RAG-Powered Support Bots on WordPress in Under a Day

The Problem With Naive AI Chatbots

Connecting a raw OpenAI GPT-4 API key directly to a website chat widget is a recipe for disaster. Customers ask about refund policies, and the bot hallucinates a '100% money-back guarantee forever' that your business never offered.

To use AI safely in production, you must implement Retrieval-Augmented Generation (RAG). The model is forbidden from answering from its general training weights; it is strictly constrained to answer using verbatim excerpts retrieved from your verified documentation.

How RAG Works in Practice on WordPress

1. Ingestion: A cron job reads your knowledge base articles, product descriptions, and shipping FAQs from WordPress.

2. Chunking & Embedding: Articles are split into 500-token chunks and converted into 1536-dimension vector embeddings using `text-embedding-3-small`.

3. Vector Storage: Embeddings are indexed in a vector database (e.g. Supabase pgvector or Pinecone).

4. Semantic Query: When a visitor asks 'Do you ship to Germany?', the system retrieves the 3 most semantically similar text passages.

5. Grounded Generation: GPT-4o-mini synthesizes a concise, polite answer citing only those 3 passages.

The Vector Retrieval SQL Function

Here is the PostgreSQL `pgvector` function used in Supabase to find semantically matching documentation chunks:

CREATE OR REPLACE FUNCTION match_documents (
  query_embedding vector(1536),
  match_threshold float,
  match_count int
)
RETURNS TABLE (
  id bigint,
  content text,
  similarity float
)
LANGUAGE sql STABLE
AS $$
  SELECT
    documents.id,
    documents.content,
    1 - (documents.embedding <=> query_embedding) AS similarity
  FROM documents
  WHERE 1 - (documents.embedding <=> query_embedding) > match_threshold
  ORDER BY similarity DESC
  LIMIT match_count;
$$;

Human Escalation: Knowing When to Hand Off

If the similarity score of the top document is below 0.72, or if the user displays frustration sentiment, the bot triggers an automated Slack alert or opens a direct WhatsApp chat with an on-call human representative.

Operational Cost & ROI

A store handling 5,000 monthly customer inquiries typically deflects 70% of repetitive tickets (order tracking, sizing, returns) automatically, saving over 40 hours of human support time every single week.