Works: • Instead of sorting the entire list each time, we only care about the middle elements. • Max Heap gives easy access to the largest number of the left half. • Min Heap gives easy access to the smallest number of the right half. By balancing them, the median is always on top of the heaps. 🔗 GitHub Implementation: https://shorturl.at/kecFp #LeetCode #DSA #ProblemSolving #Heaps #Cplusplus #CodingInterview #DataStructures #Algorithms #Coding #LinkedInCoding #SystemDesign #MedianFinder #Tech

🚀 Exploring Agent Framework for Pro Developers – Observability & Workflows Over the past week and a half, I completed several labs on #AgentFramework, and the experience reinforced how critical observability and workflow orchestration are for building robust AI agents. I initially did my lab in .NET and loved the experience but in a convo with one of my colleagues here at work Eduard van Valkenburg who's leading the charge for #Python he showed me #DevUI which is really key to the work I am doing on #Observability especially with folks creating #Agents and interested in transparency of the calls being made... tools or otherwise, quality of responses etc... and #DevUI is at the moment on only in Python... so with that said... I Forked the Agent Framework repo and on my own fork I dove deep with it & created tools that can read from my #Microsoft365 #MicrosoftGraph data exposed as #Tools (Think #MCP or #Function calls) with the Agents... See a complete walk thru below. Lab 9: Shows how to integrate #OpenTelemetry (#OTEL) with GenAI Semantic Conventions for deep traceability. Lab 10: Highlights workflow concepts that resonate with enterprise-grade agent design. I’ve shared a detailed walkthrough here: 👉 https://lnkd.in/eKK3pRkx And here’s my fork demonstrating #M365 Graph tool calls and how developers can trace agent behavior end-to-end: 👉 https://lnkd.in/e-BDyXpM Curious to hear how others are approaching observability and evals in their agent workflows. Let’s connect! #AgentFramework #GenAI #Microsoft365 #AIObservability #DeveloperExperience Tagging Andrew Connell, Sébastien Levert, Maxim Salnikov, Lee Englestone and im sure im forgetting others who have had convo with me or chats on this topic..

Introducing `marimo check`: a linter that gives agents the feedback they need to write production-ready data apps and pipelines, as well as exploratory notebooks, with marimo. https://lnkd.in/gyFYz56H Companies large and small are replacing legacy data tools with AI-native, workflows built on marimo: with agents, marimo lets companies go instantly from prompt to data app, pipeline, LLM eval notebook, and more . Developers and agents prefer marimo to Jupyter because it's code-first and productionizable, while agents wielding marimo make Python accessible to non-technical users who were previously limited by Looker's inflexible UI. Previously, agents like Claude Code could already execute marimo notebooks as scripts to get feedback on the correctness of their code. With `marimo check`, agents can get additional feedback without running a single line of code, empowering them to self-correct on the fly. In the coming weeks, we will have a lot more to say about how marimo is uniquely well-suited to empower both agents and humans to rapidly create production-ready data apps and pipelines ... see our latest newsletter for a sneak peek (https://lnkd.in/geQ85iXy).

🚀 Powering RAG with LangChain: Meet the Document Loader Class! One of the most underrated yet powerful components in the LangChain ecosystem is the Document Loader — your gateway to structured data ingestion. In this example (see image 👇), I’m using PyPDFLoader to automatically: 📥 Load and parse PDFs directly from a URL 🧩 Split content into Document objects with page_content and metadata 🔍 Prepare clean, structured inputs for downstream tasks — from embeddings to retrieval for RAG (Retrieval-Augmented Generation) pipelines Why it matters: LangChain’s loader classes make it seamless to connect raw data (like PDFs, web pages, or APIs) to LLMs, enabling smarter context retrieval and more accurate responses. Whether you’re: Building AI-powered assistants trained on custom documents Creating semantic search or chat-with-your-data apps Or just exploring LLM workflows for data extraction — the Document Loader is your essential first step. 💡 Tip: Combine PyPDFLoader with text splitters and vector stores (like Chroma or FAISS) for a production-ready RAG pipeline. #LangChain #AI #LLM #RAG #MachineLearning #Python #GenerativeAI #DataScience #VectorDatabases #OpenSource #keeplearning

