@ztimson/ai-utils 0.8.10 → 0.8.11

package/dist/index.mjs

@@ -1,40 +1,40 @@
 import * as O from "node:os";
-import { tmpdir as L } from "node:os";
-import { objectMap as R, JSONAttemptParse as _, findByProp as U, JSONSanitize as S, clean as W, Http as N, consoleInterceptor as C, fn as J, ASet as D } from "@ztimson/utils";
-import { Anthropic as F } from "@anthropic-ai/sdk";
-import { OpenAI as H } from "openai";
-import { fileURLToPath as G } from "url";
-import { join as I, dirname as B } from "path";
-import { spawn as b, execSync as K } from "node:child_process";
+import { tmpdir as U } from "node:os";
+import { objectMap as z, JSONAttemptParse as _, findByProp as R, JSONSanitize as S, clean as W, Http as C, consoleInterceptor as N, fn as J, ASet as F } from "@ztimson/utils";
+import { Anthropic as B } from "@anthropic-ai/sdk";
+import { OpenAI as D } from "openai";
+import { fileURLToPath as H } from "url";
+import { join as I, dirname as G } from "path";
+import { spawn as w, execSync as K } from "node:child_process";
 import { mkdtempSync as V } from "node:fs";
-import w from "node:fs/promises";
+import b from "node:fs/promises";
 import * as M from "node:path";
-import P, { join as q } from "node:path";
+import q, { join as A } from "node:path";
 import { createWorker as Y } from "tesseract.js";
 import * as Z from "cheerio";
 import { $Sync as j } from "@ztimson/node-utils";
-class z {
+class L {
 }
-class Q extends z {
+class Q extends L {
   constructor(r, e, t) {
-    super(), this.ai = r, this.apiToken = e, this.model = t, this.client = new F({ apiKey: e });
+    super(), this.ai = r, this.apiToken = e, this.model = t, this.client = new B({ apiKey: e });
   }
   client;
   toStandard(r) {
     const e = Date.now(), t = [];
-    for (let a of r)
-      if (typeof a.content == "string")
-        t.push({ timestamp: e, ...a });
+    for (let i of r)
+      if (typeof i.content == "string")
+        t.push({ timestamp: e, ...i });
       else {
-        const s = a.content?.filter((n) => n.type == "text").map((n) => n.text).join(`
+        const o = i.content?.filter((n) => n.type == "text").map((n) => n.text).join(`
 
 `);
-        s && t.push({ timestamp: e, role: a.role, content: s }), a.content.forEach((n) => {
+        o && t.push({ timestamp: e, role: i.role, content: o }), i.content.forEach((n) => {
           if (n.type == "tool_use")
             t.push({ timestamp: e, role: "tool", id: n.id, name: n.name, args: n.input, content: void 0 });
           else if (n.type == "tool_result") {
-            const c = t.findLast((u) => u.id == n.tool_use_id);
-            c && (c[n.is_error ? "error" : "content"] = n.content);
+            const u = t.findLast((l) => l.id == n.tool_use_id);
+            u && (u[n.is_error ? "error" : "content"] = n.content);
           }
         });
       }
@@ -55,9 +55,9 @@ class Q extends z {
   }
   ask(r, e = {}) {
     const t = new AbortController();
-    return Object.assign(new Promise(async (a) => {
-      let s = this.fromStandard([...e.history || [], { role: "user", content: r, timestamp: Date.now() }]);
-      const n = e.tools || this.ai.options.llm?.tools || [], c = {
+    return Object.assign(new Promise(async (i) => {
+      let o = this.fromStandard([...e.history || [], { role: "user", content: r, timestamp: Date.now() }]);
+      const n = e.tools || this.ai.options.llm?.tools || [], u = {
         model: e.model || this.model,
         max_tokens: e.max_tokens || this.ai.options.llm?.max_tokens || 4096,
         system: e.system || this.ai.options.llm?.system || "",
@@ -67,66 +67,66 @@ class Q extends z {
           description: m.description,
           input_schema: {
             type: "object",
