ScienceGPT — a domain-agnostic search engine wearing a physics costume

ScienceGPT sits over the Brookhaven Nuclear Science References (NSR) database — decades of citations indexed by nuclide, reaction, and measured quantity. It's the most domain-specific product in the family, and also the most portable: strip the physics vocabulary and what's left is a clean pattern for searching any large, structured reference corpus. That's why its natural home is science.nysgpt.com today, but its guts would serve a finance or contracts surface just as well.

Search with modes, not just a box

The search input is a plain, fast field — but it carries a mode (which axis of the corpus you're querying), so one component serves Home, References, and the ENDF report views without forking:

// src/components/SearchInput.tsx
export function SearchInput({ value, onChange, isLoading, placeholder = "Search NSR records..." }: SimpleSearchInputProps) {
  return (
    <div className="relative w-full">
      <Search className="absolute left-3 top-1/2 h-4 w-4 -translate-y-1/2 text-muted-foreground" />
      <input
        type="text"
        value={value}
        onChange={(e) => onChange(e.target.value)}
        placeholder={placeholder}
        className="h-11 md:h-9 w-full rounded-lg border border-input bg-background pl-10 pr-4 focus-visible:ring-2 focus-visible:ring-ring"
      />
      {isLoading && (
        <div className="absolute right-4 top-1/2 -translate-y-1/2">
          <div className="h-4 w-4 animate-spin rounded-full border-2 border-nuclear border-t-transparent" />
        </div>
      )}
    </div>
  );
}

The SearchMode type and a mode picker turn a single box into a faceted search entry point — the same idea that powers the search experiences in SportsGPT and TariffsGPT.

A typed combobox bound to real values

The most reusable piece is the reaction combobox: an autocomplete backed by a distinct_reactions view, with input normalization so users can type the notation however they like — with or without parentheses, any case:

// src/components/ReactionCombobox.tsx
/** Normalize input for matching: strip parens and lowercase */
function normalize(s: string): string {
  return s.replace(/^\(/, "").replace(/\)$/, "").toLowerCase().trim();
}

async function fetchReactions(): Promise<string[]> {
  const { data, error } = await supabase
    .from("distinct_reactions")
    .select("value")
    .order("value", { ascending: true });
  if (error) throw new Error(error.message);
  return (data ?? []).map((r) => r.value);
}

Pulling the suggestion set from a distinct_* table is the generalizable trick: the autocomplete is always in sync with what's actually in the data, and swapping distinct_reactions for distinct_tickers or distinct_clauses re-points the whole component at a new domain.

Why it matters

ScienceGPT is a reminder that a good search UX is structural, not cosmetic. Modes turn a box into facets, a distinct-values table keeps autocomplete honest, and a card/drawer pair handles detail. Dress it in physics for BNL, or point it at a different corpus tomorrow.

Brendan Stanton

Founder, NYSgpt