BunnyRAG: graph-aware retrieval and reproducible evaluation workflow 

Synthetic benchmarking and robustness testing 

Approach

I implemented a two-stage retrieval method that combines:

• semantic similarity, and

• a global graph-structure signal (effective resistance / conductance) that encourages selecting nodes spanning multiple graph communities.

Graph communities serve as a proxy for distinct disciplines.

Evaluation design 

To understand behavior under realistic uncertainty, I evaluated BunnyRAG across three synthetic regimes:

1. Seed-community regime: whether seed nodes belong to the same community (representing a single-discipline starting point) versus a more mixed seeding strategy.

2. Edge-weight regime (stochastic): uniformly random edge weights VS stochastic noise of vector space distance VS query-aware edge weights

Results and takeaways

• The global effective-resistance signal provided stronger cross-community coverage than expected.

• Contrary to the initial hypothesis that a semantic punishment is necessary to force “hops” across communities, experiments indicated that rewarding semantic similarity could still preserve strong community coverage (high entropy) across several regimes.