Failure is Feedback · Worked Example

How a failed hop becomes re-planning feedback

A multi-hop retrieval agent walks a document graph one hop at a time. When a hop lands on the wrong component, the failure isn't discarded — it's written to the action history and fed back into PLAN, which rewrites the sub-query and retries. Step through the run below.

❓

How many words is the writing behind the statue of the president who served with Vice President Hannibal Hamlin?

ANSWER24

1 / 14

Step 0 · PLANPLAN runs before any traversal — the query is decomposed into sub-queries.

Step 0 · PLANSub-query list is ready; no anchor has been chosen yet (𝒟ₐₙ꜀ = ∅).

Step 1 · TRAVERSE q₁TRAVERSE q₁ — a global hop with no anchor yet.

Step 1 · TRAVERSE q₁q₁ resolved → “Abraham Lincoln”.

Step 2 · TRAVERSE q₂TRAVERSE q₂ from anchor {Hamlin} — hop along the navigational edge.

Step 2 · TRAVERSE q₂✗ Wrong component — vector-only matched a surface similarity, not the answer.

Step 3 · TRAVERSE q₂ (retry)TRAVERSE q₂ again — the anchor stays {Lincoln}.

Step 3 · TRAVERSE q₂ (retry)✗ Wrong again — the cabinet table says nothing about the statue.

Step 4 · PLAN (re-plan)PLAN reads the failure trace (h₂, h₃): both hops failed.

Step 4 · PLAN (re-plan)q₂ → q₂′ — the sub-query is rewritten, using the failures as feedback.

✦ IntermissionThe Traverser: each hop is parameterized by τ = (hop scope, vector granularity, LLM reasoning).

Step 5 · TRAVERSE q₂′Backtrack: the anchor resets to {Hamlin}, then hops the other edge to the Memorial.

Step 5 · TRAVERSE q₂′✓ q₂′ resolved → the Lincoln Memorial.

Step 6 · STOPAll sub-queries solved — the Reranker assembles the final ranked components 𝓒_R.

Navigate with ← →, the buttons, the dots, or swipe on touch · 14 frames