Graph tool reference

Returns: Returns load status, graph_version and graph counts. Use get_graph_info for the structural summary.

• Replacement and append change the active graph. Constructed property graphs do not support append; UNSUPPORTED_PROPERTY_GRAPH_APPEND preserves the existing graph.

Returns: Direct loads return format, status and graph version. Background loads return async, job_id and status; get_job_status exposes the completed result under job.result.

• For .sql, .csv and .txt, MCP returns CONSTRUCTION_REQUIRED and directs construction through the dataset panel. Other engine file paths can detect edge lists, triples or prose; that does not replace the dataset-construction entry point.
• Prose, multi-graph datasets and some Nebula imports become background work; queued does not mean the graph is already loaded.

Returns: Returns async=true, job_id and queued status. The final job result contains extraction and graph-loading information.

• This entry point performs text-relation extraction. Skill-based dataset construction is handled by a separate construction agent with query validation and representation evolution.

Returns: Returns load status, graph version and counts. Relations and supplied evidence metadata are included in graph creation.

• relation is stored as edge_type, evidence as edge_text, and confidence as edge_feat. Triple count need not equal final edge-entry count.

No business parameters; identity and session are injected by the runtime.

Returns: Returns success, graph_version and the cleared graph state.

No business parameters; identity and session are injected by the runtime.

Returns: Returns conversion status, graph version, node and edge counts, and direction information.

• Changing direction affects path, in/out-degree and edge-count semantics; inspect get_graph_info after conversion.

No business parameters; identity and session are injected by the runtime.

Returns: Returns datasets with identifiers, versions and storage information. Construction datasets are filtered by access scope.

Returns: Returns dataset_id, dataset_version, graph version and counts. Use load_constructed_graph for construction datasets.

• Use load_constructed_graph for construction datasets. Ordinary loading can return USE_CONSTRUCTED_GRAPH_LOADER to direct you to the correct session-binding update.

Returns: Returns dataset metadata, including identifier, name, versions and user or global scope.

No business parameters; identity and session are injected by the runtime.

Returns: Returns restoration status and the active graph_version, or an explanation when no prior state is available.

No business parameters; identity and session are injected by the runtime.

Returns: Each entry includes graph_version, description, creation time, node and edge counts, and a current flag.

Returns: Returns switch status, the active graph_version, and node and edge counts.

Returns: Returns discarded_graph_version, the active graph_version and graph counts. Fails when the version or a usable fallback does not exist.

• Marked internal in the catalog, but still discoverable through the native MCP toolkit. It deletes history; use load_graph_version for ordinary switching.

Returns: Submission returns async=true, job_id, graph_id and graph_version. Completed output includes quality summaries, counts, representation=partitioned_snapshot and current_graph_replaced=false.

• No full GraphData or NetworkX graph is constructed first; a running large-graph worker and sufficient resources are required. Omitting quality_policy uses exact node spelling and last duplicate weight for edgelist, but nfkc_casefold and maximum weight for triples. An empty policy object uses the policy-class defaults and is not equivalent to omission.

Returns: Returns an asynchronous job handle. Read summary, preview and artifact_id with get_job_result after completion. Capabilities and output identify exact, approximate, bounded or mixed semantics; full materializable output is stored as artifacts.

• Paths require source/target, flow/cuts source/sink, method families method, and metric families metric. LCA additionally needs node1/node2, dominators source, and equitable coloring num_colors. The parameter object allows extension fields, but runtime validates them per algorithm and explicitly rejects unsupported modes or insufficient resources.

Returns: Returns success, graph_id and versions, including preparation/publication status and quality summaries.

• Algorithms require a published version. This list differs from ordinary active-graph history returned by list_graph_versions; matching numbers alone do not establish a shared source.

Returns: Returns success, artifact and download_url_template. artifact includes the file count, bounded path/download-index preview and full index range.

• The {partition} in download_url_template is the download index in a flat output-file list, not necessarily the input partition number. Download each file through Backend at /api/sessions/{session_id}/large-graph/artifacts/{artifact_id}/partitions/{partition}; workspace, session, and file-scope checks remain, without browser sign-in.

No business parameters; identity and session are injected by the runtime.

Returns: Returns success and capabilities, including engine version, exact/approximate/bounded semantics, complexity, parallel strategy and result types.