-            properties: m.args ? R(m.args, (o, l) => ({ ...l, required: void 0 })) : {},
-            required: m.args ? Object.entries(m.args).filter((o) => o[1].required).map((o) => o[0]) : []
+            properties: m.args ? z(m.args, (s, a) => ({ ...a, required: void 0 })) : {},
+            required: m.args ? Object.entries(m.args).filter((s) => s[1].required).map((s) => s[0]) : []
           },
           fn: void 0
         })),
-        messages: s,
+        messages: o,
         stream: !!e.stream
       };
-      let u, i = !0;
+      let l, c = !0;
       do {
-        if (u = await this.client.messages.create(c).catch((o) => {
-          throw o.message += `
+        if (l = await this.client.messages.create(u).catch((s) => {
+          throw s.message += `
 
 Messages:
-${JSON.stringify(s, null, 2)}`, o;
+${JSON.stringify(o, null, 2)}`, s;
         }), e.stream) {
-          i ? i = !1 : e.stream({ text: `
+          c ? c = !1 : e.stream({ text: `
 
-` }), u.content = [];
-          for await (const o of u) {
+` }), l.content = [];
+          for await (const s of l) {
             if (t.signal.aborted) break;
-            if (o.type === "content_block_start")
-              o.content_block.type === "text" ? u.content.push({ type: "text", text: "" }) : o.content_block.type === "tool_use" && u.content.push({ type: "tool_use", id: o.content_block.id, name: o.content_block.name, input: "" });
-            else if (o.type === "content_block_delta")
-              if (o.delta.type === "text_delta") {
-                const l = o.delta.text;
-                u.content.at(-1).text += l, e.stream({ text: l });
-              } else o.delta.type === "input_json_delta" && (u.content.at(-1).input += o.delta.partial_json);
-            else if (o.type === "content_block_stop") {
-              const l = u.content.at(-1);
-              l.input != null && (l.input = l.input ? _(l.input, {}) : {});
-            } else if (o.type === "message_stop")
+            if (s.type === "content_block_start")
+              s.content_block.type === "text" ? l.content.push({ type: "text", text: "" }) : s.content_block.type === "tool_use" && l.content.push({ type: "tool_use", id: s.content_block.id, name: s.content_block.name, input: "" });
+            else if (s.type === "content_block_delta")
+              if (s.delta.type === "text_delta") {
+                const a = s.delta.text;
+                l.content.at(-1).text += a, e.stream({ text: a });
+              } else s.delta.type === "input_json_delta" && (l.content.at(-1).input += s.delta.partial_json);
+            else if (s.type === "content_block_stop") {
+              const a = l.content.at(-1);
+              a.input != null && (a.input = a.input ? _(a.input, {}) : {});
+            } else if (s.type === "message_stop")
               break;
           }
         }
-        const m = u.content.filter((o) => o.type === "tool_use");
+        const m = l.content.filter((s) => s.type === "tool_use");
         if (m.length && !t.signal.aborted) {
-          s.push({ role: "assistant", content: u.content });
-          const o = await Promise.all(m.map(async (l) => {
-            const d = n.find(U("name", l.name));
-            if (e.stream && e.stream({ tool: l.name }), !d) return { tool_use_id: l.id, is_error: !0, content: "Tool not found" };
+          o.push({ role: "assistant", content: l.content });
+          const s = await Promise.all(m.map(async (a) => {
+            const d = n.find(R("name", a.name));
+            if (e.stream && e.stream({ tool: a.name }), !d) return { tool_use_id: a.id, is_error: !0, content: "Tool not found" };
             try {
-              const p = await d.fn(l.input, e?.stream, this.ai);
