Workspace guide

#Choose the right entry point

Opening the application enters the default workspace directly, without sign-in, registration, or account switching. Create your graph prepares construction data; Projects organize conversations around datasets; the Graph workspace shows the current conversation’s results. These are application features; the documentation site does not accept data or call models.

The home page includes Demi-Gods and Semi-Devils, financial risk, and molecular property examples. Use example graph reuses the preset graph and opens a conversation. Selecting a question topic on a home-page case sends that question and starts analysis immediately, without repeating extraction. The card indicates whether preset resources are available.

Configure the conversation model in Model services. URL and API Key fields are not filled automatically: blank fields retain the saved connection, and saved new values replace it. The construction panel has separate Construction model settings to select or customize a model and save its connection for TXT extraction and AI-assisted table construction.

Changing the UI language immediately updates example names, questions, and system labels. Responses follow the UI language at send time; a language change affects the next turn without translating or rewriting existing messages.

#Review the source before building

After Upload source, choose Build graph on the source card to open the panel, review the samples, and then start construction. Source data preview shows a bounded sample of the original input, separately from graph nodes, edges, and schema. It remains available before, during, and after construction.

• SQL: switch tables to inspect columns, types, keys, and up to five sample rows. Explanations prefer source comments and then key relationships; name-based inference is explicitly marked for review. Preview does not execute an uploaded SQL script.
• CSV: selecting one or more CSV files opens the table-structure dialog, which currently reads comma-separated headers. Check names, keys, composite-key column order, and each table’s samples. A direct single-CSV source created through another upload entry without bundle schema can select a delimiter on the construction page; this does not repair a header already configured incorrectly.
• TXT: review an excerpt preserving line breaks, up to 4000 characters. Line-based records preserve text structure; intelligent entity and relation extraction requires a configured construction model and manual review of extracted meaning.

#Build, edit, and reuse a graph

• Set the domain and construction options, then start construction. Status refreshes automatically; on failure, read the reason and correct the input or configuration before retrying.
• After construction, inspect the draft’s size and type relationships, then click Save construction result. The graph becomes usable for analysis when its state is ready. The schema overview shows node and relation types, not the full instance graph. Full construction reports are stored on the server or in CLI output.
• To edit the graph, choose an operation and Apply and preview, inspect the changes, and then save. Preview does not commit automatically; reload the current revision after a conflict.
• Load a constructed graph imports an independent copy into a project. It has its own dataset_id and subsequent revisions, and does not automatically follow edits to the personal source.
• Create a conversation in the target project and select the imported dataset. Confirm the active graph after loading; construction_revision and session graph_version are separate version numbers.

#Inspect entities and trace paths

For example, ask: “Confirm graph direction, node types, relation types, and graph_version, then compute a minimum-hop path between the two real node IDs I provide, explicitly using use_weight=false, and display the path subgraph.” Replace IDs with values copied from data or node details. Paths follow graph direction; more edges do not imply higher revenue or sales.

Partitioned large graphs use a separate snapshot and job workflow: check get_large_graph_capabilities, prepare a partitioned version with prepare_large_graph, submit computation with run_large_graph_algorithm, and read the result artifacts after completion. Ask the agent to follow these steps in chat. Large-graph algorithms may use sampling or approximation; interpret the returned scope and method.

#Training, inference, and project comparison

Submit the task, model, and parameters in chat. Node classification needs suitable features, labels, and splits. An unlabeled business graph can support applicable pretraining tasks but does not by itself provide classification evaluation. Multi-graph classification uses a separate data format and loading workflow.

• Use Machine learning to view status, loss, evaluation curves, or tables. A submitted or running job is not a completed training result.
• In a project conversation, switch between Current conversation and Project comparison to inspect the task, model name, epochs, metrics, and completion state, and confirm the dataset. Check full parameters and model_id in the chat tool result, or ask the agent to retrieve them. Metrics from different tasks or splits cannot be ranked directly.
• Use the returned model_id for inference after training completes. Displayed row count is not evaluation sample count. A dash in the training list is not zero; inference metrics are separate from metrics recorded during training.

#View reports and download files

Artifacts shows structured reports and conversation files. Report cards display analysis content; the file list includes uploads and provides download buttons for downloadable results in Derived. Report text in chat does not itself mean a downloadable file has been created.

• Ask the report to record the source, dataset_id, construction_revision, graph_version, tools, and metric definitions. The structured reporting tool is enabled by default and unavailable when the server sets MCP_ENABLE_REPORTING_TOOLS=false.
• For a download, ask for a result file such as Markdown and publication to Derived. Open Conversation files under Artifacts, select Refresh files if needed, and use the file’s download button.
• Verify that report numbers come from completed tool results and distinguish source facts, structural metrics, and model inferences.

#Manage memory, skills, and help

Memory in the sidebar lets you search, inspect, and delete user memories. Explicitly ask to save a preference as user memory if it should carry into future conversations, then reopen the memory panel to verify it. Learned knowledge uses separate storage, so save_learning results may not appear in this list.

Skills lists discovered procedures and their details, accepts .md or .zip uploads, and supports reviewing candidates. Uploading or generating a skill does not approve it automatically. Deletion still depends on its source and editability; built-in and workspace-created skills follow different rules. Help searches graph tools and graph-learning models by name, task, or purpose.

A reusable preference could be: “Check source columns and keys first; report the data source and graph versions; connection counts reflect all relationships in the current graph, not sales volume or revenue.” Ask about saved preferences in a new conversation to check retrieval.

#Confirm completion and recover

• If messages pause, inspect status or reopen the conversation. Background work may still be running; do not repeatedly submit the same write.
• Cancellation depends on server capabilities. When supported, Stopping is not yet Stopped; wait for the terminal state. Closing the page does not cancel execution.
• A complete analysis should identify its source, graph versions, completed queries, and final model or report files. Progress, an image, or model-generated prose alone is insufficient.