Methodology

1. Source Inclusion Criteria

Every source in the Priolix database is ingested from one of these authoritative providers:

• PubMed / MEDLINE — peer-reviewed biomedical literature
• ClinicalTrials.gov — registered clinical trials with results
• Cochrane Database — systematic reviews and meta-analyses
• NIH Office of Dietary Supplements (ODS) — official fact sheets
• FDA / Flockhart / DDInter — curated drug interaction databases
• National and international clinical guidelines (NICE, AHA, etc.)

We do not include blog posts, press releases, single-author opinion pieces, predatory journals, or content from non-peer-reviewed sources.

2. Quality Scoring (0–100)

Every source receives a quality score derived from a deterministic rubric combining study design, sample size, journal quality, peer-review status, and recency. Scores map to four public tiers:

Tier is displayed on every source badge so readers can calibrate confidence at a glance.

3. Interaction Data

Our 1,307 drug–supplement interactions are assembled from three authoritative curated datasets and manually normalized:

• DDInter 2.0 (Chongqing Medical University) — 658 interactions
• Flockhart CYP-450 Table (Indiana University) — 530 interactions
• Priolix-curated set (FDA labels, Cochrane reviews, guideline extracts) — 119 interactions

Each record includes: severity tier, pharmacological mechanism, clinical effect, management guidance, and evidence strength (well-established, probable, possible, theoretical, inconclusive).

4. How We Use AI

Priolix uses large language models (LLMs) to draft plain-language explanations and structured Q&A summaries on top of our curated data. We are transparent about this:

• Drafting model: Gemma 4 31B (Instruct, NVFP4 quantized), served locally on our own infrastructure via vLLM. No user question ever leaves our servers.
• Grounding: every AI-authored answer is built from a structured prompt that includes the relevant curated interaction record plus the top 8 PubMed sources for that (supplement, drug) pair. The model is instructed to cite only those PMIDs.
• Independent audit: every AI answer is passed to a second, independent model — NVIDIA Nemotron 3 Super 120B — which scores it on four axes: grounding, accuracy, safety, and hallucination detection. Answers are auto-approved only when composite ≥ 75 and hallucination ≥ 90.
• Disclosure: every AI-assisted answer on the site carries a 🔬 AI-assisted answer badge linking back to this page.
• Curated data takes priority: where a curated interaction record exists, it is considered authoritative and rendered alongside the AI answer. If they ever conflict, the curated record wins.

5. Human Review

AI answers flagged by the auditor model (composite score 50–74 or hallucination 70–89) are held back from publication and routed to the Priolix Evidence Team for manual review before they appear on the site. Answers scoring below the flagged threshold are rejected outright and never served to users.

6. Updates and Corrections

The Priolix evidence corpus is refreshed automatically each week. Interaction records are versioned, and every page carries a "last updated" date. If you believe any content is inaccurate or out of date, please contact us — we review every correction request and update the underlying data immediately when warranted.

7. What We Are Not

Priolix is an information resource. It is not a substitute for licensed medical advice, diagnosis, or treatment. Always consult your physician or pharmacist before starting, stopping, or changing any supplement — particularly while taking prescription medication. See our medical disclaimer for full detail.