Traversal explorer — agentic vs RAG, side by side

Pick a question from the dropdown. The left panel shows the agentic agent’s traversal over the wiki graph — nodes light up in call order (numbered badges) with ancestor chains highlighted. The right panel shows RAG’s retrieval concentration — paper nodes are sized by how many chunks were pulled from that paper and tinted by their max cosine similarity.

Both panels render the same wiki graph, so you see at a glance what each method actually looked at and where their footprints agree or diverge.

Loading…

Agentic — graph traversal

RAG — retrieval concentration

Reading the panels

Left (agentic). Every node with a red ring was touched by a tool call. The small number next to each ring is the call’s position in the sequence (1 = first). Red tree edges show the ancestor chain from the wiki root to each visited page. If the agent called tools in parallel (issued multiple read_page calls in one model turn), their badges will be consecutive integers even though they executed at roughly the same wall-clock time.

Right (RAG). Every blue-ringed node is a paper RAG pulled chunks from. Node size scales with chunk count (more chunks → bigger circle); stroke opacity scales with the top chunk’s cosine similarity (stronger match → solider ring). The ×N label is the chunk count. RAG sees documents only through their embedded chunks, so the right panel never touches index.md or section pages — only leaves in papers/.

What to look for

  • Overlap: do both panels land on the same papers? If yes, both strategies “see” the same evidence and the difference is only in how they assemble it. If no — one found evidence the other missed.
  • Concentration vs breadth: a single fat blue node (RAG) means the embedder latched onto one paper; many small ones mean the question lives across many papers. The same contrast shows up on the agentic side as a long call chain vs a broad fan of parallel reads.
  • Off-target retrieval: RAG nodes with high chunk counts but not in the gold PMID list hint at semantic near-misses — a sign you might need chunk filtering or query rewriting.