A maker uploads PDFs to a knowledge set feeding an AI agent, and the documents sit in Syncing indefinitely — or fail outright — after a connection used by indexing was replaced.
Symptom
One or more knowledge sources — documents added directly to an agent, or sources inside a knowledge set — stay in Syncing indefinitely, or move to Failed, and the agent can't retrieve from the new content. In a Simple agent's Knowledge list, the same states show as Syncing... and Sync Error. This can affect every source in the set at once, since they usually share the same indexing configuration.
Cause
Indexing runs each document through parsing, chunking, and embedding, and the models in that pipeline reach their provider through a connection — most commonly the Embedding Model's connection, but the same applies to any parsing, transcription, or extraction model. An Application parser can also use a connection of its own. If that connection was deleted, deactivated, or swapped for a replacement without updating the reference indexing uses, the pipeline can't reach the model or parser, and every document depending on it gets stuck or fails.
Fix
Find the embedding model (and any parsing, transcription, or extraction models) your indexing uses. Indexing can be set in several places, and the most specific one wins: the source's own Indexing tab, the knowledge set's Settings, Indexing, the agent's Settings, Indexing, or AI Setup, Agent Workflow, Indexing for workspace defaults. The Embedding Model is under Storage.
Go to AI Setup, Models (the model library) and find that model. Check its state — Connection Missing, Inactive Connection, or Connection Unavailable all mean the model currently can't be reached. The AI Models screen you open from Models in the main navigation shows whether a model's connection is active, but it can't switch the connection.
Use Switch Connection to point the model at a valid, active connection.
In the same Indexing settings, confirm the Embedding Model is still correctly selected. If the Parsing Method under Default Strategy is Application and the parser uses a connection, re-select an active connection in its Connection field — in the Default Strategy and in any per-format indexing strategy. If it's Agentic, check the parsing model the same way as in steps 2 and 3. An unset parsing method falls back to the built-in parser, so it doesn't block indexing on its own.
Re-index the affected documents. In a Simple agent's Knowledge list, hover the document that shows Sync Error and select Retry. For sources in a knowledge set, saving changes to the Indexing settings re-indexes all added knowledge — you're asked to Confirm Changes before it starts. Sync Knowledge only appears once a source is Completed, so you can't use it to restart a source that's Syncing or Failed.
To verify, confirm the source's Status moves from Syncing/Failed to Completed (or Scheduled if it's on a recurring sync), and that its Error Details no longer list connection errors for the affected documents. In a Simple agent, the document shows Synced: followed by the sync date.
Prevention
Indexing settings point at a model, and the model points at a connection. Replacing a connection doesn't update the models that used the old one — they keep pointing at it until you switch them in AI Setup, Models. Whenever you replace a connection (expired credentials, a new API key, a different provider account), switch every model that used the old connection to the new one, and re-select the connection on any Application parser. Otherwise indexing can fail silently across every source that shares that configuration.
Note: Because sources in the same knowledge set typically share indexing configuration, a single broken connection can stall or fail every document in that set at once — not just the most recently added ones.