What problem does it solve? When developing Ascend C/CANN operators for Ascend NPU hardware, engineers need verified evidence about API signatures, Tiling constraints, data movement rules, debugging fixes, and similar operator examples before writing or changing code. This Skill provides a read-only retrieval entry point over a structured knowledge base, turning a task description into short queries, ranked evidence cards, and graph neighbors so answers are grounded in documented sources instead of guesses. ## Core Features & Use Cases - Preflight retrieval planning: Converts a task, error, or question into a search route, read-first card list, relevance signals, and follow-up commands before any reasoning or coding begins. - Composable retrieval pipeline: Supports BM25/TF-IDF/tag-type/graph/dense recall, deterministic reranking, regex grep, multi-hop graph neighbors, and scoped filtering by bundle, kind, tags, or A3 platform. - Evidence-ordered API verification: Enforces a knowledge-base-first, fixed-revision-upstream-second, local-CANN-install-last evidence order for API and platform constraint checks. - Use Case: While implementing a BatchMatmulMaxSum operator kernel, run preflight with the task description to retrieve DataCopyPad usage cards, Tiling partitioning guides, similar operator examples, and known precision traps before writing the kernel. ## Quick Start Ask the assistant to run the knowledge-query preflight for your current Ascend C operator task, for example by requesting a knowledge base check on DataCopyPad alignment constraints before implementing the kernel.
