Methodology

This page describes how MedSideInfo gathers, verifies, and presents side-effect information. We want you to be able to verify our sourcing yourself before you trust the output.

How we pick sources

We classify sources into two tiers based on regulatory authority and editorial process. The complete list is on our Sources page.

• Tier 1 — Official. Government and intergovernmental health agencies, plus the primary clinical-literature index. These have regulatory or institutional authority and the most rigorous editorial process. FDA, NIH, MedlinePlus, NLM, NHS (UK), PubMed, PubMed Central.
• Tier 2 — Medical. Peer reference sites with professional medical editorial review. Mayo Clinic, Cleveland Clinic, WebMD, Drugs.com, RxList, Healthline, Medical News Today.

Tier ranking is visible in the per-medicine result table so you can weight what you read. Authenticated users can override our default tier classifications per source.

How we cross-reference

For every medicine you search, our pipeline:

1. Resolves the name. A brand name, generic name, or common misspelling is mapped to the canonical drug. Alias matches (e.g. Tylenol → acetaminophen) come from a curated list; novel inputs go through a five-layer resolver (alias table → known- medicines list → FDA fuzzy match → embeddings → LLM fallback).
2. Discovers candidate pages. We use TinyFish browser automation to surface the most relevant page on each source for the medicine in question, in parallel across all sources. Eight sources are queried by default; authenticated users can pick a different subset.
3. Extracts side effects from each page. Each fetched page goes to a structured-extraction LLM (Fireworks). The LLM returns each side effect with a frequency label, an explanatory sentence, and a verbatim quote from the source page.
4. Grounds every quote. Before we display a side effect, our grounding verifier checks that the quoted span literally appears in the source page — character-for-character. Anything that doesn't match is dropped. This is our primary defense against LLM hallucination.
5. Normalizes across sources. Side effects from all fetched sources are unified into a single table grouped by frequency (Very common · Common · Uncommon · Rare · Very rare · Unknown). Each row links back to the exact source page and shows the verbatim quote.

How we handle conflicts between sources

When sources disagree about the frequency of a side effect — say, Mayo lists nausea as “common” and the FDA label lists it as “uncommon” — we surface the discrepancy rather than picking a winner. Each row shows where the data came from, and you can compare per-source phrasing side by side.

We do not silently average frequencies, drop sources, or weight one publisher above another beyond the visible Tier classification.

Update cadence

We refresh source data on a 24-hour rolling cache. The first time anyone searches for a given medicine, we fetch live; subsequent searches inside the cache window use the cached payload. Pages that fail to extract are queued for retry by an automated cron with a progressive strategy ladder; if every strategy fails, an alert surfaces to our admin queue for manual investigation.

What we do NOT do

Limitations we want you to know about

We try to be honest about what we do not yet do well:

• Rare diseases and niche drugs: some specialty drugs have thin coverage in our default source set. Output for these may be incomplete.
• Regional variations: a drug approved in one country may have different labelling elsewhere. We surface NHS (UK) alongside US sources, but global coverage is uneven.
• Drug-specific dosing: our pipeline does not extract or compare dosing information; effects shown are not dose-adjusted.
• New approvals: a drug approved in the last 30 days may not yet appear in consumer-reference sources our pipeline fetches.
• Drug interactions: the side-effect view shows single-drug effects; interaction effects between two drugs are a separate (in-progress) feature.

How to verify our work

Every effect we display is paired with the verbatim quote that backed it and a link to the original source page. Click through. If a quote does not appear on the source page, that is a bug — please tell us and we will fix it.