Document Q&A RAG Capstone

1. Ingestion pipeline

Chunk, embed, and index a small document collection.

• Documents are chunked with stable chunk IDs
• Chunks are embedded and stored with metadata
• Re-running ingestion does not duplicate unchanged chunks

2. Grounded generation with citations

Retrieve relevant chunks and generate an answer that cites them.

• Retrieval returns a ranked, deduplicated set of chunks
• The generated answer references specific source chunks
• A minimum relevance threshold gates whether generation proceeds

3. Safety and honesty

Handle low-evidence and adversarial cases correctly.

• A query with no supporting content returns an honest 'not enough evidence' answer
• Instructions embedded inside a retrieved document are not obeyed
• The pipeline never claims an exercise passed without deterministic validation