Knowledge sources
PDFs, documents, website pages, tables, or records from an existing system.
Short Professional Guide
RAG in an Existing System
RAG is a pattern where the system retrieves relevant information from internal or external sources before the model answers. The answer is grounded in business knowledge, not only in the model’s general memory.
RAG is useful when AI must answer from documents, procedures, data, or a knowledge base that changes over time.
Official sources: Retrieval-Augmented Generation paper (Lewis et al.) · Supabase AI & vectors documentation
What does RAG solve?
A language model can write an answer, but it does not automatically know what is inside your company documents. RAG adds retrieval: the system finds relevant passages, sends them to the model, and the model answers based on that context.
It is useful for customer support, employee questions, document search, training systems, quote preparation, contract analysis, and any workflow where business knowledge changes over time.
What should be prepared before adding RAG?
Knowledge sources must be mapped, documents cleaned, content split into meaningful chunks, retrieval configured, and answers tested against real questions. Long documents without good chunking often produce weak answers even with a strong model.
Permissions matter as well. Not every user should see every document. RAG inside a business system must respect access control, privacy, and separation between clients or teams.
What makes the difference in answer quality?
Three things. Chunking by document structure, by headings and paragraphs, not by a fixed character count. Hybrid search: semantic search that understands intent, combined with keyword search that catches product codes, names, and clauses. And in Hebrew, keyword search must handle prefixes such as ו, ב, ל, מ, and ש, or it misses simple matches.
Permissions are filtered before information reaches the model, not after. If a document the user may not see enters the model’s context, an instruction to the model will not always keep it out of the answer. And it is measured: a set of dozens of real questions with correct answers, run after every change, checking whether answers are grounded in the sources and actually answer the question.
How does Itay Karkason implement RAG?
Itay Karkason connects RAG to existing systems through a staged process: knowledge mapping, indexing, user interface connection, answer testing, permission handling, and monitoring. The goal is not a generic chat tool, but useful answers grounded in business knowledge.
Components to Plan
Chunking
Splitting information into retrievable parts with enough context.
Permissions
Keeping users, clients, documents, and sensitive data separated.
Quality tests
Real questions, missing answers, contradictions, and cases where the system should say it has no information.
Frequently Asked Questions
Does RAG replace a database?
No. RAG is a retrieval and context layer over existing information. The database still manages the data.
Does RAG work in Hebrew?
Yes, but retrieval quality, document chunking, and Hebrew answers should be tested with real data.
Research Sources
This page was built as a short reference guide based on NotebookLM research and professional sources. Key sources:
Want to connect AI to your business knowledge?
Start by reviewing the documents and the current workflow.
Contact