Please add a demo of Retrieval-Augmented Generation with embedding and text-generation model.

RAG Data Not Following System Prompt / Not Restricting to Retrieved Context

Would you add a demo/example of a Retrieval-Augmented Generation (RAG) pipeline that:

Uses Markdown content as knowledge base

Converts it into embeddings via an Embedding Model

Retrieves relevant chunks

Sends them to a Text Generation Model

Ensures the model strictly answers ONLY from retrieved context

Currently, the model:

❌ Does not follow the system prompt strictly

❌ Generates answers outside of retrieved (RAG) data

❌ Hallucinates or adds extra knowledge

✅ Expected Behavior

The Text Generation Model should:

Only answer using retrieved RAG context

Refuse to answer if context is insufficient

Follow system prompt strictly?

❌ Current Behavior

Model ignores restriction instructions

Uses general knowledge instead of retrieved chunks

Responses are not grounded in RAG data.