Syncari's Pipeline AI Assistant brings conversational, bounded pipeline authoring into Sync Studio. The assistant uses SyncAI and the model provider configured for the Pipeline Assistant feature to read the active canvas, the entity schema, the connected sources, and the current chat history, then proposes graph edits you review and apply. The pipeline graph remains authored by you; the assistant's role is to draft the change so you can stage, inspect, and save it.
Pipeline AI Assistant edits one canvas at a time - an entity pipeline or a field pipeline - and matches the change to the canvas you opened. Requests that span both produce a combined response with the entity-pipeline change on the active canvas and per-field changes as cards in the chat.
Note: By using the Pipeline AI Assistant, you are agreeing to send the messages you type, the active canvas graph, the entity schema for the current entity, the list of connected sources and service credentials, and recent conversation history to the model provider selected in SyncAI settings. The assistant does not automatically mask or obfuscate data before it is sent to the provider.
Prerequisites
Before using the Pipeline AI Assistant, confirm the following settings:
- SyncAI is enabled for the instance (SyncaAI in Settings).
- A managed provider or BYOM provider is available for SyncAI model calls.
- The Pipeline Assistant SyncAI feature is enabled and mapped to a provider and model when your tenant requires explicit provider mapping.
- The user configuring the pipeline has access to Sync Studio, the pipeline, and the fields used in the request.
If the SyncAI feature is disabled, or if Pipeline Assistant is disabled in provider feature settings, the assistant icon will not appear in Pipeline Studio.
Using The Assistant
Open an entity or field pipeline in Pipeline Studio, then click the SyncAI icon to open the assistant panel. The assistant attaches to whichever canvas is currently active:
- On an entity pipeline canvas, the assistant edits at the entity scope - sources, the core entity, record-level functions and predicates, actions, and sinks. The assistant can also give a summary of what changes it would do per field pipeline, which you can review.
- On a field pipeline canvas, the assistant edits at the attribute scope - attribute sources, value transforms, the core attribute, and attribute sinks.
The active canvas is the source of truth for scope. A request that targets the other canvas is returned as a Clarification telling you which canvas to open.
The assistant classifies each request into one of the following operations:
- Create - build a new pipeline from scratch.
- Modify - add a node, change a node's configuration, route new edges, or change the core entity's dedupe and merge settings on an existing pipeline.
- Add Field Pipeline - open a new attribute pipeline for an existing field.
- Modify Field Pipeline - change one or more field pipelines on the entity. Field-pipeline changes always render as separate cards in the chat with their own Save button.
- Explain - describe what the current pipeline does, node by node.
- Clarification - ask a follow-up or give a prerequisite instruction when the request cannot be applied as-is.
The assistant resolves field references against the entity schema. If a request names a field that does not exist (for example, an invented apiName like `normalized_email` or `website_domain`), the assistant maps it back to the real apiName from the schema or returns a Clarification listing valid fields, rather than emitting a node that references a missing field.
The assistant does not currently delete nodes or edges. If a request implies removal, the assistant returns a Clarification.
Applying Changes
When the assistant proposes a change, the response includes a short summary, a reasoning note, and one or both of:
- An entity-pipeline change - staged directly on the active canvas. New nodes appear with a new-node highlight, and modified nodes appear with a modified-node highlight. Click Apply Changes in the chat to stage the change, then Save Draft on the canvas to commit.
- One or more field-pipeline cards — each card shows a per-field proposal with a deep link that opens the field pipeline canvas in a new tab with the staged graph and a Save Attribute Pipeline button. Field pipelines are reviewed and saved independently of the entity-pipeline change; you do not have to leave the entity canvas to act on them.
Highlights persist on the canvas across renders until you save the draft, at which point the assistant treats the changes as resolved and clears the highlights.
Examples
In this example we see how the LLM can suggest changes to both entity pipeline and field pipeline based on the requirement, my prompt is
"When an Account record arrives, normalize the company name (lowercase, strip Inc/LLC/Corp) and the website (strip protocol and www). Check for an existing Account matching on either normalized_name or website_domain. If found, merge incoming values in with latest-wins. If not, insert a new Account."
Here is how the output looks like
Each field-pipeline card has its own Save button and a deep link that opens that field's canvas with the staged change. The entity-pipeline filter is applied via the Apply Changes button in the chat.
Now let us see how a clarification looks like, here is a prompt
"For every new Contact, extract the domain from their email and look up an Account with that website domain. If found, set the Contact's account_id. If not, create a stub Account using enriched company data from the email domain and link the Contact to it."
Here the AI assistant does not have the full context to build out the pipeline. Hence it follows up with clarifying questions.
Logs And Troubleshooting
Each turn of the chat is stored as a message under the chat session with the operation type, the staged graph payload, the assistant's reasoning, and token counts. You can clear the session at any time to start fresh; clearing also removes the staged context the assistant carries between turns.
If the assistant returns a Clarification instead of a change, the message names the missing context - typically a field that does not exist in the schema, a source or service credential the tenant does not have, or a request that targets the wrong canvas. Provide the requested information in the next turn; the assistant continues the same session.
If the assistant icon does not appear in Pipeline Studio, check that SyncAI is enabled and the Pipeline Assistant feature is enabled in provider feature settings. If a staged change does not render on the canvas after Apply Changes, re-open the pipeline; the assistant's staged graph is preserved with the rest of your unsaved changes until you save or discard the draft.
A Word of Caution
LLMs are not accurate 100% of the time. It is always recommended you review the changes very carefully before applying and saving the pipeline. Save the current version and create a new version of the draft before applying any changes from LLM, so you have a checkpoint to go back to incase something goes wrong.