Excited to share my latest project: a multi-threaded chatbot application built with LangGraph and Streamlit! 🚀 This project showcases how to build stateful, tool-using AI agents with persistent memory. I've implemented two distinct versions to demonstrate different capabilities: 1️⃣ A simple conversational chatbot that remembers past interactions in a thread, backed by a SQLite database. 2️⃣ An advanced agentic chatbot that can use external tools like DuckDuckGo for web searches, a calculator for math, and an API to fetch real-time stock prices. One of the coolest features is the Streamlit UI, which visually indicates when the agent is using a tool (st.status), providing a transparent and interactive user experience. Key Technologies Used: Backend: LangGraph for building the stateful agent graph. LLM: Google's Gemini Pro via LangChain. Frontend: Streamlit for a fast, interactive UI. Persistence: langgraph.checkpoint.sqlite.SqliteSaver to save conversation history. Tools: DuckDuckGo Search, Alpha Vantage API, and a custom calculator tool. This project was a fantastic learning experience in building modular and powerful AI applications. The separation of the LangGraph backend from the Streamlit frontend makes the architecture clean and scalable. Check out the full code and a detailed breakdown on my GitHub! Your feedback and suggestions are welcome. 🔗 GitHub Repo: #AI #LangGraph #Streamlit #Python #LLM #GenerativeAI #Chatbot #AIAgent #LangChain #GoogleGemini #PortfolioProject #OpenSource

🚀 Built an AI Email Drafter using LangGraph & LangChain! Excited to share my latest project exploring agentic AI tools! What I built: Intelligent email generation system with customizable tones Multi-agent workflow using LangGraph's state management Streamlit interface with real-time preview and file management Integration with Groq's Llama-3.1-8b-instant model ⚓ Key tools & technologies: 💠 LangGraph: Agent workflow orchestration with conditional routing 💠 LangChain: LLM integration and message handling 💠 Streamlit: Interactive web UI with progress bars and real-time feedback 💠 Groq API: High-performance inference for fast email generation 💠 Python: Custom tools for document management and file operations 📕 Key learnings: Designing agent architectures with conditional routing Managing complex conversation flows and state Building resilient AI applications with proper error handling Creating intuitive UIs for AI-powered tools Implementing custom tools for document storage and retrieval 👉 Github:- https://lnkd.in/gZv5Dsa9 #AI #LangChain #LangGraph #MachineLearning #Python #Streamlit #Groq #AgenticAI #TechInnovation #Tools #srms #srmscet

Here is a short cheat sheet for the watsonx Orchestrate CLI, designed for both local development and remote configuration. This resource covers essential components, including: - Environments - Agents - Connections - Import LLM (watsonx.ai) via AI Gateway The CLI is a part of the "watsonx Orchestrate Agent Development Kit (ADK)" 😉 Explore the full cheat sheet: https://lnkd.in/eFh5WEcG #watsonxorchestratecli #connection #environment #aigateway #agent #watsonxorchestrate

🚀 A Multi-Agent Small Language Model system Built a multi-agent customer service system for an airline that runs locally on consumer hardware. It demonstrates multi-agent coordination with SLMs using the OpenAI Agents SDK. Features: ✅ Multi-Agent Architecture — Specialized agents for Triage, FAQ, and Seat Booking ✅ Seamless Handoffs — Agents route conversations based on user intent ✅ Local Deployment — Runs on consumer GPUs with vLLM ✅ GGUF Quantization — Uses quantized models for memory efficiency This demonstrates how to build multi-agent systems with SLMs using the OpenAI Agents SDK, keeping data local and costs low. 🔗 Medium: https://lnkd.in/gcfwyXFW 🔗 GitHub Repository: https://lnkd.in/gjMtb_gm #MultiAgent #AI #LocalAI #SmallLanguageModels #OpenSource #CustomerService #VLLM #Python #Research #OpenAIAgentsSDK