dsh-tool-bandit-search

In web searches, users often face a dilemma: quick and simple queries yield fast answers but may lack comprehensive coverage, while broader, multi-angle queries provide better results at the cost of latency. Hard-coding a single strategy fails to adapt to all scenarios, leading to over-processing of simple questions or under-delivery on complex ones. The dsh-tool-bandit-search plugin addresses this by employing a contextual multi-armed bandit mechanism to dynamically optimize search strategy selection. It features two internal strategies—'quick' for single queries capped at 5 results, and 'thorough' for parallel multi-query variants merged to up to 10 results—leveraging Thompson sampling to balance exploration and exploitation. After each call, a reward is computed based on result quality (relative to the strategy's own cap) and speed (response latency), updating the bandit's beliefs in real time. This allows the search tool to learn and adapt to different query types without manual tuning, naturally favoring the strategy that performs best under current conditions. Ideal for frequent DeepSeek Harness users such as researchers, developers, or content creators, the plugin optimizes the trade-off between result quality and response time, integrating seamlessly into workflows to enhance efficiency and let models focus on core tasks rather than strategy decisions.