-              return { type: "tool_result", tool_use_id: l.id, content: S(p) };
+              const p = await d.fn(a.input, e?.stream, this.ai);
+              return { type: "tool_result", tool_use_id: a.id, content: S(p) };
             } catch (p) {
-              return { type: "tool_result", tool_use_id: l.id, is_error: !0, content: p?.message || p?.toString() || "Unknown" };
+              return { type: "tool_result", tool_use_id: a.id, is_error: !0, content: p?.message || p?.toString() || "Unknown" };
             }
           }));
-          s.push({ role: "user", content: o }), c.messages = s;
+          o.push({ role: "user", content: s }), u.messages = o;
         }
-      } while (!t.signal.aborted && u.content.some((m) => m.type === "tool_use"));
-      s.push({ role: "assistant", content: u.content.filter((m) => m.type == "text").map((m) => m.text).join(`
+      } while (!t.signal.aborted && l.content.some((m) => m.type === "tool_use"));
+      o.push({ role: "assistant", content: l.content.filter((m) => m.type == "text").map((m) => m.text).join(`
 
-`) }), s = this.toStandard(s), e.stream && e.stream({ done: !0 }), e.history && e.history.splice(0, e.history.length, ...s), a(s.at(-1)?.content);
+`) }), o = this.toStandard(o), e.stream && e.stream({ done: !0 }), e.history && e.history.splice(0, e.history.length, ...o), i(o.at(-1)?.content);
     }), { abort: () => t.abort() });
   }
 }
-class A extends z {
-  constructor(r, e, t, a) {
-    super(), this.ai = r, this.host = e, this.token = t, this.model = a, this.client = new H(W({
+class P extends L {
+  constructor(r, e, t, i) {
+    super(), this.ai = r, this.host = e, this.token = t, this.model = i, this.client = new D(W({
       baseURL: e,
       apiKey: t || e ? "ignored" : void 0
     }));
@@ -136,17 +136,17 @@ class A extends z {
     for (let e = 0; e < r.length; e++) {
       const t = r[e];
       if (t.role === "assistant" && t.tool_calls) {
-        const a = t.tool_calls.map((s) => ({
+        const i = t.tool_calls.map((o) => ({
           role: "tool",
-          id: s.id,
-          name: s.function.name,
-          args: _(s.function.arguments, {}),
+          id: o.id,
+          name: o.function.name,
+          args: _(o.function.arguments, {}),
           timestamp: t.timestamp
         }));
-        r.splice(e, 1, ...a), e += a.length - 1;
+        r.splice(e, 1, ...i), e += i.length - 1;
       } else if (t.role === "tool" && t.content) {
-        const a = r.find((s) => t.tool_call_id == s.id);
-        a && (t.content.includes('"error":') ? a.error = t.content : a.content = t.content), r.splice(e, 1), e--;
+        const i = r.find((o) => t.tool_call_id == o.id);
+        i && (t.content.includes('"error":') ? i.error = t.content : i.content = t.content), r.splice(e, 1), e--;
       }
       r[e]?.timestamp || (r[e].timestamp = Date.now());
     }
@@ -167,53 +167,53 @@ class A extends z {
           content: t.error || t.content
         });
       else {
-        const { timestamp: a, ...s } = t;
-        e.push(s);
+        const { timestamp: i, ...o } = t;
+        e.push(o);
       }
       return e;
     }, []);
   }
   ask(r, e = {}) {
     const t = new AbortController();
-    return Object.assign(new Promise(async (a, s) => {
+    return Object.assign(new Promise(async (i, o) => {
       e.system && (e.history?.[0]?.role != "system" ? e.history?.splice(0, 0, { role: "system", content: e.system, timestamp: Date.now() }) : e.history[0].content = e.system);
       let n = this.fromStandard([...e.history || [], { role: "user", content: r, timestamp: Date.now() }]);
-      const c = e.tools || this.ai.options.llm?.tools || [], u = {
