NutriIntel: Architecture Analysis of an Agentic RAG System for Clinical Pediatric Nutrition

NutriIntel is an open-source Agentic RAG system designed specifically for clinical pediatric nutrition consultation. This article deeply analyzes its hybrid retrieval architecture, deterministic treatment engine, multi-workflow orchestration mechanism, and how to achieve safe and reliable AI-assisted decision-making in medical scenarios. The project is maintained by luch91, with source code hosted on GitHub (link: https://github.com/luch91/NutriIntel-Agentic-RAG-Clinical-Pediatric-Nutrition-Assistant), and was released on May 23, 2026.

In the field of medical AI, LLM applications are deepening into professional vertical scenarios, but medical scenarios have strict requirements for AI systems (accurate retrieval, determinism, traceability, clinical safety). In the pediatric nutrition field, factors such as patient age, weight, diagnosis, etc., affect nutrition plans, and simple Q&A-style AI is difficult to meet the needs. NutriIntel was born in this context, balancing safety and practicality through hybrid retrieval, deterministic engines, and structured dialogue management.

NutriIntel adopts a front-end and back-end separation architecture: the back-end is based on Python 3.12 and FastAPI, and the front-end uses Next.js14 + TypeScript. Core process: User message → Intent classifier (sentence-transformers + LogisticRegression, F1 score reaches 98.6%) → Route to different workflows (treatment, recommendation, comparison, general). Treatment queries initiate a multi-agent workflow: Slot-filling Agent collects clinical parameters → Therapy Gatekeeper reviews → Deterministic Engine generates plans (strictly based on authoritative data sources).

NutriIntel's retrieval system combines vector retrieval and BM25 keyword retrieval:

• Vector retrieval: Qdrant vector database, sentence-transformers/all-MiniLM-L6-v2 embedding model (384 dimensions);
• BM25 retrieval: Supplementary for exact matching of medical terms; Retrieval process: Query rewriting → Vector search (top20) + BM25 search (top20) → Merge and deduplicate → Conditional filtering → Reorder → Return top7 fragments. Knowledge base sources: 9 authoritative clinical nutrition textbooks (e.g., Clinical Pediatric Dietetics) + food composition tables (FCT) from multiple African countries, supporting cross-regional needs.

To solve the 'hallucination' problem of medical AI, NutriIntel designs a deterministic treatment engine:

• Core principle: Nutritional target values are based on authoritative data sources (DRI tables, clinical protocols), not generated by LLM;
• Calculation factors: Patient clinical parameters (age, gender, etc.) + drug-nutrient interactions + disease-specific needs;
• Interpretability: Each recommendation can trace its source;
• Supported diseases: Type 1 diabetes, cystic fibrosis, food allergies, premature infant nutrition, and other pediatric diseases.

NutriIntel implements refined dialogue state management (IDLE → SLOT_FILLING → DISPATCHING → RESPONDING):

• SLOT_FILLING phase: Directly process bare values (e.g., '10 years old') as slot answers without misunderstanding as new queries;
• State persistence: Redis (memory fallback) supports multi-session concurrency; Multi-workflows: Treatment (complex multi-agent), recommendation (hybrid retrieval), comparison (entity extraction + dual retrieval), general (open-domain questions), flexibly responding to different queries.

The project has a clear code structure (modules like agents, api, classification, etc.), including 217 test cases (unit, integration, end-to-end). Observability: Structured logging (structlog) + Prometheus metrics. PDF processing: LangChain PyPDFLoader, with PyMuPDF + Tesseract OCR fallback for scanned versions. Data ingestion: Automatically executed on the first request, indexes 34 knowledge base PDFs, taking 2-3 minutes.

Limitations: 1. Mainly optimized for African regional food composition tables; other regions require additional data integration; 2. Dependence on Qdrant and Redis increases deployment complexity. Improvement Directions: Incorporate 'soft factors' such as patients' socioeconomic factors and eating habits. Summary: NutriIntel proves that Agentic RAG + deterministic engine can meet the safety and interpretability requirements of medical scenarios. Its designs such as hybrid retrieval, state management, and multi-workflows are worth learning for medical AI developers and can be extended to other high-reliability AI-assisted decision-making fields.