Luminate AI

Corrective RAG (CRAG) Decision Pathways
Correct
The local hierarchical index contains sufficient information. The system generates the response using local documents only.
Ambiguous
The local index is partially helpful but lacks detail. The system retrieves fresh context from arXiv and merges both sources.
Incorrect
The local index is irrelevant. The system completely bypasses local data, executing a live search on arXiv for a fresh answer.

Try Pre-computed Evaluation Queries (Offline Cache)

Select any query below to see its cached answer, CRAG corrective fallback status, and complete retrieval trace log immediately.

Category A: In-Index RAG Queries
Category B: Out-of-Index Queries (Live arXiv Fallback)
✦ Hierarchical Indexing
Constructs a recursive tree of paper abstracts and cluster summaries (RAPTOR-style). Surfaces high-level summaries for theme-based queries to provide better contextual synthesis.
✦ Corrective RAG (CRAG)
Grades retrieved context relevance. If it is evaluated as insufficient or out-of-index, the assistant discards it and falls back to a live arXiv search to prevent hallucinations.
✦ Execution Trace Log
Reveals the internal execution path step-by-step. Inspect cosine similarities, CRAG grader evaluations, live arXiv query terms, and the compiled context passed to the LLM.