The Problem with Vector RAG in High-Stakes Healthcare

In clinical pharmacology, precision is non-negotiable. When healthcare providers evaluate patient regimens, failing to catch a multi-hop metabolic interaction or an additive anticholinergic burden can lead to acute toxicity, severe cognitive impairment, or preventable hospitalizations.

When developers build medical AI assistants using standard Vector RAG (Retrieval-Augmented Generation), they typically chunk unstructured medical guidelines and search them via cosine similarity. But Vector RAG breaks down completely under multi-hop reasoning. If an AI agent has to calculate multi-step CYP450 enzyme bottlenecks, parse proprietary multi-ingredient combination brands, or flag look-alike sound-alike (LASA) transcription typos, flat-text chunk retrieval either misses the connection entirely or hallucinates plausible-sounding nonsense.

To solve this, we built Polypharmacy Guardian—a deterministic, graph-grounded clinical triage agent powered by FalkorDB.

Why FalkorDB? (Architectural Overview)

Pharmacological data is inherently network-driven: drugs, metabolic enzymes, adverse effects, and validated clinical scoring scales exist as interconnected relationships.

We used FalkorDB as our high-speed graph database backbone because it allows us to execute complex multi-hop graph traversals in under 70 milliseconds.

Core Architectural Layers:

1. Graph Ingestion & Entity Modeling (app/ingestion.py, drugs_seed.csv):

   We ingested over 46 verified drug ingredients (:Ingredient) categorized by therapeutic classes and mapped with explicit edges such as [:INHIBITS], [:SUBSTRATE_OF], and [:HAS_ACB] (Anticholinergic Burden). Every edge is backed by rigorous literature provenance (e.g., Flockhart CYP450 Table 2024, Boustani ACB Scale 2008).

2. Indian Fixed-Dose Combination (FDC) Decomposition (:BrandFDC):

   Proprietary combination brands (such as Solvin Cold, Ultracet, Ecosprin AV, and Combiflam) are modeled as discrete :BrandFDC nodes. When a clinician inputs a brand name, FalkorDB instantly traverses [:CONTAINS_INGREDIENT] edges to decompose it into its active constituents (Solvin Cold \(\rightarrow\) Paracetamol, Chlorpheniramine, Phenylephrine), exposing hidden anticholinergic stacking that flat search lists miss.

3. Parameterized Cypher Agent Loop (app/agent.py, cypher_tools.yaml):

   Our 3-step AI agent translates unstructured clinical notes into optimized Cypher queries executed directly against FalkorDB, ensuring 100% deterministic math and zero LLM hallucination during safety checks.

4. Pre-Execution Safety Guardrails:

   • LASA Near-Misses: Transcription typos (e.g., Hydralizine) trigger an AMBIGUOUS warning halt using official ISMP Tall-Man lettering standards (hydrALAZINE), strictly excluding ambiguous guesses from active calculations.

   • Zero-Trust Unknown Coverage: Unindexed entities (e.g., Lecanemab) return UNKNOWN_COVERAGE and generate a review ticket without hiding active co-prescribed conflicts among known drugs.

5. Closed-Loop Pharmacist State Mutation:

   Verdicts persist as idempotent pharmacist review tickets (TICK-XXXXXXXX) inside FalkorDB. Clinical teams can click Approve Swap or Override in the UI, triggering live state mutations directly on the graph nodes.

Architectural Design Comparison

To evaluate how graph-backed RAG compares against standard unstructured retrieval in clinical settings, we measured performance across key clinical tasks:

Rigorous Verification & Testing

To ensure absolute safety, our test suite (test_api.py) runs 14 strict clinical assertions—covering negative controls, unknown entities, prodrug bioactivation failures (e.g., Codeine + Fluoxetine), and state mutations—passing successfully with Exit Code 0.

Links & Resources

• GitHub Repository: https://github.com/debugwsded316/Graphs-hacks

• Demo Video: https://youtu.be/IMQOEou6MpA