Before saving

A final review greatly reduces branch, variable, and JSON errors.

Structure

[ ] Start is connected to the first useful block.
[ ] Every block is reachable from the entry point.
[ ] Every route ends with a coherent answer or action.
[ ] After success, After failure, and Else branches are handled.
[ ] Agent loops have an exit route.

Naming

[ ] Names describe each block’s business role.
[ ] Output aliases are clear and unique.
[ ] Created variables contain no spaces or special characters.
[ ] Renaming did not leave old references behind.

Variables

[ ] References were selected through the {{ helper when possible.
[ ] The producer block is upstream on the same route.
[ ] REQUEST is the default and CONVERSATION is used only when required.
[ ] Temporary references such as {{http-1.body}} match actual identifiers.
[ ] Sensitive codeblock diagnostics were removed.

JSON and HTTP

[ ] Headers are valid JSON objects.
[ ] Accepted codes are arrays of numbers.
[ ] The body matches raw, formData, or x-www-form-urlencoded.
[ ] Content-Type and response format are consistent.
[ ] Timeouts cover normal service duration.

LLM, documents, and MCP

[ ] A model is selected for every LLM instruction and request.
[ ] The instruction variable points to the correct block.
[ ] For RAG, retriever_retrieve_information comes from the MCP catalog.
[ ] Retriever metadata matches the expected connector and sources.
[ ] MCP tools were fetched and limited to what is needed.
[ ] tools, genericToolCalls, and toolServerRouting come from the correct blocks.
[ ] The loop returns to the LLM after MCP - Call tools and has an exit route.
[ ] Tokens are supplied at runtime and never stored as plain text.

Final experience

[ ] The normal path was tested in Semantics Chat.
[ ] Every important failure was triggered and checked.
[ ] Final answers contain no raw technical data.
[ ] The assistant clearly acknowledges missing information.
[ ] Sensitive actions require confirmation.
[ ] The workflow is saved with the expected status.