+      const u = e.tools || this.ai.options.llm?.tools || [], l = {
         model: e.model || this.model,
         messages: n,
         stream: !!e.stream,
         max_tokens: e.max_tokens || this.ai.options.llm?.max_tokens || 4096,
         temperature: e.temperature || this.ai.options.llm?.temperature || 0.7,
-        tools: c.map((o) => ({
+        tools: u.map((s) => ({
           type: "function",
           function: {
-            name: o.name,
-            description: o.description,
+            name: s.name,
+            description: s.description,
             parameters: {
               type: "object",
-              properties: o.args ? R(o.args, (l, d) => ({ ...d, required: void 0 })) : {},
-              required: o.args ? Object.entries(o.args).filter((l) => l[1].required).map((l) => l[0]) : []
+              properties: s.args ? z(s.args, (a, d) => ({ ...d, required: void 0 })) : {},
+              required: s.args ? Object.entries(s.args).filter((a) => a[1].required).map((a) => a[0]) : []
             }
           }
         }))
       };
-      let i, m = !0;
+      let c, m = !0;
       do {
-        if (i = await this.client.chat.completions.create(u).catch((l) => {
-          throw l.message += `
+        if (c = await this.client.chat.completions.create(l).catch((a) => {
+          throw a.message += `
 
 Messages:
-${JSON.stringify(n, null, 2)}`, l;
+${JSON.stringify(n, null, 2)}`, a;
         }), e.stream) {
           m ? m = !1 : e.stream({ text: `
 
-` }), i.choices = [{ message: { role: "assistant", content: "", tool_calls: [] } }];
-          for await (const l of i) {
+` }), c.choices = [{ message: { role: "assistant", content: "", tool_calls: [] } }];
+          for await (const a of c) {
             if (t.signal.aborted) break;
-            if (l.choices[0].delta.content && (i.choices[0].message.content += l.choices[0].delta.content, e.stream({ text: l.choices[0].delta.content })), l.choices[0].delta.tool_calls)
-              for (const d of l.choices[0].delta.tool_calls) {
-                const p = i.choices[0].message.tool_calls.find((f) => f.index === d.index);
-                p ? (d.id && (p.id = d.id), d.type && (p.type = d.type), d.function && (p.function || (p.function = {}), d.function.name && (p.function.name = d.function.name), d.function.arguments && (p.function.arguments = (p.function.arguments || "") + d.function.arguments))) : i.choices[0].message.tool_calls.push({
+            if (a.choices[0].delta.content && (c.choices[0].message.content += a.choices[0].delta.content, e.stream({ text: a.choices[0].delta.content })), a.choices[0].delta.tool_calls)
+              for (const d of a.choices[0].delta.tool_calls) {
+                const p = c.choices[0].message.tool_calls.find((f) => f.index === d.index);
+                p ? (d.id && (p.id = d.id), d.type && (p.type = d.type), d.function && (p.function || (p.function = {}), d.function.name && (p.function.name = d.function.name), d.function.arguments && (p.function.arguments = (p.function.arguments || "") + d.function.arguments))) : c.choices[0].message.tool_calls.push({
                   index: d.index,
                   id: d.id || "",
                   type: d.type || "function",
@@ -225,11 +225,11 @@ ${JSON.stringify(n, null, 2)}`, l;
               }
           }
         }
-        const o = i.choices[0].message.tool_calls || [];
-        if (o.length && !t.signal.aborted) {
-          n.push(i.choices[0].message);
-          const l = await Promise.all(o.map(async (d) => {
-            const p = c?.find(U("name", d.function.name));
+        const s = c.choices[0].message.tool_calls || [];
+        if (s.length && !t.signal.aborted) {
+          n.push(c.choices[0].message);
