Inference Scaling for Long-Context Retrieval Augmented Generation

_{scroll ↓ to Resources}

Note

investigate how performance scales with increasing magnitude of the inference compute?
they consider 2 advanced RAG modifications
- Demonstration RAG (DRAG) combines RAG with few-shot examples; its inference compute scales with both number of documents and number of queries
- Iterative Demonstration-Based RAG (IterDRAG)
  - Decomposes the query into simpler sub-queries.
  - For each sub-query, performs retrieval and uses fetched context to generate intermediate answers.
  - After all sub-queries are resolved, the retrieved context, sub-queries, and their answers are combined to synthesize the final answer.

table file.inlinks, file.outlinks from [[]] and !outgoing([[]])  AND -"Changelog"