Dave Page: Building Ask Ellie: A RAG Chatbot Powered by pgEdge

Why It Matters

Ellie proves that enterprises can add RAG capabilities using existing PostgreSQL investments, reducing infrastructure complexity and cost. It also demonstrates a secure, extensible architecture for AI‑assisted knowledge bases.

Key Takeaways

• Ellie uses PostgreSQL as core RAG engine.
• Vectorizer auto‑generates embeddings via OpenAI.
• RAG Server streams answers with Claude Sonnet model.
• Frontend validates input, monitors prompt injection attempts.
• Deployment uses Ansible, Cloudflare Tunnel, no separate vector DB.

Pulse Analysis

The rise of Retrieval‑Augmented Generation has pushed vendors to stitch together disparate vector stores, API gateways, and orchestration layers. pgEdge flips that script by leveraging PostgreSQL 18 and its extensibility to host the entire pipeline. Docloader crawls documentation, cleans HTML, and stores structured text in a relational table, while the Vectorizer extension watches for changes and automatically pushes content to OpenAI for embedding, persisting vectors with pgvector. This tight integration eliminates the need for a dedicated vector database, simplifying data governance and reducing latency.

On the generation side, the pgEdge RAG Server acts as a thin orchestration layer. It performs similarity search against the stored embeddings, selects the top‑20 chunks within an 8,000‑token budget, and forwards them along with the user query to Anthropic's Claude Sonnet. The server streams partial results via Server‑Sent Events, delivering a responsive chat experience. Security is baked in: the RAG Server runs on localhost, accessed only through a Cloudflare Tunnel and a proxy function that sanitises headers and masks secrets, limiting the attack surface compared to open LLM endpoints.

From an operations perspective, the solution is fully automated with Ansible playbooks that provision Debian EC2 instances, install pgEdge Enterprise Postgres, configure extensions, and manage secrets via Ansible Vault. Frontend logic, written in vanilla JavaScript, adds conversation compaction, prompt‑injection monitoring, and local storage persistence, turning the widget into a lightweight observability layer. For organizations already running PostgreSQL, Ellie demonstrates a cost‑effective path to embed AI‑driven support directly into existing documentation portals, accelerating knowledge retrieval while maintaining strict data controls.