+          const a = await Promise.all(s.map(async (d) => {
+            const p = u?.find(R("name", d.function.name));
             if (e.stream && e.stream({ tool: d.function.name }), !p) return { role: "tool", tool_call_id: d.id, content: '{"error": "Tool not found"}' };
             try {
               const f = _(d.function.arguments, {}), g = await p.fn(f, e.stream, this.ai);
@@ -238,17 +238,17 @@ ${JSON.stringify(n, null, 2)}`, l;
               return { role: "tool", tool_call_id: d.id, content: S({ error: f?.message || f?.toString() || "Unknown" }) };
             }
           }));
-          n.push(...l), u.messages = n;
+          n.push(...a), l.messages = n;
         }
-      } while (!t.signal.aborted && i.choices?.[0]?.message?.tool_calls?.length);
-      n.push({ role: "assistant", content: i.choices[0].message.content || "" }), n = this.toStandard(n), e.stream && e.stream({ done: !0 }), e.history && e.history.splice(0, e.history.length, ...n), a(n.at(-1)?.content);
+      } while (!t.signal.aborted && c.choices?.[0]?.message?.tool_calls?.length);
+      n.push({ role: "assistant", content: c.choices[0].message.content || "" }), n = this.toStandard(n), e.stream && e.stream({ done: !0 }), e.history && e.history.splice(0, e.history.length, ...n), i(n.at(-1)?.content);
     }), { abort: () => t.abort() });
   }
 }
 class X {
   constructor(r) {
     this.ai = r, r.options.llm?.models && Object.entries(r.options.llm.models).forEach(([e, t]) => {
-      this.defaultModel || (this.defaultModel = e), t.proto == "anthropic" ? this.models[e] = new Q(this.ai, t.token, e) : t.proto == "ollama" ? this.models[e] = new A(this.ai, t.host, "not-needed", e) : t.proto == "openai" && (this.models[e] = new A(this.ai, t.host || null, t.token, e));
+      this.defaultModel || (this.defaultModel = e), t.proto == "anthropic" ? this.models[e] = new Q(this.ai, t.token, e) : t.proto == "ollama" ? this.models[e] = new P(this.ai, t.host, "not-needed", e) : t.proto == "openai" && (this.models[e] = new P(this.ai, t.host || null, t.token, e));
     });
   }
   defaultModel;
@@ -270,27 +270,27 @@ class X {
     };
     const t = e.model || this.defaultModel;
     if (!this.models[t]) throw new Error(`Model does not exist: ${t}`);
-    let a = () => {
+    let i = () => {
     };
-    return Object.assign(new Promise(async (s) => {
+    return Object.assign(new Promise(async (o) => {
       if (e.history || (e.history = []), e.memory) {
-        const c = async (i, m, o = 10) => {
-          const [l, d] = await Promise.all([
+        const u = async (c, m, s = 10) => {
+          const [a, d] = await Promise.all([
             m ? this.embedding(m) : Promise.resolve(null),
-            i ? this.embedding(i) : Promise.resolve(null)
+            c ? this.embedding(c) : Promise.resolve(null)
           ]);
           return (e.memory || []).map((p) => {
-            const f = (l ? this.cosineSimilarity(p.embeddings[0], l[0].embedding) : 0) + (d ? this.cosineSimilarity(p.embeddings[1], d[0].embedding) : 0);
+            const f = (a ? this.cosineSimilarity(p.embeddings[0], a[0].embedding) : 0) + (d ? this.cosineSimilarity(p.embeddings[1], d[0].embedding) : 0);
             return { ...p, score: f };
-          }).toSorted((p, f) => p.score - f.score).slice(0, o).map((p) => `- ${p.owner}: ${p.fact}`).join(`
+          }).toSorted((p, f) => p.score - f.score).slice(0, s).map((p) => `- ${p.owner}: ${p.fact}`).join(`
 `);
         };
         e.system += `
 You have RAG memory and will be given the top_k closest memories regarding the users query. Save anything new you have learned worth remembering from the user message using the remember tool and feel free to recall memories manually.