• The catalog describes implementation contracts and prerequisites, not successful execution of every method on the current data. Community names currently map to weighted synchronous label propagation, and some centralities use sampling or approximations. For strict comparisons, set strict_semantics and inspect actual output.

No business parameters; identity and session are injected by the runtime.

Returns: Returns a graph summary, node and edge previews, num_edge_entries, num_unique_edges, semantic_type, storage_type, and label, feature and GraphML-readiness information.

• num_edge_entries counts stored entries, while num_unique_edges counts deduplicated logical edges. Interpret undirected bidirectional storage according to its graph semantics.

Returns: Returns degree, in-degree, out-degree, neighbors, predecessors, incident edges, labels and types, plus text when present. Lists are bounded previews.

• Use search_node for descriptive queries, then call this tool with an exact name.

Returns: Returns existence, weights, relationship names, labels and types, with preview and truncation information for multiple relationships.

• Distinguish parallel relationships using returned relationship IDs; an endpoint pair does not imply a single relationship.

Returns: Returns status fields including exists and directed.

Returns: Returns neighbors for the requested direction and traversal metadata, bounded by limit.

• Nebula handles direction in the native query; the result covers only one hop.

Returns: Returns best, candidates, ambiguous, semantic_enabled and retrieval_mode. Candidates contain exact node_name values and matching evidence.

• Resolve candidates when ambiguous=true. Without an embedding provider, exact names and valid node IDs can still match; other text queries receive no lexical-similarity score, so structural filters and sorting primarily determine candidates. Inspect semantic_enabled and retrieval_mode.
• Entity semantic search depends on Graph Engine’s own embedding configuration. EMBEDDING_PROVIDER explicitly selects sentence-transformers, openai or lexical. Automatic selection first tries local Sentence Transformers, requiring its dependency and model weights, then OpenAI when OPENAI_API_KEY is set, and otherwise lexical hashing. The current OpenAI branch calls the fixed api.openai.com/v1/embeddings endpoint; Agent Runtime chat-model or EMBEDDING_BASE_URL settings do not override it. Lexical mode does not provide semantic ranking for entities. Index-encoding failures also disable semantic search; inspect semantic_enabled and retrieval errors. Explicit openai selection requires OPENAI_API_KEY; failure to initialize a forced provider does not run the automatic selection sequence. Initial entity indexing encodes node and edge documents for the entire graph, so remote reachability, latency and quota also affect availability.

Returns: Returns best, candidates, ambiguous and retrieval mode; candidates include exact endpoints and relationship information.

• Provide at least one of query, src_query or dst_query. Resolve relationship or endpoint candidates when ambiguous=true.
• Exact endpoints resolve only the last relationship, not every parallel relationship. Use actual edge_id values in the saved attributed graph to inspect individual relationships.
• Semantic queries use the same Graph Engine embedding provider as search_node. Without an available provider, rely on exact resolution, structural filters and sorting; rerank does not invoke a separate reranking model.

Returns: A single-path result contains found, path, hop_count, total_length, edges and weighted. K-path mode returns paths; all-target mode returns reachable_count, distance statistics and nearest_paths.

• Weights represent distance or cost here. For multigraphs, a private calculation view uses the minimum weight per endpoint pair without deleting stored relationships.

Returns: Returns length, paths, count_returned and truncated, showing up to 10 paths by default. paths is empty when no path exists.

• This tool returns equal-cost optimal paths; yen_k_shortest can include more expensive alternatives. The public schema has no limit parameter and defaults to 10 results.

Returns: Depending on the mode, returns has_path, paths/count/truncated, or an average-distance result. Bounded Nebula analysis includes scope metadata.

• Average shortest paths use weights. Directed input must be strongly connected or the call fails; disconnected undirected input uses its largest component. The result does not establish reachability across other components.

Returns: Returns node or closure-edge previews, counts and truncation flags. Incomplete closure enumeration reports num_edges_at_least.

• Nebula queries are bounded by max_steps, max_nodes and max_edges. Absence within that scope does not prove unreachability across the entire graph.

Returns: Returns max_flow_value, total_nonzero_edges, top_flow_edges and truncated.

• weight acts as capacity, with parallel capacities summed. Cost weights do not automatically represent throughput capacity.

