GRASP: Plan-Guided Graph Retrieval with Adaptive Fusion and Reranking on Semi-Structured Knowledge Bases

summary

The gist

The gist: GRASP is a three-stage SKB retrieval framework that unifies plan-based graph retrieval, plan-conditioned fusion with a dense retriever, and a fine-tuned reranker over the fused candidates.

In short

GRASP is a three-stage retrieval framework for semi-structured knowledge bases that combines graph planning with dense text retrieval. It uses an LLM to create a graph plan, executes it against a database, fuses the results with a text retriever based on the plan's risk level, and finally uses a fine-tuned reranker to reorder candidates. This hybrid approach significantly improves retrieval accuracy across benchmarks.

Key concepts

Graph Plan Generation: An LLM is used to translate a natural language query into an executable graph plan. This plan defines the necessary steps, such as which nodes to link and which relations to follow, providing a structured path for the system to search the knowledge base.
Plan-Conditioned Fusion: This stage combines two different rankings: one from the graph retrieval and one from a dense text retriever. The fusion method adapts its weighting dynamically based on the query's complexity (risk level) reported by the planner, ensuring both aspects of the data are considered optimally.
Fine-tuned Reranker: A specialized model is trained to reorder a smaller set of top candidates. This reranker analyzes each candidate using a multi-turn chat involving system instructions and the candidate's context, refining the final ranking for higher precision.
Dynamic RRF Weighting: Instead of using a fixed weight for fusing graph and text scores, GRASP uses Dynamic Reciprocal Rank Fusion (RRF). This method calculates a query-specific weight based on the number of candidates found and the risk level determined by the planner, allowing for adaptive fusion.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Sample Complexity — explained in 18 episodes
Federated Learning — explained in 17 episodes
Quantum Fisher Information (QFI) — explained in 16 episodes
Retrieval-Augmented Generation (RAG) — explained in 16 episodes
Interpretability — explained in 15 episodes

This episode discusses

The paper

GRASP: Plan-Guided Graph Retrieval with Adaptive Fusion and Reranking on Semi-Structured Knowledge Bases · Read on arXiv

University of Michigan

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "GRASP: Plan-Guided Graph Retrieval with Adaptive Fusion and Reranking on Semi-Structured Knowledge Bases".

Jane: The gist: GRASP is a three-stage SKB retrieval framework that unifies plan-based graph retrieval, plan-conditioned fusion with a dense retriever, and a fine-tuned reranker over the fused candidates.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So to wrap up, "GRASP: Plan-Guided Graph Retrieval with Adaptive Fusion and Reranking on Semi-Structured Knowledge Bases" is this three-stage retrieval framework that converts the query into a structured plan, executes it against a graph database with dense rescoring, fuses with a separately trained multi-field retriever via plan-conditioned RRF, and reranks the fused top candidates with a fine-tuned reranker >

Lu: It sets a new state of the art on every metric across three STaRK datasets, lifting average Hit@one from sixty-two point zero to seventy-three point nine > <ref:2605.30237#pg1,lifting average Hit@1 from 62.0 to 73.9>

Meng: The paper shows that the graph and dense retrieval branches are complementary and that the fused candidate pool supplies high-quality hard negatives that a downstream reranker can exploit to push top-rank precision substantially higher >

Jane: This means for someone who only listens to the show, this paper shows that combining structural information with semantic text information, when done intelligently through plan conditioning, is much more effective than using either branch in isolation >

Tom: The headline performance of GRASP is stable across reasonable choices of hyperparameters, making it easy to deploy without per-dataset hyperparameter expertise >

Lalam: It's a three-stage retrieval framework that converts the query into a structured plan, executes it against a graph database with dense rescoring, fuses with a separately trained multi-field retriever via plan-conditioned RRF, and reranks the fused top candidates with a fine-tuned reranker >

Conclusion: Tom: So, we’ve been talking about GRASP, this new framework for searching messy knowledge bases that combines graph structures and text descriptions >

Jane: Right, Tom, so the authors are focusing on how to actually get both of those things working together without just throwing them into a blender >

Lu: It’s about this three-stage pipeline where you first get an AI planner to make a graph plan, then you fuse that with a dense retriever using some dynamic weighting based on the planner’s risk assessment >

Meng: And then they fine-tune a reranker on the top candidates from that fusion pool so you don't just get noisy results from the initial search >

Tom: Exactly, so it’s not just one big model doing everything, it’s a clever sequence of steps where each part handles a specific kind of information and then refines what came before >

Jane: I mean, what this really means for someone listening is that when you ask a complex question about something structured like a knowledge base, you get results that aren't just random text matches or just rigid graph paths >

Lu: The big idea here is showing that these two approaches—the precise path following and the broad semantic understanding—are actually complementary and when you fuse them smartly, you get better recall >

Tom: It’s about proving that the graph part gives you precision on connections while the text part gives you breadth on meaning, and GRASP manages that trade-off >

Jane: And it suggests that we can build smarter AI retrieval systems by having a planner guide the whole process instead of just relying on one method to do all the heavy lifting >

Lu: It’s a solid framework because it shows how you can dynamically adjust your search strategy based on what kind of information you need, like using that risk level to weight the graph vs. text retrieval >

Tom: So, it’s a method that lets you deploy something relatively stable without needing deep expertise in every single dataset where you use it >

Jane: It really points toward systems that can adapt their search strategy on the fly depending on the nature of the query and how much risk is involved >

Lu: Next up, we're going to look at some of those specific results across those three benchmarks and see exactly how much better this is compared to what was out there before >

More episodes

← Home