Searchbox
An airgapped, closed-corpus question-answering demo where a self-hosted Qwen3.6 agent explores a zipped document collection with local tools, such as bash, grep, embeddings and rerankers, under a fixed token budget, to test how test-time compute affects search.
Given a query, a zip archive of documents called a dataroom, and a token budget, Searchbox has a local Qwen3.6-35B-A3B model, run in a minimal harness, explore the archive with local tools such as bash, grep, embeddings and rerankers to compose an answer, treating search itself as test-time compute rather than a single fixed retrieval step. The project frames itself as a testbed for questions such as which tools a model reaches for during agentic search, whether a dense retriever adds anything a keyword search like grep does not, and whether spending more token budget produces better answers, especially on harder questions.
The archive is unzipped read-only and indexed lazily on first use, with embeddings cached per text and role for the life of a run so identical text is never embedded twice; the task is appended to the model's system prompt with an explicit instruction to answer only from the archive, using no network access, and to write the final answer to a file. Two companion projects feed Searchbox: one that crawls the web agentically to produce a dataroom zip, and one that extracts entity relations from a dataroom to generate non-trivial test questions.
It runs as a self-hosted demo built around local embeddings, rerankers and the Qwen3.6 model family.