DingWu1021/Promsa — reverse-engineered prompt

Build me a working local version of this visual question answering research agent. I want to give it an image and a question, and have it act like a budget aware researcher that can decide step by step whether to do an image search, do a text search, or stop and answer once it has enough evidence.

Please wire up the full flow so it can search against a Wikipedia style knowledge source, summarize what it finds back into the conversation, avoid repeating the same retrieval results, and handle cases where the first search was wrong by trying again with better evidence. I also want the training and evaluation flow to work with the included configs, scripts, and sample data, including the cold start stage and the reinforcement learning stage described in the README.

Make it easy to run locally, with clear commands to start the needed services, test inference on an example image and question, and run eval. If anything is unclear, look up current docs online and fill in the missing setup details.