Not for merge: Demo - Deep Search + InstructLab RAG integration

[nerdalert]

Example 1 (Components: Deep Search RAG + InstructLab 4-bit quantized)

• Perform a RAG query against a realtime collection of documents of data not included in the trained model.
• This could also represent private organizational documents that need to stay on premise, air-gapped or any other data privacy policy since all components are local.
  -Select the collection corresponding to the topic or have a generic data collection for all topics depending on the desired data policy.
• Without RAG, the model attempts to answer the question but its a hallucination since it is not trained on US Q1Y2024 economic data.

rag-demo1-economic-data.mp4

Example 2 (Components: Deep Search RAG + InstructLab 4-bit quantized)

• Perform a RAG query against a collection of Annual Report filings.
• The question is left generic to show the ability to query an entire corpus of documents vs drilling down into the collection and querying a specific document.
• The first query will use the top result (chunk) and feed it back to InstructLab for answering which happens to be Disney's filing.
• In the second query, we choose the exact document(s) for querying which happens to be Amazon's filing, thus returning information from that filing.
• There are many patterns to optimize RAG with sending larger and/or more chunks but this is a simple example exhibiting the importance of managing data sourcing and documents.

rag-demo2-accounting-data.mp4

[nerdalert]

Closing since there is a newer demo PR with docling running as a standalone service.