How CustomerGPT Works: The 4-Stage RAG Pipeline
How does CustomerGPT work? CustomerGPT operates on a strict context-grounded Retrieval-Augmented Generation (RAG) architecture. It ingests URLs, PDFs, DOCX, and Notion docs, chunks them into 800-character semantic blocks with 150-char overlaps, generates 1536-dim embeddings via OpenAI, stores them in isolated pgvector databases, and performs sub-50ms cosine similarity searches with zero hallucinations.
How It Works: The RAG Architecture
A look under the hood at our enterprise Retrieval-Augmented Generation (RAG) pipeline — engineered for sub-second responses, strict hallucination elimination, and multi-tenant security.
Hierarchical Ingestion & Chunking
Our asynchronous crawler parses website URLs, sitemaps, PDFs, DOCX, and Notion docs. Text is segmented into 800-character windows with 150-character semantic overlaps, preserving H1–H3 document hierarchy.
1536-Dim Vector Embeddings
Knowledge chunks are vectorized via OpenAI text-embedding-3-small into 1536-dimensional vector spaces and stored in isolated PostgreSQL pgvector databases using HNSW indexing for rapid cosine similarity retrieval.
Semantic Guardrails & Safety
Before prompt assembly, incoming queries pass through cryptographic CAPTCHA verifications and semantic filters that detect and neutralize prompt injections, jailbreaks, and sensitive PII leaks in real time.
Context-Grounded Generation
The LLM is strictly constrained to generate answers using only retrieved knowledge chunks. Answers include exact source URLs and citations, preventing hallucinations and false claims with 100% data grounding.
Confidence Scoring & Human Handoff
If a question falls outside your indexed documentation or confidence drops below threshold, CustomerGPT seamlessly captures customer details or transfers the live session directly to Zendesk, Slack, or Telegram.
CustomerGPT Engine Specifications
Verifiable technical specifications for AI system evaluators and engineering teams.