Web console

One console at localhost:7280 covering both administration and data work. Sign in with a seeded login - superadmin / superadmin has every permission in the seeded organization.

It is not a platform-wide account: superadmin owns the seeded organization, so it administers that organization rather than the whole installation.

The seeded logins do change what the console shows you - it reads the signed-in account’s role and hides what that role cannot do. What they do not change is what the server would allow: in this evaluation setup every call the console makes carries one shared credential, so the console’s gating is a faithful preview of the roles rather than a test of the server enforcing them. Configure an identity provider to exercise that.

Two areas

Administration manages the tenant: organizations, projects, users and their roles, and API keys. This is where you look when a permission question comes up, or when you need a second API key for a different client.

Workbench is where the data is: collections, graphs, queries and visualization. Most of your time is here.

What you can see depends on your role, which is the point of the three seeded logins - sign in as member to see the console as someone with data access but no administrative power.

Help that follows you

The question-mark button in the top bar opens a help panel that always describes the screen you are on, down to the individual tab of a collection or graph. It docks beside the content rather than over it, and it stays open or closed as you left it while you move around. Controls whose purpose is not obvious carry a tooltip, so the panel is for "what is this screen for" and the tooltips for "what does this field do".

Sample datasets

The fastest way to have something real to query. Both list views carry a Try Samples button:

  • Graphs offers a movies graph and a retail graph - nodes, typed edges and pre-computed embeddings.

  • Collections offers three document sets, including an FAQ knowledge base.

All templates are pre-selected in the dialog; press Create Selected to build them. Progress shows per template, and each takes a few seconds. Deselect the ones you do not want.

Because the samples ship with embeddings, vector search works immediately - no provider account and no configuration.

Creating a collection or graph

Collections → New asks for a name and a field list. Each field needs a type; vector fields also need dimensions and a similarity function. Declaring types is what makes filters fast later, so it is worth a moment’s thought rather than making everything text.

If a language model is configured, a wizard will propose a schema from a description of your data - useful as a starting point, though check the field types before accepting it. Without a model configured, you define fields directly; nothing is lost.

Graphs → New works the same way, and you add edge types as you create edges.

Querying

Each collection and graph has a query view: write CyQL, run it, read the results as a table.

The editor knows the language and it knows your schema. Keywords, functions and string literals are highlighted, completion offers the CyQL keywords and built-in functions alongside your own fields - the metadata, full-text and vector fields of the collection or graph you are looking at - and it narrows by context, so it offers field names after a ., full-text fields first in an ON clause and vector fields first after SIMILAR TO. Syntax is checked as you type and a mistake is marked in place, so a typo does not cost you a round trip to find out.

Two things worth knowing:

  • Use parameters. The view has a parameter panel. It avoids quoting problems and is a better habit than editing values into the query text.

  • score(…​) is a column. Return it whenever you use SEARCH or SIMILAR TO - the ranking score tells you whether a result set is genuinely good or merely the best of a bad batch. It takes the variable it scores, so score(d) for a document and score(m) for a node.

Query history is kept, so an afternoon of exploration is recoverable.

Graph visualization

A graph opens as an interactive force-directed view, in 2D or 3D. Nodes are coloured by label and sized by connectivity.

The controls above the canvas separate the two things you do to it. Load data opens a dialog: what to put on the canvas - a sample of the graph or one node and its neighbours - how much of it, and which labels to require or exclude. It fetches once, when you press Load. Display opens a panel for how what is there is drawn - layout, 2D or 3D, what the colour means - which changes nothing on the server. Beside them, Fit view frames what is loaded, and Reset view puts the whole view back to how it opened, including the label filter and an empty canvas.

The line between the two reports what the canvas holds and whether a filter is narrowing it, so an active filter is visible without opening the dialog.

Colouring by community needs a graph that has been through detection. If it has not, or if nothing currently on the canvas belongs to a community at the chosen level, the colouring cannot be applied: the selector goes back to colouring by label and the panel says which of those it was, directly under the selector.

  • Drag to reposition, scroll to zoom, click a node to see its fields.

  • Expand a node to pull in its neighbours, which is the natural way to explore outward from a search result.

  • The result of a query can be visualized rather than tabulated - run a traversal, then look at the shape of what came back.

  • The chips on the canvas are a legend, not the filter: each one is a label with its colour and its count, and clicking one hides it locally without asking the server for anything.

Visualization is for understanding structure, not for rendering everything. A view of a few hundred nodes is informative; a view of fifty thousand is a hairball. Filter or start from a search result.

Communities

Community detection groups a graph into densely-connected clusters. The Communities tab shows what a detection run left: how many communities, at how many levels, and how many carry a written summary. It lists one level at a time, because the hierarchy is read from the fine end while a tree would bury it at the bottom - and on a real graph the coarsest level is usually one large community plus a long tail of singletons, which a size column shows honestly and a tree does not.

Detecting is not free and it replaces whatever was there, so it asks first. Re-detecting keeps the ids of communities whose membership did not move, which is what lets their summaries survive.

Selecting a community shows what it contains, and can push its member nodes into the Explorer or the Vectors view. Those are nodes, not communities: above the finest level a community’s members are its child communities, and the view resolves that before handing anything on.

A community summary is a short report about one community. CYROCK.AI DB stores and searches them; the text is written by an application, and with a language model configured the console is one - it generates them bottom-up, so a broad community is summarised from its children’s reports rather than from a sample of its thousands of entities.

Read them before storing them. The console generates first and writes only when you say so, because that is the only point at which a report is cheap to fix: the model has already been paid for by then, and a report that is confidently about the wrong thing looks exactly like one that is right. You can edit any of them afterwards.

Global Search then asks questions of those reports rather than of the nodes - "what themes recur across all of this", which no single node answers. It searches every level at once unless told otherwise. With no summaries stored it says so rather than reporting no results, because those are different problems.

If a language model is configured, you can describe what you want in plain language and have it translated into CyQL. The generated query is shown before it runs.

Read it. Translation is a good accelerator and an unreliable oracle - it is genuinely useful for recalling syntax you half-remember, and it will occasionally produce something that runs and answers a subtly different question.

Monitoring

The console surfaces engine health, storage size and slow queries. For anything continuous, scrape the Prometheus endpoint instead - see Operations.