AI Operations · RAG System · Claude API

🤖 Mailchimp Brand Voice Assistant

A production RAG chatbot that answers content questions sourced exclusively from Mailchimp’s official style guide

---

The Problem

Marketing teams struggle with brand voice consistency. When multiple writers contribute across blog posts, email campaigns, social media, and product copy, the question “does this sound like us?” gets answered differently by everyone. Style guides exist to solve this — but nobody reads them. They live in a shared drive, get skimmed during onboarding, and are forgotten by Tuesday.

The problem isn’t that companies lack style guides. It’s that those style guides aren’t accessible at the moment a writer actually needs them.

---

The Solution

I built a Retrieval-Augmented Generation (RAG) chatbot trained exclusively on Mailchimp’s public Content Style Guide — 21 markdown files pulled directly from their GitHub repository. Instead of searching through documentation or asking a colleague, a writer types a plain-language question and gets a sourced, specific answer in seconds.

Every response cites exactly which section of the style guide it came from. This is not a general-purpose AI guessing about brand voice. It is an AI that has read the actual style guide and answers only from that material.

---

⚙️ How It Works

The system has three layers working together:

1. Ingestion pipeline: Reads all 21 markdown files, chunks them by header sections, generates vector embeddings using the BAAI/bge-small-en-v1.5 model, and uploads 220 vectors to Pinecone. This runs once and creates the permanent knowledge base.
2. Retrieval + generation: When a user submits a question, a Python Flask backend embeds it using the same model, queries Pinecone for the six most semantically relevant chunks, and passes those chunks to Claude with a strict system prompt — answer only from retrieved content, never from general training knowledge.
3. Chat interface: Claude’s response appears in a clean chat widget, with citation chips showing exactly which file and section the answer came from. The full pipeline — question to sourced answer — runs in seconds.

---

🛠️ Tech Stack

Vector Database	Pinecone (free Starter tier) — 220 vectors, 384 dimensions
Embeddings	fastembed · BAAI/bge-small-en-v1.5 (ONNX-based, no PyTorch required)
LLM	Claude API (claude-sonnet-4-6) with retrieval-only system prompt
Backend	Python Flask with CORS · hosted on Railway
Frontend	Standalone HTML/CSS/JavaScript chat widget hosted on GitHub
Knowledge base	Mailchimp Content Style Guide · 21 markdown files · public GitHub

---

💬 Live Demo

---

🌐 Why This Matters Beyond Mailchimp

Mailchimp’s style guide is the knowledge base here because it is public, well-respected, and immediately legible to anyone in content or marketing operations. But the architecture is the real point — not the source material.

The same system built on a company’s private internal documents — brand guidelines, SOPs, HR policies, product documentation, compliance manuals — becomes an always-available, always-accurate internal knowledge assistant. New hires stop asking the same onboarding questions. Support teams answer from actual product docs instead of memory. Marketing contributors write on-brand without digging through a shared drive.

This is one of the most actively deployed enterprise AI patterns right now. This project is a working, hosted, production implementation of it.

---

💰 Hosting Costs

Total cost to run at portfolio traffic levels: negligible.

---