Skip to main content

Workflows: the knowledge-graph update node

For users who maintain a knowledge graph (an AMPG, Augmented Multi-resolution Property Graph, in the builder) and want its next version trained from a workflow's results. The AMPG update step builds a training document from the connected steps' results and the base version's documents and creates the next version from it; it does not ingest a PDF you supply (for that, build a version from the graph's page: Knowledge graphs: build a version). The graph needs a Ready version with training documents; the workflow needs a connected Task step.

Steps​

  1. Click AMPG update in the Add a step rail and drag from each Task whose learnings should feed the graph into the card.
  2. Open the drawer and pick the graph under Your AMPGs ("Currently vN; this step creates vN+1.").
  3. Read Base: "Starts from vN · k documents" with the document names. The base is always the graph's newest Ready version when the step runs, never the first version or a merge, so a weekly run builds on last week's version.
  4. Under Sources, every connected step is a ticked checkbox with its handle; you can add a workspace pattern such as projects/conversations/**/memo*.pdf. At least one source is required.
  5. Choose a Mode. Rewrite (default): "Opus reads version N's documents and the learnings from the connected steps and rewrites them into one new training document. Version N+1 is trained on that document." Append: "Version N's documents are kept as they are; a new document with only the learnings is added."
  6. Leave Hold the new PDF for my review before creating the version on (the default) to read the document first: "An Inbox card shows the built PDF with Approve · Send back · Stop before the AMPG version is created."
  7. Switch on Update all agents using this AMPG (off by default) to move every pinned agent to the new version on creation; the drawer lists Agents using this AMPG.
  8. Note Destination: users/{me}/ampg/{kb}/training/{YYYYMMDD}_v{N+1}.pdf, the graph's training folder.
  9. Save the version, activate it and run (Run, schedule, approve).

What you should see​

With the hold on, the card and your Inbox show the built PDF with Accept, Send back and Stop before any version exists. When the run finishes, the Summary's AMPG changes block reads "Created name vN from vM + learnings · Rewrite" (or "· Append") with Open the PDF, KB versions and, when adoption is yours, "adoption left to you" and Publish & adopt; the AMPG tab reads "AMPG version created · graph vN" and "· adopted k agents" when agents moved. The version builds like any other on the graph's page (Knowledge graphs); when adoption is yours the Inbox also shows "AMPG «name» vN is ready" with Publish & adopt, Publish only and Review files.

Notes​

  • Rewrite trains vN+1 on one consolidated document; Append trains it on the base documents plus one new document of learnings.
  • Agents keep their pinned version until moved: tick Update all agents using this AMPG, click Publish & adopt afterwards, or move them from the graph's page or the agent's Configure dialog (use, move agents, delete).
  • One version builds at a time per graph: the step cannot create its version while another version of the same graph is still building; 50 versions per graph.
  • "Choose the AMPG to update.": pick a graph. "Pick at least one source: the learnings come from the connected steps.": connect or tick one. "No training documents found for vN": build a version with documents first; the step fails rather than rewrite nothing.
  • The step failed: the Summary's Failed block names it and the versions table shows the failed attempt; fix and Re-run. You sent the PDF back: the step rebuilds with your note and holds again. Agents still read the old version: click Publish & adopt or move them by hand.