Returns: Returns min_cut_value, partition sizes, and previews of the smaller partition and cut edges.

• Source–sink cuts use capacity semantics; parallel relationships are summed by capacity in the calculation view.

Returns: Returns min_cut_value, algorithm, partition sizes, cut-edge previews and disconnected. Disconnected graphs have a cut value of 0.

• Directed input is rejected when treat_as_undirected=false. Use minimum_cut for a source–sink problem that retains direction.

No business parameters; identity and session are injected by the runtime.

Returns: Connected input returns total_weight, num_edges and mst_edges_preview. Disconnected input returns num_components and forest_summary.

• Uses an undirected view, selecting minimum-weight relationships per pair for property multigraphs. A forest summary is not a complete edge list.

Returns: Returns connectivity flags, component counts, sizes and node previews. Defaults to 5 components with 10 preview nodes each.

Returns: Returns cycles, truncated and total_cycles_at_least. With truncation, the count is a confirmed lower bound.

• Simple cycles can grow exponentially in number. limit=all removes the preview enumeration bound; when truncated=true, returned count is not the total cycle count.

Returns: Returns cycle, cycle_length, cycle_weight, edges, cycles_examined, search_exhausted and truncated.

• Longest simple cycle is a costly combinatorial problem. search_exhausted=false guarantees only the best examined candidate; multigraphs are unsupported.

Returns: Returns a topological-order preview, total node count and truncated. Cyclic input cannot produce a valid topological order.

• Topological order may be non-unique; a preview is a prefix of one valid order, not a complete execution plan.

Returns: find_cliques returns clique previews and a count lower bound; clique_number returns the maximum size; number_of_cliques returns participation counts.

• A maximal clique cannot be enlarged; a maximum clique has the greatest size among all cliques. Enumeration can be exponential, and multigraphs are unsupported.

Returns: Returns method-specific value, is_planar, hash, score_sequence, order/orders or traversal-tree edges_preview, with applicable totals and truncation. Does not change the active graph.

• Tree/forest, AT-free, bipartite and planarity checks use an undirected view. BFS/DFS require source; topological ordering requires a DAG. score_sequence checks tournament-score-sequence conditions, not whether the input itself is a tournament. Equal WL hashes do not prove isomorphism.
• The ordinary-graph entry rejects multigraphs with UNSUPPORTED_MULTIGRAPH_OPERATION; the corresponding large-graph operation has separate capability constraints.

Returns: Returns method-specific value, nodes, items, components or condensation nodes/edges; size/count and truncated describe totals and list scope.

• Provide source and target together for local connectivity/cuts. Connectivity, cuts, articulation, bridge and biconnected methods use an undirected view; condensation requires a directed graph. Global node/edge cuts require nonempty connected input and differ from capacity-weighted minimum_cut.
• The ordinary-graph entry rejects multigraphs with UNSUPPORTED_MULTIGRAPH_OPERATION; the corresponding large-graph operation has separate capability constraints.

Returns: Returns method-specific edges/total_weight, ancestor, items, path, value, antichains or structures. Collider results include count_is_exact; count=null means no exact total was obtained.

• Spanning trees use a connected undirected view. LCA needs node1/node2; dominators need source. DAG longest paths use weight, and DAG-specific methods require a DAG. is_aperiodic requires a nonempty strongly connected directed graph; v_structures excludes connected predecessor pairs.
• The ordinary-graph entry rejects multigraphs with UNSUPPORTED_MULTIGRAPH_OPERATION; the corresponding large-graph operation has separate capability constraints.

Returns: Returns matching or edges with size/total_weight, value for matching validation, or coloring and num_colors. Lists can be truncated.

• A maximal matching cannot accept another edge but need not have maximum cardinality. Minimum edge cover rejects isolates; equitable_color requires num_colors above the maximum degree. Greedy coloring depends on strategy and does not guarantee a minimum color count.
• The ordinary-graph entry rejects multigraphs with UNSUPPORTED_MULTIGRAPH_OPERATION; the corresponding large-graph operation has separate capability constraints.

Returns: Scalars return value, node sets return nodes/count, and mappings return values/count with applicable truncated. Acyclic girth returns value=null and acyclic=true.

