DeepMed Search: An Open-Source Agentic Platform for Medical Deep Research with Introspective Verification

1 Introduction

Large Language Models (LLMs) have revolutionized clinical question answering [Singhal et al., 2025; Tu et al., 2023; Yang et al., 2022], yet their deployment in highstakes evidence-based medicine remains hindered by limited transparency and auditability [Ji et al., 2023; Rudin, 2019]. Medical deep research requires synthesizing heterogeneous data from academic literature, constantly updating clinical guidelines, and structured patient records. While RetrievalAugmented Generation (RAG) is widely adopted to ground model outputs [Lewis et al., 2020; Karpukhin et al., 2020; Izacard and Grave, 2021], standard RAG implementations often rely on static, linear pipelines that are vulnerable to Retrieval-Induced Reasoning Drift, especially in complex, long-tail scenarios such as rare disease diagnosis [Mallen et al., 2023]. In these contexts, standard semantic retrievers suffer from commonality bias [Kandpal et al., 2023], preferentially surfacing generic evidence that supports common mimic conditions and causing the LLM to be hijacked by plausible but non-discriminative distractors. Although commercial AI research tools attempt to address this issue, their black-box nature prevents clinicians from auditing the causal logic behind retrieved evidence, which remains a nonnegotiable requirement in clinical workflows.

∗Corresponding Author.

To democratize trustworthy and auditable AI in healthcare, we introduce DEEPMED SEARCH, a fully opensource, agentic platform engineered for transparent medical deep research. Moving beyond simple RAG frameworks, DEEPMED SEARCH introduces a source-adaptive router that autonomously orchestrates sub-queries across heterogeneous endpoints, including PubMed APIs, live web search, and a local Neo4j-based knowledge graph. Furthermore, to ensure the causal alignment and faithfulness of the retrieved evidence, the platform features an introspective verification module [Kiciman et al., 2024]. This engine employs contrastive hypothesis expansion to filter confounding noise and utilizes a multi-agent adversarial debate framework [Du et al., 2024; Yao et al., 2022] to rigorously stress-test reasoning paths before final synthesis.

In this paper, we showcase the system's glass-box interactivity, while clinicians can observe DEEPMED SEARCH as it autonomously decomposes complex queries, routes searches, and conducts real-time multi-agent deliberations. Ultimately, the system generates structured, citation-backed research reports while maintaining complete interpretability [Xiong et al., 2024].

2 SystemArchitecture

DEEPMED SEARCH is engineered as a modular, agentic workflow designed to transform high-level clinical inquiries into structured research insights. As illustrated in Figure 1, the platform transitions from autonomous data orchestration to rigorous causal verification across three primary stages: Source-Adaptive Routing, Causal Introspective Filtering, and Multi-Agent Synthesis.

2.1 Source-Adaptive Routing

A core pillar of DEEPMED SEARCH is its ability to navigate heterogeneous data landscapes. Unlike monolithic RAG systems, our platform incorporates a Source-Adaptive Router that treats the information retrieval process as a sequential decision-making task.

Upon receiving a clinical query, the system first performs Atomic Decomposition, breaking the input into distinct clinical constraints (e.g., symptomatic triad, temporal progression, and treatment resistance). The Router then dynamically evaluates the information density of each sub-query to dispatch them to optimal endpoints:

• Scientific Evidence: Direct API integration with PubMed for fetching peer-reviewed literature and clinical trial abstracts [Qiu et al., 2025].
• Pathophysiological Logic: Querying a local NEO4J knowledge graph (integrating UMLS and SNOMEDCT) to identify multi-hop causal links between symptoms [Pearl, 2009].
• Real-time Guidelines: Utilizing a Web Search crawler to retrieve the latest consensus statements and medical news.

2.2 Causal Introspective Filtering

To mitigate the Commonality Bias identified in clinical LLMs [Kandpal et al., 2023], the platform utilizes a Causal Introspective Filtering mechanism. This stage aligns the retrieved evidence with diagnostic logic via two steps: 1) Contrastive Hypothesis Expansion: The system proactively formulates "confounding hypotheses", prevalent diseases that mimic the target long-tail condition (e.g., generating "Essential Hypertension" when analyzing potential "Pheochromocytoma") [Park and Lee, 2024]. 2) Evidence Downweighting: Chunks are filtered based on a discriminative score Si = sim(ci,Htarget) sim(ci,Hmimic)+ϵ . Here, sim(·, ·) denotes the cosine similarity between BGE-M3 sentence embeddings, and ϵ

is set to 10-6 to avoid division by zero. A candidate chunk is regarded as discriminative when Si is larger than the threshold τs; chunks failing to meet a dynamic threshold are tagged as "non-discriminative noise" and visually dimmed in the interface, ensuring the model's attention remains on the longtail signals.

2.3 Multi-Agent Synthesis

The final reasoning phase is handled by an introspective verification module, instantiated as a multi-agent adversarial debate framework [Chan et al., 2023]. The system deploys three distinct agents: a Proponent (Apro) that constructs reasoning chains from evidence to diagnosis, an Opponent (Acon) that identifies missing causal links or inconsistencies, and a Judge (Ajudge) that delivers the final verdict once reasoning converges (Information Gain < τ ).