 `;
-        const u = await c(r);
-        u.length && e.history.push({ role: "tool", name: "recall", id: "auto_recall_" + Math.random().toString(), args: {}, content: `Things I remembered:
-${u}` }), e.tools = [{
+        const l = await u(r);
+        l.length && e.history.push({ role: "tool", name: "recall", id: "auto_recall_" + Math.random().toString(), args: {}, content: `Things I remembered:
+${l}` }), e.tools = [{
           name: "recall",
           description: "Recall the closest memories you have regarding a query using RAG",
           args: {
@@ -298,9 +298,9 @@ ${u}` }), e.tools = [{
             query: { type: "string", description: "Search memory based on a query, can be used with or without subject argument" },
             topK: { type: "number", description: "Result limit, default 5" }
           },
-          fn: (i) => {
-            if (!i.subject && !i.query) throw new Error("Either a subject or query argument is required");
-            return c(i.query, i.subject, i.topK);
+          fn: (c) => {
+            if (!c.subject && !c.query) throw new Error("Either a subject or query argument is required");
+            return u(c.query, c.subject, c.topK);
           }
         }, {
           name: "remember",
@@ -309,31 +309,31 @@ ${u}` }), e.tools = [{
             owner: { type: "string", description: "Subject/person this fact is about" },
             fact: { type: "string", description: "The information to remember" }
           },
-          fn: async (i) => {
+          fn: async (c) => {
             if (!e.memory) return;
             const m = await Promise.all([
-              this.embedding(i.owner),
-              this.embedding(`${i.owner}: ${i.fact}`)
-            ]), o = { owner: i.owner, fact: i.fact, embeddings: [m[0][0].embedding, m[1][0].embedding] };
-            return e.memory.splice(0, e.memory.length, ...e.memory.filter((l) => !(this.cosineSimilarity(o.embeddings[0], l.embeddings[0]) >= 0.9 && this.cosineSimilarity(o.embeddings[1], l.embeddings[1]) >= 0.8)), o), "Remembered!";
+              this.embedding(c.owner),
+              this.embedding(`${c.owner}: ${c.fact}`)
+            ]), s = { owner: c.owner, fact: c.fact, embeddings: [m[0][0].embedding, m[1][0].embedding] };
+            return e.memory.splice(0, e.memory.length, ...e.memory.filter((a) => !(this.cosineSimilarity(s.embeddings[0], a.embeddings[0]) >= 0.9 && this.cosineSimilarity(s.embeddings[1], a.embeddings[1]) >= 0.8)), s), "Remembered!";
           }
         }, ...e.tools || []];
       }
       const n = await this.models[t].ask(r, e);
-      if (e.memory && e.history.splice(0, e.history.length, ...e.history.filter((c) => c.role != "tool" || c.name != "recall" && c.name != "remember")), e.compress) {
-        const c = await this.ai.language.compressHistory(e.history, e.compress.max, e.compress.min, e);
-        e.history.splice(0, e.history.length, ...c);
+      if (e.memory && e.history.splice(0, e.history.length, ...e.history.filter((u) => u.role != "tool" || u.name != "recall" && u.name != "remember")), e.compress) {
+        const u = await this.ai.language.compressHistory(e.history, e.compress.max, e.compress.min, e);
+        e.history.splice(0, e.history.length, ...u);
       }
-      return s(n);
-    }), { abort: a });
+      return o(n);
+    }), { abort: i });
   }
   async code(r, e) {
     const t = await this.ask(r, { ...e, system: [
       e?.system,
       "Return your response in a code block"
-    ].filter((s) => !!s).join(`
-`) }), a = /```(?:.+)?\s*([\s\S]*?)```/.exec(t);