• Center, periphery, eccentricity, Wiener index and barycenter use the largest undirected component; barycenter is weighted and the other four are unweighted. Efficiency uses an unweighted undirected view; rich-club coefficients use normalized=false. flow_hierarchy requires a directed graph and uses weights.
• The ordinary-graph entry rejects multigraphs with UNSUPPORTED_MULTIGRAPH_OPERATION; the corresponding large-graph operation has separate capability constraints.

Returns: Returns centrality_type, total_nodes, value_stats, top_nodes, bottom_nodes and truncated.

• On multigraphs, PageRank/eigenvector/Katz sum parallel weights; betweenness/closeness/harmonic/load use minimum distances. voterank/percolation do not support multigraphs.
• Distance weights on property/multigraphs must be nonnegative; betweenness and load require strictly positive weights. Iterative centralities can fail on numerical or convergence conditions.

Returns: Returns degree statistics, distributions and high/low-degree node previews. With node specified on Nebula, returns that node's degree directly.

Returns: Returns metric, value and defined. Diameter/radius use unweighted distances in the largest undirected component. Nonfinite values become value=null and defined=false; undefined assortativity also includes a reason code and explanation.

• density retains the original direction semantics; diameter/radius use the largest undirected component. Multigraph support is explicit only for density, diameter and radius.

Returns: Returns the algorithm, community count, modularity, sizes and node previews, with communities_truncated.

• Leiden falls back to Louvain if dependencies are unavailable or execution fails. Girvan–Newman uses the first split; asyn_fluidc processes only the largest component of disconnected input.
• Multigraphs support only louvain and greedy_modularity using an undirected sum-of-parallel-weights view. Nodes within each community remain a preview.

Returns: Returns subgraph counts and node/edge previews. create_version=true also returns a new graph version and activates it.

• create_version defaults to true. Pass false for inspection only; after activation, subsequent tools use the subgraph.
• Multigraphs explicitly support ego_graph and node_subset; core, truss and corona modes are not supported on multigraphs.

Returns: Returns a score for the specified pair or the top_k potential relationships ranked by score.

• Uses undirected neighborhoods. Global scoring still scans non-edges; top_k bounds retained results, not candidate computation. Multigraphs are unsupported.

Returns: Returns indexing information such as document ID, chunk count and retrieval mode.

• The document index belongs to the current session. Changing embedding dimensionality may require reindexing existing documents.

Returns: Returns relevant passages, document and chunk IDs, scores and retrieval mode.

Returns: Returns nodes, edges and optional community/highlight data. Bounded views include fields such as truncated; the active graph is unchanged.

• The overview combines highlights, node types, community representatives, high-degree nodes and connecting paths. VISUALIZE_MAX_EDGES defaults to 2,000. Inspect truncated, total_nodes/total_edges and overview.required_nodes_truncated; excessive highlights can also be truncated. Obtain graph metrics from analysis tools.

Returns: Returns report structure and display data. Statistics must come from source_results, the graph, or data explicitly supplied by the user.

• Controlled by MCP_ENABLE_REPORTING_TOOLS and enabled by default; 0, false, no or off disables agent registration.
• Returns an analysis_report.v2 structure for display; it does not automatically write Markdown/PDF or Derived files. For a download, create content under the Runtime Ephemeral workspace and use workspace_publish(path=..., name=...), with the correct extension in name.

Returns: Returns dataset identity, profile, graph count, labels and split information. MCP loading runs as a background job.

• Uses fixed dataset contracts, splits and metrics. Loading one constituent graph as the active graph does not replace the multi-graph training dataset.

Returns: Returns graph_version, graph counts, split, graph_label, previews and analysis readiness.

Returns: Returns an architecture diagram, layer summary and key hyperparameters.

Returns: MCP returns a job handle. The completed result contains model identity, training status and metrics or errors; artifacts are associated with their session and training-data identity.

• Load a dataset before graph classification and use its fixed split/metric. feature_source=embedding requires upstream_model_id; inspect readiness before training.
• Dotted parameter names show nested paths; send actual requests as data, train_config and algo_config objects. MCP supplies the internal background-job marker for training.
• The MCP layer can return a background handle; GraphMCPTools in the native runtime monitors the job and delivers its final result to the main agent.
• Single graphs without node_feat use normalized in-degree, out-degree, total-degree and constant features; business fields are not automatically encoded. GNN pretraining has no separate masking/augmentation implementation or early stopping. Node2Vec runs on CPU, defaults to 128 embedding dimensions and ignores weight_decay and early stopping.

