toolnexus 0.23.0 → 0.25.0

package/src/mcp2026/toolsearch.ts

@@ -0,0 +1,56 @@
+// Lazy tool search: with client.toolSearch the LLM sees two small tools instead of all N:
+// search_tools(query) -> ranked {name, description, inputSchema, server}, and
+// call_tool(name, arguments) to run one it found. Ranking is BM25 over name (x2) + description.
+import type { Obj } from "./protocol.js"
+
+/** BM25 (k1 1.2, b 0.75) over name tokens twice + description tokens; best first, ties keep list order. */
+export function rank(tools: Obj[], query: string, limit: number): Obj[] {
+  const docs = tools.map((t) => {
+    const nt = tokens(String(t.name ?? ""))
+    return [...nt, ...nt, ...tokens(String(t.description ?? ""))]
+  })
+  const df = new Map<string, number>()
+  let total = 0
+  for (const d of docs) {
+    total += d.length
+    for (const w of new Set(d)) df.set(w, (df.get(w) ?? 0) + 1)
+  }
+  const avg = total / Math.max(1, tools.length)
+  const qs = tokens(query)
+  const hits: { i: number; score: number }[] = []
+  docs.forEach((d, i) => {
+    const tf = new Map<string, number>()
+    for (const w of d) tf.set(w, (tf.get(w) ?? 0) + 1)
+    let s = 0
+    for (const q of qs) {
+      const f = tf.get(q) ?? 0
+      if (f === 0) continue
+      const n = df.get(q) ?? 0
+      const idf = Math.log(1 + (tools.length - n + 0.5) / (n + 0.5))
+      s += (idf * f * 2.2) / (f + 1.2 * (0.25 + (0.75 * d.length) / avg))
+    }
+    if (s > 0) hits.push({ i, score: s })
+  })
+  hits.sort((a, b) => b.score - a.score) // stable
+  return hits.slice(0, Math.max(0, limit)).map((h) => tools[h.i])
+}
+
+/** Lowercases and splits on non-alphanumerics and camelCase boundaries. */
+export function tokens(s: string): string[] {
+  const out: string[] = []
+  let cur = ""
+  let prevLower = false
+  for (const r of s) {
+    if (/\p{Lu}/u.test(r) && prevLower) {
+      if (cur) out.push(cur.toLowerCase())
+      cur = r
+    } else if (/[\p{L}\p{N}]/u.test(r)) cur += r
+    else if (cur) {
+      out.push(cur.toLowerCase())
+      cur = ""
+    }
+    prevLower = /\p{Ll}/u.test(r)
+  }
+  if (cur) out.push(cur.toLowerCase())
+  return out
+}