Live RAG Playground

Agentic RAG Evidence Hunt Playground

Load a small corpus, fire a natural-language query, and watch a real multi-hop agentic loop out-hunt a one-shot baseline on a live Cytoscape evidence graph. The comparison runs on actual fuse.js retrieval — not a fake score.

Enter run hunt
Corpus Size 0docs — no corpus loaded —
One-Shot Evidence 0nodes awaiting run
Agentic Hops 0hops awaiting run
Confidence Lift 0.00 agentic − one-shot
▍ Corpus empty
▍ Query type a natural-language question
presets:
One-Shot single fuse.js search · no loop
hops 1 evidence 0 conf
load a corpus · run a hunt
graph will appear here
query doc
No evidence yet — run a hunt to populate.
Agentic search · read · score · refine · re-search
hops 0 evidence 0 conf
load a corpus · run a hunt
graph grows hop by hop
query doc active hop
No evidence yet — run a hunt to populate.

How it hunts

Each hop runs fuse.js over the corpus, scores the top-K, then refines the query by extracting the highest-signal terms from the retrieved snippets. Loops while composite confidence is below 0.55 or until 5 hops are spent.

Composite confidence

conf = Σ (avg_hop_score × |unique_docs_hop|) ÷ |unique_docs_total|. Hops that add no new evidence stop contributing weight. Both panes use the same retriever — only the loop policy differs.

What this proves

The one-shot pane answers part of the question. The agentic pane stitches cause → mitigation → impact into a connected evidence chain. Same retriever, same data, different search policy.