RAG Knowledge Base

How RAG Works in BPract Agents

When a visitor asks a question, BPract Agents does not just pass it to the LLM and hope for the best. The system first converts the question into a vector embedding using the same model that indexed your content. It then performs a similarity search across your entire knowledge base using pgvector in PostgreSQL, retrieving the most relevant chunks of content. These chunks are injected into the LLM prompt as context, ensuring the AI response is grounded in your actual business data rather than general training knowledge. This dramatically reduces hallucinations and ensures accuracy.

Multiple Ingestion Methods

• Website crawler automatically follows links and indexes up to your configured page limit per crawl session
• File upload handles PDF, DOCX, plain text, and CSV formats with automatic text extraction
• Direct text input lets you paste FAQs, product specs, or policy documents right in the admin panel
• Scheduled re-crawling keeps your knowledge base in sync as your website content changes

Embedding and Retrieval Quality

BPract Agents uses state-of-the-art embedding models to convert text into high-dimensional vectors. The pgvector extension in PostgreSQL provides fast approximate nearest-neighbor search, returning the most relevant chunks in milliseconds even across thousands of documents. You can configure chunk size and overlap in the admin panel to balance between retrieval precision and context breadth. Larger chunks preserve more context around key information, while smaller chunks improve precision for specific factual queries.