I recently switched my agent from Haiku 4.5 to Sonnet 4.6. RAG is configured at the Model Settings level with Message Processing set to No Processing in the Chat block — this configuration worked correctly on Haiku. Since switching to Sonnet, the agent no longer retrieves the data from my knowledge base during full workflow runs. However, when I test directly in Sandbox on Sonnet, RAG retrieves correctly. What is different between Sandbox and a full workflow Chat block run in terms of RAG execution, and what do I need to change to make Sonnet retrieve correctly in the full workflow?
Hi @Neil,
The Sandbox in Model Settings is where you can run prompts to quickly review answers. It’s available for you as the builder, running RAG directly from those settings. The full workflow Chat block uses its own Message Processing setting, and since yours is set to No Processing, the Data Source is not queried.
Changing from Haiku to Sonnet doesn’t affect RAG retrieval. Enabling RAG in the Chat block’s Message Processing setting should fix it.
Thanks Alex. To clarify my current setup: I have RAG enabled via the toggle in Model Settings beneath the model selector, and Message Processing in the Chat block is set to No Processing. This configuration worked correctly on Haiku 4.5 — the doc was retrieving in full workflow runs. After switching to Sonnet 4.6, retrieval stopped working in full workflow runs but works correctly in Sandbox. Should I leave my Model Settings RAG toggle as-is and also enable RAG in the Chat block Message Processing, or should I switch entirely to the Chat block approach? I want to make sure I don’t break a configuration that was working.
Hi Neil,
That’s correct, the fix here is to enable RAG in the Chat block’s Message Processing settings. That’s what actually controls retrieval in chat during full workflow runs.
You can keep the RAG enabled in the Model Settings, as it won’t affect the end users’ workflow.