Crucially, rather than producing a single-line answer, the synthesis engine generates a structured research report. These reports are citation-backed and organized into sections: Differential Diagnosis, Supporting Evidence (from both Graph and Text), and Rejected Hypotheses. This "glass-box" output allows clinicians to audit the platform's logic in real-time.

3 Implementation Details

The DEEPMED SEARCH backend is engineered for highthroughput concurrent processing to support real-time clinical deep research workflows. Frontend & Orchestration: Built on NEXT.JS, the frontend integrates a WebSocket-driven UI for streaming real-time agent deliberations. A dedicated orchestration layer handles the concurrent dispatch of sub-queries via the Source-Adaptive Router. Knowledge Engines: Structured multi-hop reasoning is powered by a NEO4J graph database integrating UMLS and SNOMED-CT. For unstructured evidence, we use MILVUS to index the MIRAGE corpus (Textbooks and PubMed) [Xiong et al., 2024], utilizing BGE-M3 for dense vector similarity coupled with a lightweight cross-encoder for precise re-ranking. LLMServing: To meet the computational demands of deploying three concurrent agents (Proponent, Opponent, Judge), we serve models like QWEN2.5-14B and DEEPSEEK-V3 via VLLM,

Query: "Patient presents with paroxysmal hypertension, headache, palpitations, and sweating. Standard anti-hypertensive drugs ineffective."

× Vanilla RAG Response:

Diagnosis: Essential Hypertension. Treatment involves lifestyle changes...

[Error: Fixates on common disease; ignores paroxysmal & sweating cues]

✓DEEPMED SEARCH (Ours):

Diagnosis: Pheochromocytoma.

Reasoning: 1. Contrastive Filter: 'Essential Hypertension' downweighted (cannot explain sweating). 2. Debate Verdict: Symptom triad strongly indicates catecholamine-secreting tumor.

Table 1: Case study: DEEPMED SEARCH actively filters common disease noise (Hypertension) to correctly identify the rare condition.

leveraging PagedAttention to optimize memory footprint and maximize throughput. Report Generation Engine: A custom synthesis module compiles the verified evidence into a structured markdown/PDF report, automatically mapping reasoning steps to specific reference nodes (PubMed IDs or Graph edges) to ensure full traceability.

4 Demonstration Scenarios

We present using real-world rare disease cases from the MedR-Bench dataset, as exemplified in Table 1. The web interface allows clinicians to input complex case reports and interactively audit the deep research process.

4.1 Scenario 1: Autonomous Routing &Causal Filtering

Consider a patient presenting with episodic hypertension, headaches, and palpitations. A standard RAG system typically retrieves generic guidelines for "Primary Hypertension", leading to reasoning drift. In DEEPMED SEARCH: 1) Routing in Action: The Router identifies the temporal cue ("episodic") and dispatches parallel queries to Neo4j and PubMed. 2) Noise Filtering: The "Evidence Panel" marks generic hypertension evidence with a Low Relevance tag (visualized as dimmed text), explicitly stating it was downweighted due to overlap with common confounders. Conversely, evidence pointing to Pheochromocytoma is dynamically highlighted.

4.2 Scenario 2: Glass-Box Debate &Report

For complex diagnoses, the clinician can activate the introspective debate mode. 1) Real-Time Deliberation: As illustrated in Figure 2, a streaming panel opens where Apro argues for Pheochromocytoma citing paroxysmal symptoms, while Acon challenges the evidence density. 2) Human-inthe-Loop: Crucially, the clinician can inject missing parameters (e.g., "Add lab result: elevated metanephrines") directly into the debate stream. Structured Output: The Ajudge finalizes the verdict, and the system instantly compiles a deep research report, linking the diagnosis to verified graph nodes.

5 Evaluation

As summarized in Table 2, DEEPMED uniquely integrates source-adaptive routing and multi-agent verification, bridging the gap between vanilla RAG and commercial black-box tools.

Performance Gains: We evaluated on the MIRAGE benchmark [Xiong et al., 2024]. DeepMed Search consistently outperforms standard RAG baselines. On the challenging PubMedQA subset, our method improves accuracy by approximately 9.0% over vanilla RAG (using Qwen3-8B). For rare diseases, experiments on the MedR-Bench dataset show that our approach achieves an accuracy of 76.43% (using DeepSeek-v3.2), bridging the gap between automated systems and expert-level reasoning.

6 Conclusion

DEEPMED SEARCH provides a transparent, agentic platform for medical deep research. It combines source-adaptive routing, causal-consistent filtering, and multi-agent verification to mitigate retrieval-induced reasoning drift in medical LLMs [Park and Lee, 2024] and generate structured, citation-backed reports. Current limitations include additional computational and latency overhead and the lack of a formal clinician user study. Future work will improve efficiency, involve clinicians in evaluation, and extend the framework to multimodal data such as medical imaging [Wang et al., 2025].

Acknowledgments

This work was supported by the National Natural Science Foundation of China under Grant No. 62306173.