Make.com RAG Agent Build

The client needed an experienced Make.com automation builder who could turn a RAG-agent architecture into a working automation flow inside a short three to four day delivery window. The scope was well beyond a simple Make.com scenario: it needed a structured data flow with two RAG agents and three LLM generation steps working together to produce the final output. Retrieval and generation also had to work together rather than being bolted on separately, since OpenAI embeddings had already been selected for the retrieval layer while Gemini was preferred for the generation layer, and those two parts needed connecting cleanly inside Make.com. Database mapping was required for every entry, so each incoming item could be stored, retrieved and processed consistently instead of being lost after a run. The architecture was visual: the client planned to share a Miro board showing how the full data flow should work, and that visual structure had to be translated into a practical Make.com implementation rather than treated as a reference document.

The solution was designed as a Make.com RAG automation framework connecting retrieval, LLM generation, database mapping and final output delivery into one structured product workflow. The data flow architecture was built around the client planned design with clear stages for input intake, retrieval, LLM processing, database mapping and final output generation, which made the scenario easier to test, maintain and expand in the next product-building phase. The RAG layer was organized around OpenAI embeddings so knowledge-base content could be searched and retrieved before the LLM steps generated the final answer, grounding the output in the provided knowledge base rather than generic model responses. Two separate RAG agent steps were planned inside the flow so different retrieval or reasoning stages could support the final output, giving the workflow a modular structure where each agent handles its own part of the information processing. The three LLM generation steps were aligned around Gemini for the non-embedding model layer, with each step able to generate a specific part of the final response before the workflow combined the outputs into one usable result. Database mapping was structured so each processed entry saved with the right fields, schema alignment and output values, keeping data organized beyond a single run and making the workflow suitable for product use. Integration points were prepared for the client-provided agent prompts and knowledge-base content.

A product-ready Make.com automation structure now runs a RAG-based data flow with retrieval, generation, database storage and final response handling. The Make.com workflow architecture covers the full product flow, with two RAG agent stages connected to the retrieval process and an OpenAI embeddings layer for knowledge-base retrieval. Three Gemini LLM generation steps produce the final output, and database mapping is structured for each processed entry so data stays consistent across runs. Prompt and knowledge-base integration points were prepared for the client-supplied content, and workflow documentation shows how data moves from input to final output. The business effect is a clear implementation path for turning a visual RAG architecture into a working automation, with the technical complexity separated into retrieval, generation, mapping and output stages so each part is easier to build and test, OpenAI embeddings and Gemini used together in one workflow, and the database layer prepared so entries process consistently.

Client retains ownership of the Make.com scenario, the agent prompts, the knowledge-base content, the embedding configuration, the Gemini model settings, the database schema and field mapping, and the Miro board the flow was translated from.