Most people use vector databases for chatbots and RAG pipelines. 𝗦𝗲𝗻𝗾𝗶 𝗔𝗜 𝘂𝘀𝗲𝘀 ...

𝗧𝘂𝗿𝗻𝘀 𝗼𝘂𝘁, 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗽𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝘄𝗼𝗿𝗹𝗱 𝗶𝘀 𝗰𝗼𝗺𝗽𝗹𝗲𝘁𝗲𝗹𝘆 https://t.co/x4X58ZB3zS" / X

Most people use vector databases for chatbots and RAG pipelines. 𝗦𝗲𝗻𝗾𝗶 𝗔𝗜 𝘂𝘀𝗲𝘀 𝗠𝗶𝗹𝘃𝘂𝘀 𝘁𝗼 𝗽𝗼𝘄𝗲𝗿 𝗽𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗿𝗼𝗯𝗼𝘁𝘀. 𝗧𝘂𝗿𝗻𝘀 𝗼𝘂𝘁, 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗽𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝘄𝗼𝗿𝗹𝗱 𝗶𝘀 𝗰𝗼𝗺𝗽𝗹𝗲𝘁𝗲𝗹𝘆 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗳𝗿𝗼𝗺 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮 𝗰𝗵𝗮𝘁𝗯𝗼𝘁. Senqi AI's CEO Song Zhi walked us through exactly where the difficulty lies: • 𝗧𝗮𝘀𝗸𝘀 𝗻𝗲𝘃𝗲𝗿 𝗰𝗹𝗲𝗮𝗻𝗹𝘆 𝗲𝗻𝗱. Ask a chatbot to summarize a doc — it processes once and returns a result. Ask a robot to patrol a warehouse — it's navigating, scanning, dodging obstacles, and adjusting in real time for the entire run • 𝗧𝗵𝗲 𝗲𝗻𝘃𝗶𝗿𝗼𝗻𝗺𝗲𝗻𝘁 𝗰𝗵𝗮𝗻𝗴𝗲𝘀 𝗺𝗶𝗱-𝘁𝗮𝘀𝗸. A hallway clear at 9am has pallets blocking it by noon. Each change can break a task chain that worked five minutes ago, so the system has to re-plan on the fly • 𝗡𝗮𝘁𝘂𝗿𝗮𝗹 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗶𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝗶𝗼𝗻𝘀 𝗵𝗶𝗱𝗲 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝘁𝗿𝗲𝗲𝘀. "Go check that area and flag anything unusual" sounds like one task, but the robot needs to resolve which area, what counts as unusual, and what to do when the photo is blurry or the path is blocked • 𝗠𝗶𝘀𝘁𝗮𝗸𝗲𝘀 𝗵𝗮𝘃𝗲 𝗽𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗰𝗼𝘀𝘁𝘀. In a chatbot, a bad answer costs a re-prompt. In a robot, a bad decision costs time, battery, and sometimes a human walking over to physically recover it 𝗛𝗼𝘄 𝗠𝗶𝗹𝘃𝘂𝘀 𝘀𝗼𝗹𝘃𝗲𝘀 𝘁𝗵𝗲 𝗺𝗲𝗺𝗼𝗿𝘆 𝗽𝗿𝗼𝗯𝗹𝗲𝗺: A robot that can't remember is a robot that repeats mistakes. Senqi AI built a long-term memory layer inside their robot system RoboBrain using Milvus: • 𝗦𝘁𝗼𝗿𝗲𝘀 𝗳𝗶𝗲𝗹𝗱 𝗲𝘅𝗽𝗲𝗿𝗶𝗲𝗻𝗰𝗲 𝗮𝘀 𝘃𝗲𝗰𝘁𝗼𝗿𝘀 — correction strategies, failure records, human-override decisions, post-inspection summaries, zone-level rules like "nighttime glare in this area, adjust camera angle" • 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗲𝘀 𝗯𝘆 𝗺𝗲𝗮𝗻𝗶𝗻𝗴, 𝗻𝗼𝘁 𝗸𝗲𝘆𝘄𝗼𝗿𝗱𝘀 — when a robot hits glare in Zone B, Milvus finds the most similar past incident from Zone D three weeks ago, not an exact keyword match • 𝗖𝗼𝗺𝗽𝗼𝘂𝗻𝗱𝘀 𝗼𝘃𝗲𝗿 𝘁𝗶𝗺𝗲 — every completed task, every human correction, every resolved edge case feeds back as new vectors. The robot's experience base grows with each deployment • 𝗕𝗿𝗶𝗱𝗴𝗲𝘀 𝘁𝗵𝗲 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗴𝗮𝗽 — relational DBs handle task IDs and timestamps, but "find the most relevant past failure in a similar scene" is a similarity problem. That's what Milvus is built for 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁: 𝗿𝗼𝗯𝗼𝘁𝘀 𝘁𝗵𝗮𝘁 𝗮𝗱𝗮𝗽𝘁 𝗳𝗿𝗼𝗺 𝗱𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 𝗵𝗶𝘀𝘁𝗼𝗿𝘆 𝗶𝗻𝘀𝘁𝗲𝗮𝗱 𝗼𝗳 𝘀𝘁𝗮𝗿𝘁𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝘀𝗰𝗿𝗮𝘁𝗰𝗵 𝗲𝘃𝗲𝗿𝘆 𝘁𝗶𝗺𝗲. 𝗦𝗲𝗲 𝗳𝘂𝗹𝗹 𝗯𝗿𝗲𝗮𝗸𝗱𝗼𝘄𝗻 𝗵𝗲𝗿𝗲: milvus.io/blog/how-robob