> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs-beta.getzep.com/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs-beta.getzep.com/_mcp/server. # Graph visualizer assistant > Use the graph visualizer assistant in the Zep dashboard to ask read-only questions about the graph you are viewing. The graph visualizer assistant answers questions about the graph you have open in the dashboard. It searches the graph, reads entities, facts, and episodes, and shows each retrieval step as it runs. Use the assistant to: * Inspect what Zep extracted from your data * See how two entities connect * Find the source episodes behind a fact * Find the earliest or most recent episodes, entities, or facts, or count entities of one type * Check a graph before you write retrieval code against it The assistant is also a working example of how an agent can use a Zep graph. Your agent can use the same graph search and graph read operations through the [API and SDKs](/searching-the-graph). Each panel step shows the operation, its arguments, and its result. This shows how an agent turns a question into graph retrievals. > **Note** > > The assistant is in limited rollout. When it is not enabled for your account, the visualizer does not show **Ask graph**. ## What the assistant can and cannot do The assistant is read-only. It cannot add, update, or delete graph data. Each conversation targets the one graph that is open in the visualizer. The assistant cannot read a different graph, project, or account. When you close the visualizer, the conversation is discarded. The assistant grounds each answer in data that it retrieved from the graph. The assistant does not answer from general model knowledge. If a search returns no matches, the assistant says that it found no evidence. ## Open the assistant #### Open a user or shared Context Graph In the Zep dashboard, open a user page or a shared Context Graph page in your project. #### Open the graph visualizer Open the relationship graph for that user or graph. #### Use the chat panel The chat panel opens next to the canvas when the visualizer opens. Drag the boundary to resize the panel. Click **Hide chat** to close the panel, and click **Ask graph** to open it again. On a narrow screen, the panel stays closed until you click **Ask graph**. ## Ask a question Type a question and press Enter, or click **Send**. Press Shift+Enter to insert a new line. Ask about one topic at a time. The assistant runs a bounded number of retrievals for each question, so a narrow question gets a better answer than a broad one. You can ask follow-up questions that refer to earlier answers in the same conversation. Click **Stop** to cancel a turn. The panel keeps the steps that already ran. ## Read the answer Before it retrieves graph data, the assistant submits a plan. The panel shows the plan as a **Plan** step with the number of retrieval steps. The assistant then runs retrieval steps with **Search**, **List entities**, **List facts**, **Neighbors**, and **Details**. Up to four steps that do not depend on each other run at the same time. Each step shows the operation, a summary of its arguments, a summary of its result, and the time the operation took in milliseconds. Expand a step to read its full arguments and result. When the answer cites graph evidence, the panel lists the cited nodes and facts under the answer. It shows six items first. Select **Show all** to expand the list, or **Show on graph** to highlight the cited evidence. When a turn completes, the panel applies the cited evidence as the canvas highlight. The steps and the canvas are linked: * Select a step to highlight the entities and facts it retrieved on the canvas. * Select an entity name under the answer to focus that entity on the canvas. The canvas shows a bounded view of a large graph. When a step retrieves an entity that is not loaded on the canvas, the panel lists it under **Not on canvas**. The panel shows a result marker for each retrieval step. Search starts with `ranked sample:` and returns the most relevant items. Search does not return every match or support pages. **List entities**, **List facts**, and **Neighbors** return matches page by page. The first line shows the item positions. `items -, more on page cN:` means that more matches remain. The assistant uses the page handle to get the next items. `complete, items - of :` marks the final page. An empty list shows `complete, 0 items:` and `- no matches`. The assistant keeps list text within a 4,000-character budget when possible. If one item exceeds the budget, the assistant returns that item and adds `[truncated]` to the clipped text. The page handle stays on the first line when more matches remain. Each Neighbors item uses one line for one matching fact and its neighbor entity, with the entity's labels after its name. An entity with more than one matching fact appears once for each fact. The item does not include the entity's summary or attributes. A turn stops after 60 seconds or when the assistant reaches its model-call limit. The assistant answers in the language of the current question. ## Availability With a Bring Your Own Model (BYOM) configuration, the assistant runs only when the provider is OpenAI. Assistant turns then use your OpenAI key. With another provider, including Azure OpenAI, **Ask graph** is disabled. Its tooltip says: “The BYOM configuration does not have a compatible model configured.” ## Limits * A question can be up to 2,000 characters. * A turn stops after 60 seconds. * The conversation keeps the last 10 turns. The panel tells you when earlier turns left context. * Each member can send 10 turns per minute, and each account 30 turns per minute. When a turn cannot complete, the panel states the reason, such as a timeout, a rate limit, or a model that did not respond. Retry, or ask a narrower question. The assistant is available only in the dashboard. The assistant is not available through the public API or SDKs. Use [graph search](/searching-the-graph) and [reading data from the graph](/reading-data-from-the-graph) for programmatic reads.