-    return a ? a[1].trim() : null;
+    ].filter((o) => !!o).join(`
+`) }), i = /```(?:.+)?\s*([\s\S]*?)```/.exec(t);
+    return i ? i[1].trim() : null;
   }
   /**
    * Compress chat history to reduce context size
@@ -343,17 +343,17 @@ ${u}` }), e.tools = [{
    * @param {LLMRequest} options LLM options
    * @returns {Promise<LLMMessage[]>} New chat history will summary at index 0
    */
-  async compressHistory(r, e, t, a) {
+  async compressHistory(r, e, t, i) {
     if (this.estimateTokens(r) < e) return r;
-    let s = 0, n = 0;
+    let o = 0, n = 0;
     for (let d of r.toReversed())
-      if (n += this.estimateTokens(d.content), n < t) s++;
+      if (n += this.estimateTokens(d.content), n < t) o++;
       else break;
-    if (r.length <= s) return r;
-    const c = r[0].role == "system" ? r[0] : null, u = s == 0 ? [] : r.slice(-s), i = (s == 0 ? r : r.slice(0, -s)).filter((d) => d.role === "assistant" || d.role === "user"), m = await this.summarize(i.map((d) => `[${d.role}]: ${d.content}`).join(`
+    if (r.length <= o) return r;
+    const u = r[0].role == "system" ? r[0] : null, l = o == 0 ? [] : r.slice(-o), c = (o == 0 ? r : r.slice(0, -o)).filter((d) => d.role === "assistant" || d.role === "user"), m = await this.summarize(c.map((d) => `[${d.role}]: ${d.content}`).join(`
 
-`), 500, a), o = Date.now(), l = [{ role: "tool", name: "summary", id: "summary_" + o, args: {}, content: `Conversation Summary: ${m?.summary}`, timestamp: o }, ...u];
-    return c && l.splice(0, 0, c), l;
+`), 500, i), s = Date.now(), a = [{ role: "tool", name: "summary", id: "summary_" + s, args: {}, content: `Conversation Summary: ${m?.summary}`, timestamp: s }, ...l];
+    return u && a.splice(0, 0, u), a;
   }
   /**
    * Compare the difference between embeddings (calculates the angle between two vectors)
@@ -363,10 +363,10 @@ ${u}` }), e.tools = [{
    */
   cosineSimilarity(r, e) {
     if (r.length !== e.length) throw new Error("Vectors must be same length");
-    let t = 0, a = 0, s = 0;
-    for (let c = 0; c < r.length; c++)
-      t += r[c] * e[c], a += r[c] * r[c], s += e[c] * e[c];
-    const n = Math.sqrt(a) * Math.sqrt(s);
+    let t = 0, i = 0, o = 0;
+    for (let u = 0; u < r.length; u++)
+      t += r[u] * e[u], i += r[u] * r[u], o += e[u] * e[u];
+    const n = Math.sqrt(i) * Math.sqrt(o);
     return n === 0 ? 0 : t / n;
   }
   /**
@@ -377,25 +377,25 @@ ${u}` }), e.tools = [{
    * @returns {string[]} Chunked strings
    */
   chunk(r, e = 500, t = 50) {
-    const a = (u, i = "") => u ? Object.entries(u).flatMap(([m, o]) => {
-      const l = i ? `${i}${isNaN(+m) ? `.${m}` : `[${m}]`}` : m;
-      return typeof o == "object" && !Array.isArray(o) ? a(o, l) : `${l}: ${Array.isArray(o) ? o.join(", ") : o}`;
-    }) : [], n = (typeof r == "object" ? a(r) : r.toString().split(`
-`)).flatMap((u) => [...u.split(/\s+/).filter(Boolean), `
-`]), c = [];
-    for (let u = 0; u < n.length; ) {
-      let i = "", m = u;
+    const i = (l, c = "") => l ? Object.entries(l).flatMap(([m, s]) => {
+      const a = c ? `${c}${isNaN(+m) ? `.${m}` : `[${m}]`}` : m;
+      return typeof s == "object" && !Array.isArray(s) ? i(s, a) : `${a}: ${Array.isArray(s) ? s.join(", ") : s}`;
+    }) : [], n = (typeof r == "object" ? i(r) : r.toString().split(`
+`)).flatMap((l) => [...l.split(/\s+/).filter(Boolean), `
+`]), u = [];
+    for (let l = 0; l < n.length; ) {
+      let c = "", m = l;
       for (; m < n.length; ) {
-        const l = i + (i ? " " : "") + n[m];