Returns: Supervised tasks return predictions and applicable evaluation; pagination does not narrow the evaluated set. Pretraining returns an embeddings array of node and embedding entries. PCBA displays only a few high-probability tasks by default.

• edges must be a list of two-element arrays. Prediction pagination does not mean evaluation covers only that page; PCBA returns a few high-probability tasks by default.
• Pretraining inference queries embedding.npy saved during training rather than running a GNN on a new graph. Link prediction uses a symmetric inner-product decoder; candidates can include existing edges, so inspect exists_in_graph. Node-classification query.split reads original graph masks, not unsaved random training splits.

Returns: Returns success, job and progress_updates. Ordinary jobs contain status, progress, detail and final result; large-graph jobs contain persisted stage, summary and artifact_id fields. Use get_job_result for the large-graph result summary.

• For ordinary jobs, after_seq replays progress by cursor; omission drains pending updates. Large-graph progress_updates is empty. Read ordinary completed output from job.result or get_job_result, and large-graph output with get_job_result. Check terminal status and business success in either case.

Returns: Returns success and job_id indicating acceptance of the cancellation request. Confirm a terminal state with get_job_status; request acceptance alone does not prove computation has exited.

• Cancellation intent and execution exit are confirmed separately. Queued jobs do not start computation; running jobs require worker exit and resource cleanup.

No business parameters; identity and session are injected by the runtime.

Returns: Returns Nebula availability, connection and graph-storage status.

No business parameters; identity and session are injected by the runtime.

Returns: Synchronization can return a background job_id. Read the final synchronization result and graph-version information through job status.

• Requires an available Nebula service. Agents normally query nodes and paths through the unified get_node_info, get_neighbors and shortest_path tools.

Returns: Returns Nebula node, attribute and connection information, or a storage error when the graph is unavailable or unsynchronized.

• This is a backend-specific entry point. Use get_node_info to let the engine select a path for the active storage backend.

Returns: Returns Nebula neighbors and traversal information, with the result count bounded by limit.

• The unified get_neighbors can also route to Nebula primary storage. Calling both is unnecessary when they answer the same query.

Returns: Returns bounded path-query results, node sequences and traversal information. No path within the bound does not establish global unreachability.

• Has no use_weight parameter and does not minimize weighted cost. Use shortest_path for the unified interface; native Nebula traversal corresponds to use_weight=false.

Returns: Returns outer success and result. Ordinary terminal result contains job_id, status, result and error; large-graph result includes status, graph_id, graph_version, summary, preview, artifact_id and error.

• success=true for queued/running means successful lookup only. Confirm status=done, then inspect the inner business result; failed/cancelled returns success=false. Unknown or inaccessible jobs return no result.

No business parameters; identity and session are injected by the runtime.

Returns: Returns source-type capabilities and optimization rules. Unregistered algorithms report algorithm_unavailable.

• Initial source construction is started from the dataset panel. Report unavailable algorithms as such; a text-extraction result is not a skill-construction result.

Returns: Returns status, revision, source_type and counts. Only ready indicates an available saved graph.

Returns: graph contains directed, nodes(id,type,properties) and edges(id,source,target,type,properties), alongside revision information.

• Read the graph before referring to element IDs. revision belongs to the saved dataset and differs from the session graph_version namespace.

Returns: Returns load status, session graph version and dataset binding. Replaces the active graph without rerunning construction.

• After loading, inspect analysis semantics and features with get_graph_info, and display the result with visualize_subgraph.

Returns: Returns a preview graph and diff for reviewing affected nodes, relationships and properties. Saved state and revision are unchanged.

• operations[].op is required in every item; other fields depend on the operation. Use get_constructed_graph first to obtain real node, relationship and property names.

Returns: Returns commit status and a new revision. On a revision conflict, read and preview again; then use load_constructed_graph to update the active session.

• Keep base_revision and operations identical to the preview and obtain user confirmation. Set remove_from_node and cascade explicitly to express the intended behavior.