Skip to main content
Version: 1.0.11

Knowledge Graphs (AMPG)

A Knowledge Graph — branded AMPG (Augmented Multi-Resolution Property Graph) — turns a document you upload into a structured graph of connected information that agents can retrieve from when answering questions. This page covers creating a graph from a PDF, following the build, working with versions, and managing the agents that depend on a graph.

What a Knowledge Graph is​

You give AlphaAgent a document; AlphaAgent builds it into a knowledge graph and stores that graph in your own knowledge-graph database, inside your account. When an agent that is wired to the graph answers a question (in Deep Traversal mode), it retrieves relevant information from the graph as context. Your document and the resulting graph never leave your account.

A Knowledge Graph is also a prerequisite for Deep Traversal: every Deep Traversal flow begins with AMPG grounding against a graph, so an agent without a Knowledge Graph wired in can only use Express mode. Wire a graph into an agent's configuration to unlock Deep Traversal for it.

This page describes what you put in, what you get out, and the lifecycle of your document. For the data-handling and residency detail, see Knowledge graphs in depth.

What makes AMPG different​

AMPG is built to retrieve by meaning, not keywords. Two things make it well suited to dense, specialized documents:

  • Multiple levels of detail. Your document is structured at several resolutions at once — whole document, sections, passages, and individual phrases — so an agent can pull either broad context or the single precise line that answers a question.
  • Relationship-aware retrieval. When an agent searches the graph, it first casts a wide net by meaning (so paraphrases, synonyms, and even misspelled or never-before-seen terms still match), then re-ranks what it found by the specific relationship the question is about.

The payoff shows up on documents where the same entity means different things in different contexts — for example, a bank that is an underwriter in one place, a counterparty in another, and a prime broker in a third. AMPG keeps a single, connected view of that entity and adapts it to the role in play, instead of forcing you to maintain a rigid schema. For the conceptual model and a worked example, see Knowledge graphs in depth.

Creating a Knowledge Graph​

From the Knowledge Graphs list, click Create and provide:

  1. Name and description. The description is what teammates see in the catalogue and in the chat composer's knowledge picker, so make it identifying.
  2. A PDF. A single PDF document is the source for the graph.

The PDF is validated in your browser before anything is uploaded:

  • Size: up to 50 MiB.
  • Pages: up to 100 pages.

If a file is too large or has too many pages, you'll see a message stating both the limit and your file's actual size or page count.

Following the build​

Building a graph runs through five visible stages, shown as a progress indicator:

  1. Preparing Knowledge Graph
  2. Uploading Document
  3. Indexing Document
  4. Building Graph Embeddings
  5. Finalizing Knowledge Graph

Live logs​

While the build is indexing the document and building graph embeddings, you can open a logs view to watch progress in real time. The other stages are quick and produce little output, so the logs button is active only during those two stages.

Builds run in the background​

A build can take anywhere from an hour to several hours depending on the document. You don't have to wait on the page:

  • If you navigate away, the build continues on the server. Studio warns you before you leave.
  • If you reopen a graph that is still building (or refresh the page), Studio drops you straight back into the live build view rather than a stale screen.

Stopping a build​

You can Stop & delete a build only while it is still in the early ingestion phase (before embeddings begin). Once embeddings are being built, the build can no longer be interrupted — let it finish, then delete the graph from the list if you don't want it. Stopping permanently deletes the graph and everything created for it so far, and you'll be asked to confirm.

Versions​

Each Knowledge Graph is versioned. To refresh a graph with updated source material, open it and click Update, then upload a replacement PDF (same size and page limits). The update runs through its own staged build (Uploading, Indexing, Building embeddings, Finalizing) and publishes a new version. Agents pin to a specific version, so existing agents keep using the version they were wired to until you re-point them.

Viewing a graph and its dependents​

Open a graph to see its details and the KG Map — the agents whose active configuration uses this graph.

Updating a graph that agents depend on​

If active agents depend on a graph you want to update, Studio guides you through a safe sequence so those agents aren't disrupted:

  1. In the KG Map, click Disassociate All. This removes the graph reference from each dependent agent while leaving the agents themselves active.
  2. Update the graph to create a new version.
  3. Click Re-associate All to reconnect the agents to the graph.

When you click Update on a graph with active dependents, an informational panel reminds you of these steps and lists the affected agents and swarms.

Deleting a graph​

Deleting checks for impact first. If agents depend on the graph, a guard lists them and the delete is blocked until you remove those references. If nothing depends on it, a confirmation lets you delete.

  • Agents — wire a Knowledge Graph to an agent (max one per agent).
  • Chat — Deep Traversal retrieves from a graph and shows the path in the Thought Map.
  • Knowledge graphs in depth — where your document lives and how retrieval works.