Andrii Bidochko

• Updated: June 16, 2026
• 6 min read

RAG-Coding: Enhancing LLM Medical Coding with Structured External Knowledge

Direct Answer

RAG‑Coding is an agentic framework that combines multiple large language model (LLM) agents with real‑time retrieval from official ICD‑10‑CM knowledge bases to automate medical coding. By grounding each coding decision in the latest tabular list and coding guidelines, the system lifts LLM accuracy into the range required for clinical compliance and reimbursement.

Background: Why This Problem Is Hard

Medical coding translates free‑text clinical documentation into standardized ICD‑10‑CM codes, a process that directly influences hospital revenue, quality reporting, and public health analytics. The difficulty stems from three intertwined factors:

• Granular taxonomy: The ICD‑10‑CM system contains over 70,000 codes, many differing by a single character that reflects disease severity, laterality, or encounter type.
• Dynamic guidelines: The Centers for Medicare & Medicaid Services (CMS) releases quarterly updates, and the 2025 revision introduced finer‑grained categories that older models were never trained on.
• Contextual nuance: Clinicians write notes in shorthand, with abbreviations, misspellings, and implicit reasoning that require deep domain knowledge to interpret correctly.

Traditional rule‑based coders struggle to keep pace with guideline changes, while pure LLM approaches excel at language understanding but lack a reliable mechanism to verify that a suggested code complies with the official tabular list. This verification gap leads to hallucinated codes, reduced recall, and potential audit failures—issues that have kept many health systems hesitant to adopt fully automated solutions.

What the Researchers Propose

The authors introduce RAG‑Coding, a modular pipeline that orchestrates four specialized LLM agents:

1. Query Generator: Converts a clinical note into concise, searchable queries targeting the ICD‑10‑CM tabular list.
2. Retriever: Pulls the most relevant sections from the official coding manual and accompanying guideline PDFs.
3. Validator Agent: Cross‑checks the candidate codes against the retrieved knowledge, flagging inconsistencies.
4. Aggregator: Synthesizes the final code set, resolves conflicts, and formats the output for downstream billing systems.

Each agent operates on the same LLM backbone (e.g., GPT‑4, Claude, or LLaMA) but is prompted with role‑specific instructions that shape its behavior. The key innovation is the explicit grounding step: before a code is emitted, the system retrieves the exact paragraph from the official source that justifies the choice, ensuring traceability and auditability.

How It Works in Practice

The end‑to‑end workflow can be visualized as a loop of generation and verification:

1. Input ingestion: A discharge summary or progress note is fed into the Query Generator.
2. Knowledge retrieval: The Retriever issues the generated queries to a vector‑search engine built over the ICD‑10‑CM tabular list (e.g., using Chroma DB integration).
3. Cross‑reference: Retrieved passages are passed to the Validator Agent, which asks the LLM to explain why each candidate code matches the passage.
4. Decision aggregation: The Aggregator consolidates validated codes, applies business rules (e.g., primary vs. secondary diagnosis), and produces a final, standards‑compliant output.

What sets RAG‑Coding apart from prior “LLM‑only” pipelines is the enforced feedback loop: the system never finalizes a code without a supporting citation from the official documentation. This design reduces hallucination, improves recall, and provides a transparent audit trail that regulators can inspect.

Evaluation & Results

The researchers benchmarked RAG‑Coding on the newly released MDACE‑2025 dataset, an expert‑reannotated version of the original MDACE corpus aligned with the 2025 ICD‑10‑CM guidelines. Evaluation focused on micro‑ and macro‑averaged F1 scores, as well as precision and recall trade‑offs.

• Baseline comparison: Across multiple LLM backbones, RAG‑Coding outperformed the strongest LLM‑only baseline by 8–13 % in micro‑F1 and 2–8 % in macro‑F1.
• Against PLM‑ICD: The state‑of‑the‑art pretrained language model PLM‑ICD achieved higher micro‑precision (+6 %), but RAG‑Coding delivered a superior micro‑recall (+11 %), resulting in comparable overall F1 scores.
• Ablation insights: Removing the Retriever dropped micro‑F1 by ~7 %, while omitting the Validator reduced precision by 9 %, confirming that each component contributes uniquely to performance.

These results demonstrate that grounding LLM decisions in structured external knowledge not only boosts raw accuracy but also aligns the system with the compliance expectations of health‑care auditors.

Why This Matters for AI Systems and Agents

For AI practitioners building enterprise‑grade agents, RAG‑Coding offers a concrete blueprint for integrating retrieval‑augmented generation (RAG) with domain‑specific compliance requirements. The framework illustrates how to:

• Modularize agent responsibilities: Assigning distinct roles (query, retrieval, validation, aggregation) simplifies debugging and enables independent scaling of each component.
• Guarantee traceability: By attaching source citations to every decision, downstream systems can automatically generate audit logs, a feature rarely addressed in generic LLM pipelines.
• Adapt to evolving standards: Updating the knowledge base (e.g., the ICD‑10‑CM tabular list) instantly propagates to the system without retraining the underlying LLM, reducing maintenance overhead.

Enterprises looking to embed AI into regulated workflows can leverage the same pattern for other taxonomies—such as CPT coding, SNOMED CT, or even financial reporting standards. The approach aligns with the broader trend of agentic AI, where multiple specialized agents collaborate under a supervisory orchestrator to achieve higher reliability.

For teams interested in rapid prototyping, the UBOS platform overview provides a low‑code environment to spin up LLM agents, connect them to vector stores, and define orchestration logic without deep engineering effort.

What Comes Next

While RAG‑Coding marks a significant step forward, several open challenges remain:

• Scalability of retrieval: Real‑time search over the full ICD‑10‑CM corpus can become a bottleneck in high‑throughput settings; future work may explore hierarchical indexing or hybrid dense‑sparse retrieval.
• Multi‑modal evidence: Incorporating imaging reports or lab results could further improve coding accuracy for complex cases.
• Human‑in‑the‑loop workflows: Designing intuitive interfaces that let coders review and override AI suggestions while preserving the audit trail.
• Generalization to other jurisdictions: Extending the methodology to ICD‑11 or country‑specific coding systems will test the framework’s adaptability.

The authors have released the MDACE‑2025 dataset under an open license, inviting the community to benchmark alternative retrieval strategies and to explore cross‑domain applications. As the healthcare industry continues to digitize, the demand for transparent, compliant AI assistants will only grow.

Organizations ready to experiment with agentic AI can join the UBOS partner program to access pre‑built connectors for medical knowledge bases, compliance tooling, and workflow automation studios.

References

RAG‑Coding paper on arXiv

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.