Hasnat Ahmad

Hasnat AhmadOct 21, 2026

Most people still think RAG just means “search your docs, then ask the model.” That used to be enough. In 2026 it’s a whole toolbox. Here are the 6 RAG patterns I’d learn first: 1) Naive RAG: retrieve chunks, send them to the LLM, get an answer. Simple and fast but retrieval quality can let you down. 2) Retrieve and Rerank: adds a reranking step so the model only sees the most relevant context. 3) Multimodal RAG: pulls from text, images, audio and video not just documents. 4) GraphRAG: uses a knowledge graph to understand how things connect, which helps with complex questions. 5) Agentic RAG (Router): an agent decides which source or tool to use for each question. 6) Multi-Agent RAG: several specialist agents handle retrieval, web search and the final answer together. You don’t need all six on day one. Start with the basics add reranking when answers feel off and move to agents when one question needs more than one source. Which one are you working with right now? Tell me in the comments. ------------------------------------------------------------------------------------- 𝗜 𝘁𝗲𝗮𝗰𝗵 𝗔𝗜 𝘀𝘁𝗲𝗽 𝗯𝘆 𝘀𝘁𝗲𝗽 𝘄𝗶𝘁𝗵 𝗿𝗲𝗮𝗹 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀 𝗮𝗻𝗱 𝗻𝗼 𝗰𝗼𝗻𝗳𝘂𝘀𝗶𝗼𝗻. 𝗪𝗮𝗻𝘁 𝟭:𝟭 𝗰𝗼𝗮𝗰𝗵𝗶𝗻𝗴? 𝗗𝗠 𝗺𝗲.

Hasnat AhmadOct 17, 2026

I just started a new open source project called agent-failure-lab. Agents rarely break in dramatic ways once they leave the notebook. They break quietly. A retried tool call charges a customer twice. A loop calls the same tool until the token budget is gone. A line of text hidden in a tool result changes what the agent does next. I kept running into these while building AI agents for clients and each time the fix was a small, boring pattern that someone had to think of first. So I’m cataloguing them, with each failure reproduced in a deterministic scenario that runs without an API key. You run the broken version, watch it fail, then run the fixed one. Every fix ships as a small guard you can drop into your own agent loop. The first release is planned to cover 12 failure modes. It’s early, so the repo is still taking shape. Link is in the comment.