Discovery & validation

Three meta functions let humans and agents discover the API and check calls without reading server source.

cugraph_list_algorithms

Inventory of every registered function.

SELECT * FROM cugraph_list_algorithms();

Columns: function_name, algorithm, source_vertex_policy, weight_col_policy, output_schema_kind, signature, allowed_arg_counts_json, summary, metadata_schema_version.

cugraph_describe_algorithm

Machine-readable contract for one function. The optional second argument 'verbose' adds full option, schema, argument, and limitation metadata.

SELECT * FROM cugraph_describe_algorithm('cugraph_bfs');
SELECT * FROM cugraph_describe_algorithm('cugraph_bfs', 'verbose');

The per-function pages in this section are generated from this verbose output.

cugraph_validate_call

Static dry-run of a specific call. Returns the concrete, table-specific output schema and a structured verdict — without scanning edges, building a graph, or launching CUDA.

SELECT * FROM cugraph_validate_call(
  'cugraph_pagerank',
  'your_edges_table',
  '{"src_col":"src","dst_col":"dst","weight_col":"weight"}'
);

Output columns (in order):

Column	Meaning
valid	whether the call is statically valid
error_code	structured code when invalid (for example, table_not_found)
message	human-readable explanation
function_name	the function checked
algorithm_enabled	whether the algorithm is enabled on this server
table_resolved	whether the edge relation resolved
src_col_resolved	whether src_col exists
dst_col_resolved	whether dst_col exists
weight_col_resolved	whether weight_col exists (when given)
src_dtype	resolved source-column dtype
dst_dtype	resolved destination-column dtype
weight_dtype	resolved weight-column dtype
output_schema_json	the concrete output schema for this call
normalized_options_json	options after defaulting and normalization
would_execute_gpu	whether execution would dispatch to the GPU
metadata_schema_version	metadata contract version

What validation does and does not do

Does: confirm the function exists and is enabled; resolve the relation; check that src/dst/weight/predicate columns exist; check vertex dtype compatibility; validate JSON option names, types, and values; enforce conditional requirements (for example, BFS path mode needs a target relation); and compute the concrete output schema.

Does not: scan edge rows, materialize Parquet, build the graph, launch CUDA, or prove the source vertex or target row actually exists.