What problem does it solve? Finding relevant research papers across a 14k-paper AI interpretability corpus and a 157M-node cross-disciplinary citation graph requires careful query construction; this Skill turns a research question into a structured, multi-ranker cloud search that returns a usable paper list. ## Core Features & Use Cases - Decomposed Query Mode: Builds a structured JSON query with at most one sub-query per database, including closed-enum interpretability tags, keywords, temporal filters, and an Agent-written HyDE abstract. - Flat Query Fallback: Submits a free-form English query when task context is too thin, letting the cloud splitter handle decomposition and HyDE generation. - Iterative Refinement: Supports up to 3 refinement rounds and parallel calls when initial results are unsatisfactory. - Use Case: While investigating how ROME edits factual associations in GPT models, submit a decomposed query targeting the interp_db with causal_attribution technique tags and receive a ranked paper list written to a local JSON file. ## Quick Start Ask the agent to search the mechanic database for papers on sparse autoencoder feature splitting in large language models.
