siruignaw-sys/dsh-tool-bandit-search

README excerpt

dsh-tool-bandit-search A DeepSeek Harness plugin that replaces the standard web search tool with a search tool that learns which search strategy to use through a contextual multi-armed bandit, instead of relying on a single hardcoded approach. Why Every web search has a tradeoff: a fast, narrow query gets you an answer quickly, but a broader, multi-angle query gets you better coverage at the cost of latency. Hardcoding one strategy means always overpaying for simple questions or always underdelivering on complex ones. This plugin lets the tool discover, from real usage, which strategy tends to pay off — and keeps adapting as conditions change. How it works The search tool has two internal strategies ("arms"): - quick — a single search query, capped at 5 results. Fast, good for simple factual lookups. - thorough — three query variants (the original plus two reframed angles) run in parallel and merged/deduplicated, capped at 10 results. Slower, better for open-ended or multi-perspective questions. On every call, the plugin uses Thompson sampling to pick an arm: each arm has a Beta(α, β) distribution representing its estimated reward, the plugin samples from both distributions, and wh…

Category