# Grounded by Experience: Gen Healthcare Prediction Augmented with Hierarchical Agentic Retrieval

Source: https://www.youtube.com/watch?v=cmo1jU1cuVU
Recap page: https://rapidrecap.app/video/cmo1jU1cuVU
Generated: 2025-11-21T20:05:05.247+00:00

---
## Quick Overview

The Hierarchical Agentic Retrieval (HAR) system outperforms traditional Large Language Models (LLMs) in medical reasoning tasks, such as predicting hospital readmission, by employing two specialized agents—Agent Top (the primary physician) and Agent Low (the consultant)—that work collaboratively to ensure answers are grounded in evidence and contextually relevant, thereby avoiding hallucinations and improving accuracy across rare disease predictions.

**Key Points:**
- The research tackles the challenge of making LLMs trustworthy in high-stakes areas like healthcare prediction.
- The proposed solution is the Hierarchical Agentic Retrieval (HAR) system, which uses two agents: Agent Top (primary decision-maker) and Agent Low (consultant/retriever).
- HAR solves the problem of LLMs hallucinating by forcing the agents to collaborate and ground responses in evidence, unlike models that rely solely on internal parameters.
- In tests predicting hospital readmission (a 2-day prediction), HAR scored 9/10 correct, significantly outperforming baseline models that scored 3/10.
- Agent Top is rewarded for accuracy, while Agent Low is rewarded for quality and efficiency (fewer steps/less latency) in retrieving relevant sub-graphs of knowledge.
- This coordinated approach, leveraging both reasoning and retrieval, ensures the final answer is both accurate and ethically sound, avoiding bias toward common diseases.
- The system's success is demonstrated by its ability to handle rare conditions where large datasets often fail older, non-hierarchical models.

![Screenshot at 0:09: The hosts introduce the core problem: how to make large language models \(LLMs\) truly trustworthy when stakes are life and death, leading into the discussion of the research paper.](https://ss.rapidrecap.app/screens/cmo1jU1cuVU/00-00-09.png)

**Context:** The video discusses a research paper presenting the Hierarchical Agentic Retrieval (HAR) system, designed to improve the trustworthiness and accuracy of Large Language Models (LLMs) when applied to critical tasks like healthcare predictions. The core issue addressed is the tendency of standard LLMs to 'hallucinate' or produce plausible but incorrect answers, especially when dealing with rare data points, which is unacceptable when patient lives are at stake.

## Detailed Analysis

The video details a research paper introducing the Hierarchical Agentic Retrieval (HAR) system to improve the reliability of LLMs in high-stakes medical prediction. The primary challenge is that current LLMs, while excellent predictors generally, often fail when dealing with rare events or lack the ability to cite verifiable sources, leading to potentially dangerous hallucinations. HAR overcomes this by structuring the AI system into two cooperating agents: Agent Top, who acts as the primary physician and decision-maker, and Agent Low, who acts as the consultant, specializing in information retrieval. Agent Top asks a question, and Agent Low searches the knowledge base (like looking up a specific chapter in a medical textbook) and returns relevant information. Crucially, the agents are trained with different reward structures: Agent Top is rewarded for correct clinical outcomes, and Agent Low is rewarded for efficient, high-quality retrieval paths. The researchers found that this system performed exceptionally well in predicting hospital readmission, achieving 9 out of 10 correct predictions, compared to baseline models only achieving 3 out of 10. This success stems from the system's ability to use evidence-based reasoning and avoid superficial pattern matching, ensuring that even when dealing with rare diseases, the resulting answer is accurate, explainable, and ethically sound, as the agents are incentivized to collaborate toward a single goal rather than acting in silos.

### The Core Problem

- Making LLMs Trustworthy in Medicine: High stakes (life and death)
- LLMs hallucinate and lack verifiable accuracy
- Need for explainable reasoning.

### The HAR Solution

- Dual-Agent Architecture: Agent Top (Primary Physician/Decision Maker)
- Agent Low (Consultant/Information Retriever)
- Agents trained with different rewards.

### Agent Reward Structures

- Agent Top rewarded for correct clinical outcomes
- Agent Low rewarded for efficient retrieval paths and quality of retrieved sub-graphs.

### Performance Results

- HAR achieved 9/10 accuracy on 2-day hospital readmission prediction
- Baseline models scored only 3/10
- HAR successfully handles rare disease prediction.

### Mechanism of Improvement

- HAR avoids isolated optimization by forcing collaboration and using evidence-based reasoning (semantic similarity) rather than just keyword matching.

### Conclusion

- The HAR framework allows AI to perform human-like, structured problem-solving, leading to accurate and ethically sound medical decision support.

![Screenshot at 0:09: The hosts introduce the core problem: how to make large language models \(LLMs\) truly trustworthy when stakes are life and death, leading into the discussion of the research paper.](https://ss.rapidrecap.app/screens/cmo1jU1cuVU/00-00-09.png)
![Screenshot at 0:48: The term G.H.A.R. \(Hierarchical Agentic Retrieval\) is introduced on screen, representing the proposed solution.](https://ss.rapidrecap.app/screens/cmo1jU1cuVU/00-00-48.png)
![Screenshot at 1:16: Visual representation of the two main approaches: the older, less effective models stuck between a rock and a hard place.](https://ss.rapidrecap.app/screens/cmo1jU1cuVU/00-01-16.png)
![Screenshot at 3:06: A visual analogy is made where the system is described as slow, introducing latency and noise, contrasting with the HAR approach.](https://ss.rapidrecap.app/screens/cmo1jU1cuVU/00-03-06.png)
![Screenshot at 4:16: The dual-agent system is explained, comparing Agent Top \(primary physician\) and Agent Low \(consulting specialist\).](https://ss.rapidrecap.app/screens/cmo1jU1cuVU/00-04-16.png)