Why We Built Grounded Retrieval

The Problem Every Oncologist Knows

Dr. Gary Takahashi is a practicing medical oncologist. He treats patients with lung cancer, hepatocellular carcinoma, melanoma, lymphoma, and cholangiocarcinoma — each governed by complex treatment algorithms that change throughout the year.

ESMO Clinical Practice Guidelines update on a rolling basis and provide a structured oncology reference set across major cancer types. Each update can introduce new first-line therapies, remove options based on safety signals, or restructure entire treatment algorithms based on new trial data.

The tools available to stay current are inadequate. Manual guideline review is time-consuming. Conference abstracts trail real-world approvals. And large language models — despite their promise — fail in ways that are particularly dangerous in clinical oncology.

What Frontier Models Get Wrong

Dr. Takahashi built Grounded Retrieval because he tested the frontier models on real clinical questions and found their answers unacceptable for patient care.

In our study, we posed six clinical questions to five foundation models: GPT-5.4, Claude 4.6, Gemini 3.1 Pro, DeepSeek-V3.2, and Perplexity.¹ The results were consistent:

Every model missed zongertinib as first-line therapy for ERBB2-mutant NSCLC — a drug that received FDA accelerated approval on February 26, 2026, with a 76% objective response rate.² When tested 25 days later, none of the five frontier models included it in their recommendations.
Every model missed sevabertinib for FGFR2-positive cholangiocarcinoma, approved November 19, 2025.
Perplexity recommended camrelizumab + rivoceranib for hepatocellular carcinoma — a combination the FDA rejected twice with complete response letters.
Claude recommended tazemetostat for follicular lymphoma — a drug voluntarily withdrawn on March 9, 2026 after a safety signal for myelodysplastic syndrome and acute leukemia.
DeepSeek recommended sintilimab — an anti-PD-1 antibody available only in China, not approved or available in the United States.

These aren't edge cases. These are routine clinical questions that oncologists face daily.

The Architecture: Retrieval Over Recall

Grounded Retrieval was designed around a core insight: the LLM should extract meaning from the query, not recall facts about treatment. Clinical facts belong in authoritative sources. The LLM's role is semantic understanding — parsing what the clinician is actually asking.

┌─────────────────────────────────────────────────────────┐
│                    Clinical Query                         │
└──────────────────────────┬──────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────┐
│           Semantic Extraction (GPT-4.1)                   │
│   Cancer type · Biomarkers · Treatment line · Context    │
└──────────────────────────┬──────────────────────────────┘
                           │
              ┌────────────┴────────────┐
              ▼                         ▼
┌──────────────────────┐   ┌──────────────────────┐
│  ESMO TOC Navigation │   │  Internet Search     │
│  Section ID + Page   │   │  (DuckDuckGo)        │
│  Content Extraction  │   │  Late-breaking       │
└──────────┬───────────┘   └──────────┬───────────┘
           │                          │
           └────────────┬─────────────┘
                        ▼
┌─────────────────────────────────────────────────────────┐
│         Synthesis with Page-Level Citations               │
└─────────────────────────────────────────────────────────┘

The system uses a Thread Pool Executor to run ESMO guideline navigation and internet search in parallel. ESMO navigation isn't vector similarity — it's structured traversal through the table of contents, identifying the correct guideline, locating the relevant section, and extracting content with specific page numbers.

Why a Mid-Tier Model Wins

A common assumption in AI is that bigger models produce better results. Our findings challenge this directly: GPT-4.1 — a mid-tier model by 2026 standards — with structured retrieval architecture outperformed every frontier model on clinical accuracy.

The reason is architectural. Frontier models fail on clinical questions not because they lack capability, but because they:

Rely on training data that becomes stale within weeks of guideline updates
Cannot verify currency — they have no mechanism to know whether a drug has been withdrawn or a new therapy approved
Hallucinate with clinical plausibility — their recommendations sound authoritative but may reference drugs unavailable in the U.S. or rejected by regulators

Grounded Retrieval sidesteps these failure modes entirely. The LLM never needs to "know" the current standard of care — it only needs to understand what the clinician is asking, then retrieve the answer from authoritative sources.

What This Means for Clinical Practice

The zongertinib case illustrates the stakes. On February 26, 2026, the FDA granted accelerated approval to zongertinib for adult patients with unresectable or metastatic NSCLC with activating ERBB2 (HER2) mutations.² The approval was based on HERIZON-Lung02, demonstrating a 76% objective response rate.

Twenty-five days later, no frontier model included this as a first-line option. For the patient sitting across from an oncologist in March 2026, the gap between "AI-assisted" care and evidence-based care is not theoretical — it's the difference between receiving an effective targeted therapy and receiving an outdated recommendation.

Grounded Retrieval was built to close this gap — not by building a better model, but by building a better system around the model.

References

Takahashi G, Mengistu E, Tarshizi E, Van Benschoten A. Building a Large Language Model Assistant to Maintain Currency with Medical Oncology Guidelines. Under review, npj Digital Medicine.
FDA grants accelerated approval to zongertinib for non-small cell lung cancer with an activating ERBB2 (HER2) mutation. FDA.gov. February 26, 2026.
ESMO Clinical Practice Guidelines. ESMO.org.