🧬 Brand Voice RAG & Anti-Slop Quality Gate

Most AI writing tools let you paste a "brand voice" description into a system prompt — something like "write in a professional but approachable tone". The problem is that language models do not reliably follow open-ended stylistic instructions, especially across long articles. The result is content that reads like a slightly better version of GPT-default output: technically correct, but hollow.

IndexPine's Brand Voice RAG works differently. Instead of describing your voice in a prompt, it learns your voice from your actual best content.

How Brand Voice RAG works

1. You provide IndexPine with examples of your best content: URLs to your top articles, uploaded PDFs of brand guidelines or whitepapers, or both.
2. IndexPine chunks this content into overlapping passages and creates dense vector embeddings for each chunk.
3. When generating a new article, IndexPine's embedding model retrieves the most stylistically relevant chunks from your content library in real time — the chunks whose tone, vocabulary, and structure most closely match the topic being written.
4. These retrieved passages are injected into the generation context so the model writes while actively referencing the way your real human writers have addressed similar topics.

What it captures that prompts cannot

• Sentence rhythm and length distribution — whether your brand favours short punchy sentences or longer flowing ones.
• Vocabulary choices — the specific words your writers choose when multiple synonyms exist.
• How you handle technical jargon — whether you define it immediately, use it freely, or avoid it entirely.
• Metaphor and analogy patterns — whether your brand explains concepts through analogies, data, or step-by-step.
• How you open and close sections — your characteristic transitions and rhetorical moves.

The Anti-Slop Quality Gate

After the Brand Voice RAG generation pass, every article goes through an automated quality gate before it is saved to your dashboard:

• Banned phrase check — any word or phrase in your banned list triggers a targeted rewrite of that sentence.
• LLM cliché detection — a secondary model scans for the 200 most common LLM filler phrases and flags them for rewriting.
• Passive voice threshold — if more than 20% of sentences are in passive voice, the gate fails and the most passive sections are rewritten.
• Factual consistency check — claims are cross-referenced against the research sources fetched during the research pass. Inconsistencies are flagged.
• Structural completeness — the article must contain all elements required by your structure rules (e.g. FAQ, TL;DR, key takeaways) or the gate fails and the missing elements are generated.