Day 4: Your First AI Employee — Brain/Memory/Tools/Prompt, a Pinecone RAG Support Agent in n8n, and When NOT to Use an Agent

Brain = the LLM doing the reasoning. Memory = a store of previous interactions so the agent has continuity - and it is a judgment call: a website chatbot needs it because users expect the session to remember; an email agent does not, because 'the full thread is already retrievable' and conversations resume weeks later. Tools (skills in the hiring analogy) are what let the agent act - read a doc, query a database, create a draft; give a narrow, high-quality set, 'specialists not generalists', and split into sub-agents when the list grows. The system prompt encodes guardrails, how to interpret queries, and when to call which tool; it is 'probably the most important element' and is refined for months against varied real queries.

The explicit design decision to leave memory out of the support agent while insisting the same brand's website bot would need it is the lesson: components are choices, not a checklist.

The problem: pasting a whole policy document into a prompt overflows context and invites hallucination. So: (1) chunk the document, (2) embed each chunk into a numeric vector with an embedding model - 'machines understand binary language. Numbers', (3) store chunk text + vector in a vector database, (4) at query time embed the question into the same space and retrieve the closest chunks. Dimensionality is explained with a 2-D fruit plot, then 'these balls are operating at 1024 dimensions', which the human brain 'cannot even comprehend.'

Ingestion is a separate, one-time workflow re-run only when the policy changes: Chat Trigger (file upload on) -> Extract from File (PDF, binary field data0) -> Pinecone Vector Store 'Insert Documents' with OpenAI text-embedding-3-small at 1024 dims, Recursive Character Text Splitter, chunk 500 / overlap 100 / batch 100 - 13 chunks into namespace 'ecom policy' in index 'customer support policy doc' (Pinecone free tier, 5 indexes). Chunk-size rule: ~500 characters when paragraphs answer discrete questions; larger when a whole page is needed. Overlap exists so an answer is never split across two chunks. The embedding model and dimension at query time MUST match ingestion.

Gmail Trigger 'On message received' (OAuth; poll every minute in production; Simplify ON to strip payload noise; PIN the test email so a security alert arriving mid-build does not replace your fixture). Edit Fields extracts email body (from snippet), sender email, thread ID, subject. OpenAI Message a Model (GPT-4.1 mini) classifies against a topic list (tracking, refunds/returns, subscription, technical, billing) and returns {"customer_support": true/false} with Output Format = JSON Object. Switch on that boolean: 'no' terminates - no tokens spent on irrelevant mail. AI Agent node (GPT-4.1 mini brain, no memory, trigger source 'defined below', input = subject + body) with two tools: Vector Store Question Answer (Pinecone index/namespace, same embedding model, synthesis model GPT-4.1 mini) and Gmail Create Draft (subject/message 'let AI define', threadId = extracted thread ID so the draft lands IN the thread, to = sender). System prompt: role, read the email, use the knowledge-base tool, use the draft tool, then produce a concise summary; sign as 'Kelly from Sleepy Owl'. Slack Send Message posts that summary as the human-in-the-loop notification.

Live test: 'I've received a defective product. Can you explain what your return policy is?' -> the agent retrieved the 30-day return/refund policy and drafted a threaded reply. The team then assigns a simpler expense-tracker agent for the breakout and promises the CS agent JSON, prompt and policy PDF with the recording (import via the three-dot menu -> Import from file).

Whenever an LLM's answer feeds program logic (If, Switch) rather than a human, it must be structured. A chatty 'I have checked your email and this definitely looks like...' cannot be routed on; nor can JSON-shaped text inside a plain string. Setting Output Format = JSON Object on the OpenAI node makes the field a real, draggable variable. JSON Object suits a single pre-known field (the boolean here); JSON Schema is the right mode when you extract several fields with a structure defined up front. Verbosity was left at 'medium'.

The companion fundamental, re-taught mid-build: system prompt vs user prompt. The system prompt is the invisible behaviour/guardrail layer (why ChatGPT refuses certain requests on the same GPT model you can call via API); the user prompt is the content. A supplement brand's bot should refuse a ramen recipe not because it cannot answer but because the system prompt scopes it.

Agent nodes earn their cost when the step is genuinely dynamic - interpreting free text and deciding which tool to call, in what order. When input and output are already known (classify this, format that, append a row), a plain LLM call, filter or switch is cheaper, faster and more predictable. Harshad names this as the most common mistake of less experienced builders, and the office hour reinforces it: use 'OpenAI send message' for fixed tasks, the agent node only when tool choice is real.

Related office-hour rulings: n8n vs Make.com (Make bills per module execution, expensive at scale); save workflows with Cmd/Ctrl+S, export JSON via the three-dot menu, push to GitHub; never share API keys - the shared test key is a liability, get your own at platform.openai.com (developer API, per-request billing) as distinct from chat.com (consumer subscription).

Video: Sora 2 and Google Veo reject real-person uploads (Sora's 'Cameo' is the sanctioned self-avatar route); HeyGen permits real people. Pixel upscalers (free online) preserve consistency better than latent upscalers when you only want resolution. Seamless loops: Sora has a native start/end match; otherwise split at a middle frame. Watermarks: free 'video inpainting' removers. Character consistency: image-to-video with reference images, never text-to-video, and an iterative feedback loop.

n8n: to keep a form on your own domain use a Webhook trigger instead of the n8n form node; Gmail-node failures with Google Sheets are usually a wrong Google account - unlink, relink, grant full access; production vs test URL on the form trigger. RAG: fully offline knowledge base via Msty 'knowledge stacks' (parallel stacks per domain, or Pinecone/Chroma if you want a real DB) paired with a vibe-coded front end later in the week. Ebooks: write and format by hand (Pages), Nano Banana for the cover. Email deliverability: convert video to GIF for HTML emails to avoid spam flags. Closing note from the mentors: no hand-holding - practice the worksheets, do Option 2, screenshot errors into chat.