Build Custom Workflows Automatically with GPT-4o, RAG, and Web Search

## What the "Agent Builder" template does Need to **turn a one-line chat request into a fully-wired n8n workflow template—complete with AI agents, RAG, and web-search super-powers—without lifting a finger?** That's exactly what *Agent Builder* automates: 1. **Listens to any incoming chat message** (via the *Chat Trigger*). 2. **Spins up an AI architect** that analyses the request, searches the web, reads n8n docs from a Pinecone vector store, and designs the smallest possible set of nodes. 3. **Auto-generates a ready-to-import JSON template** and hands it back as a downloadable file—plus all the supporting assets (embeddings, vector store etc.) so the next prompt is even smarter. Think of it as your personal "workflow chef": you shout the order, it shops for ingredients, cooks, plates, and serves the meal. All you do is eat. --- ## Who will love this? * **No-code builders / power users** who don't want to wrestle with AI node wiring. * **Agencies & consultants** delivering lots of bespoke automations. * **Internal platform teams** who need a "workflow self-service portal" for non-technical colleagues. --- ## How it's wired | Sub-process | What happens inside | Key nodes | | -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | -------------------------------------------------------------------------------------- | | **Web Crawler** (optional) | Firecrawl scrapes docs.n8n.io (or any URL you drop in) and streams raw markdown back. | `Set URL → HTTP Request (Extract) → Wait & Retry` | | **RAG Trainer** | Splits the scraped docs, embeds them with OpenAI, and upserts vectors into Pinecone. | `Recursive Text Splitter → Embeddings OpenAI → Train Pinecone` | | **Agent Builder** | The star of the show - orchestrates GPT-4 (via OpenRouter), SerpAPI web-search, your Pinecone index and a Structured Output Parser to **produce → validate → prettify** the final n8n template. | `Chat Trigger → AI Agent → OpenAI (validator) → Code (extract) → Convert to JSON file` | *Every arrow in the drawn workflow is pre-connected, so the generated template always passes n8n's import check.* --- ## Getting set up (5 quick creds) | Service | Credential type | | --------------------------------------------------- | ---------------------------------------------------------- | | **OpenAI / Azure OpenAI** - embeddings & validation | *OpenAI API* | | **Pinecone** - vector store | *Pinecone API* | | **OpenRouter** - GPT-4 LLM | *OpenRouter API Key* | | **SerpAPI** - web search | *SerpAPI Key* | | **Firecrawl** (only if you plan to crawl) | Generic **Header Auth** → `Authorization: Bearer YOUR_KEY` | Each node already expects those creds; just create them once, select in the dropdown, hit **Activate**. --- ## What a typical run looks like 1. **User says:** "Build me a workflow that monitors our support inbox, summarises new tickets with GPT and posts to Slack." 2. *Chat Trigger* captures the message. 3. **AI Agent**: * queries Pinecone for relevant n8n docs, * fires a SerpAPI search for "n8n gmail trigger example", * sketches an architecture (Gmail Trigger → GPT Model → Slack). 4. The agent **returns JSON** — OpenAI node double-checks field names, connections, type versions. 5. A tiny JS Code node slices the JSON out of the chat blob and **saves it as `template.json`** ready for download. 6. You download, import, and done. --- ## Customising * **Switch the LLM** - plug in Claude 3, Gemini 1.5, or a local model; just swap the *OpenRouter Chat Model* node. * **Point the RAG at your own docs** - change the crawl URL or feed PDFs via the *Default Data Loader*. * **Hard-code preferred nodes** - edit the "User node preferences" in the system message so the agent always chooses *Notion* for databases, etc. --- ## Take-away notes * It's a **prototype** feel free to experiment with it to improve its capabilities. * **Have fun building!**