skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
ultimatile/cuda-x-skills80 installs

cuda-webdoc-search

Search CUDA-X library documentation (cuBLAS, cuTENSOR, cuTensorNet, cuSOLVER, etc.) to find API symbols, functions, and types. Use when you need to look up CUDA library APIs, discover available functions, or find documentation URLs for specific operations.

How do I install this agent skill?

npx skills add https://github.com/ultimatile/cuda-x-skills --skill cuda-webdoc-search
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill provides tools for searching and extracting CUDA-X library documentation. It includes capabilities to fetch content from remote URLs and local file paths. While the primary data sources are official NVIDIA documentation domains, the ability to read local files and process arbitrary HTML content presents minor security risks regarding data exposure and indirect prompt injection.

  • Socketpass

    No alerts

  • Snykwarn

    Risk: MEDIUM · 1 issue

What does this agent skill do?

CUDA-X Documentation Search Guide

Overview

Search and discover APIs across CUDA-X library documentation by querying Sphinx inventory files. Use this skill when you need to find specific CUDA library functions, check available APIs, or get documentation links.

Setup

uv tool install "cuda-webdoc-search @ git+https://github.com/ultimatile/cuda-x-skills.git#subdirectory=skills/cuda-webdoc-search"

After installation, cws is available as a standalone command. For one-shot usage without installing:

uvx --from "cuda-webdoc-search @ git+https://github.com/ultimatile/cuda-x-skills.git#subdirectory=skills/cuda-webdoc-search" cws search cublas --stats

When to Use

  • Finding CUDA library APIs (e.g., "what SVD functions does cuTensorNet have?")
  • Checking if a specific function exists in a library
  • Getting documentation URLs for CUDA APIs
  • Exploring available operations in a library

Quick Reference

Check Available Domains First

Before searching, check what domains (API types) are available:

cws search <library> --stats

This returns domain counts: cpp (C++ APIs), c (C APIs), py (Python bindings), std (doc labels).

Search for APIs

cws search <library> --domains <domain> --keywords "<terms>" --fuzzy --limit 20

Keyword Syntax (fzf subset)

  • Space-separated terms are AND (all must match): --keywords "SVD batch"
  • Use | for OR (requires shell quoting): --keywords "SVD | QR"
  • AND binds tighter than OR: --keywords "a b | c" = (a AND b) OR c
  • Note: Only AND/OR via | is supported. fzf operators ^, ', !, $ are not available.

Available Libraries

See registry.toml for the full list. Common ones:

  • cuquantum - cuTensorNet, cuStateVec (quantum/tensor network)
  • cublas - BLAS operations
  • cusolver - Dense/sparse solvers
  • cusparse - Sparse matrix operations
  • cufft - FFT operations
  • cudnn - Deep learning primitives
  • cutlass - GEMM templates
  • cuda_math - Math functions

Workflow

  1. Identify the library: Determine which CUDA library likely contains the API
  2. Check domains: Run --stats to see available domain types
  3. Search: Use --keywords with --fuzzy for flexible matching
  4. Filter by domain: Use --domains cpp for C++ APIs, --domains c for C APIs
  5. Extract details: Use cws get to get full documentation content

Examples

Find tensor decomposition APIs in cuTensorNet

cws search cuquantum --stats
# Shows: cpp:3179, py:1172, std:3378, c:4

cws search cuquantum --domains cpp --keywords "SVD" --fuzzy --limit 10
# Returns: cutensornetTensorSVD, etc. (functions ranked above enumerators)

cws search cuquantum --domains cpp --keywords "SVD | QR" --fuzzy
# Returns: entries matching SVD OR QR

Search across multiple libraries

# "Where is SVD in CUDA-X?" — search cuSOLVER and cuDSS together
cws search cusolver cudss --keywords "svd" --fuzzy --limit 10

# Mix any number of sources
cws search cusolver cusparse cudss --keywords "solve" --fuzzy

Multi-source output uses "sources" (list) instead of "source" (string), and "total_found" becomes a per-source dict.

Find GEMM functions in cuBLAS

cws search cublas --stats
# Shows: std:36 (only doc labels, no cpp/c domain)

cws search cublas --domains std --keywords "gemm"
# Returns: cublas-t-gemm, cublas-t-gemmex, etc. (doc section labels)

Extract API documentation details

After finding a function URL, extract its full documentation:

cws get <url> --section <function_name>

Example:

cws get \
  "https://docs.nvidia.com/cuda/cuquantum/latest/cutensornet/api/functions.html" \
  --section "cutensornetTensorSVD"

Output is a brace-delimited text tree:

{
cutensornetTensorSVD
{
cutensornetStatus_t cutensornetTensorSVD {
const cutensornetHandle_t handle
const cutensornetTensorDescriptor_t descTensorIn
...
}
Performs SVD decomposition of a tensor. { ... }
Parameters { ... }
}
}

List available sources

cws list                     # table format
cws list --json              # JSON format

Audit registry health

cws audit                    # audit all sources
cws audit --source cublas    # audit single source

Output Format

cws search

JSON output includes:

Single source:

  • source: Library name (string)
  • total_found: Total APIs in filtered domains
  • filtered_count: APIs matching keywords
  • candidates: List of matching APIs with group (name), url, domain, role

Multi-source (cws search A B):

  • sources: Library names (list)
  • total_found: Per-source totals (dict)
  • filtered_count: Total APIs matching keywords across all sources
  • candidates: Merged list (k-way merge by score if --fuzzy, else concatenated)

cws get

Brace-delimited text tree where:

  • { } denote hierarchy
  • Text content is preserved
  • Inline elements (code, links) are flattened to text

Options:

  • --section <id>: Extract only a specific section
  • --main-only: Extract only the main content area

Files

  • cli.py - Unified CLI entry point (cws)
  • search.py - Search APIs across library inventories
  • audit.py - Audit registry entries for endpoint health
  • get.py - Extract documentation content as text tree
  • fetchers.py - Data fetching for Sphinx/Doxygen sources
  • scoring.py - Search ranking and filtering logic
  • registry.py - Registry loading utility
  • registry.toml - Library metadata (URLs, doc types)

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/ultimatile/cuda-x-skills/cuda-webdoc-search">View cuda-webdoc-search on skillZs</a>