-        if (this.estimateTokens(l.replace(/\s*\n\s*/g, `
-`)) > e && i) break;
-        i = l, m++;
+        const a = c + (c ? " " : "") + n[m];
+        if (this.estimateTokens(a.replace(/\s*\n\s*/g, `
+`)) > e && c) break;
+        c = a, m++;
       }
-      const o = i.replace(/\s*\n\s*/g, `
+      const s = c.replace(/\s*\n\s*/g, `
 `).trim();
-      o && c.push(o), u = Math.max(m - t, m === u ? u + 1 : m);
+      s && u.push(s), l = Math.max(m - t, m === l ? l + 1 : m);
     }
-    return c;
+    return u;
   }
   /**
    * Create a vector representation of a string
@@ -404,39 +404,39 @@ ${u}` }), e.tools = [{
    * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
    */
   embedding(r, e = {}) {
-    let { maxTokens: t = 500, overlapTokens: a = 50 } = e, s = !1;
+    let { maxTokens: t = 500, overlapTokens: i = 50 } = e, o = !1;
     const n = () => {
-      s = !0;
-    }, c = (i) => new Promise((m, o) => {
-      if (s) return o(new Error("Aborted"));
-      const l = [
-        I(B(G(import.meta.url)), "embedder.js"),
+      o = !0;
+    }, u = (c) => new Promise((m, s) => {
+      if (o) return s(new Error("Aborted"));
+      const a = [
+        I(G(H(import.meta.url)), "embedder.js"),
         this.ai.options.path,
         this.ai.options?.embedder || "bge-small-en-v1.5"
-      ], d = b("node", l, { stdio: ["pipe", "pipe", "ignore"] });
-      d.stdin.write(i), d.stdin.end();
+      ], d = w("node", a, { stdio: ["pipe", "pipe", "ignore"] });
+      d.stdin.write(c), d.stdin.end();
       let p = "";
       d.stdout.on("data", (f) => p += f.toString()), d.on("close", (f) => {
-        if (s) return o(new Error("Aborted"));
+        if (o) return s(new Error("Aborted"));
         if (f === 0)
           try {
             const g = JSON.parse(p);
             m(g.embedding);
           } catch {
-            o(new Error("Failed to parse embedding output"));
+            s(new Error("Failed to parse embedding output"));
           }
         else
-          o(new Error(`Embedder process exited with code ${f}`));
-      }), d.on("error", o);
-    }), u = (async () => {
-      const i = this.chunk(r, t, a), m = [];
-      for (let o = 0; o < i.length && !s; o++) {
-        const l = i[o], d = await c(l);
-        m.push({ index: o, embedding: d, text: l, tokens: this.estimateTokens(l) });
+          s(new Error(`Embedder process exited with code ${f}`));
+      }), d.on("error", s);
+    }), l = (async () => {
+      const c = this.chunk(r, t, i), m = [];
+      for (let s = 0; s < c.length && !o; s++) {
+        const a = c[s], d = await u(a);
+        m.push({ index: s, embedding: d, text: a, tokens: this.estimateTokens(a) });
       }
       return m;
     })();
-    return Object.assign(u, { abort: n });
+    return Object.assign(l, { abort: n });
   }
   /**
    * Estimate variable as tokens
@@ -455,8 +455,8 @@ ${u}` }), e.tools = [{
    */
   fuzzyMatch(r, ...e) {
     if (e.length < 2) throw new Error("Requires at least 2 strings to compare");
-    const t = (n, c = 10) => n.toLowerCase().split("").map((u, i) => u.charCodeAt(0) * (i + 1) % c / c).slice(0, c), a = t(r), s = e.map((n) => t(n)).map((n) => this.cosineSimilarity(a, n));
-    return { avg: s.reduce((n, c) => n + c, 0) / s.length, max: Math.max(...s), similarities: s };
+    const t = (n, u = 10) => n.toLowerCase().split("").map((l, c) => l.charCodeAt(0) * (c + 1) % u / u).slice(0, u), i = t(r), o = e.map((n) => t(n)).map((n) => this.cosineSimilarity(i, n));
+    return { avg: o.reduce((n, u) => n + u, 0) / o.length, max: Math.max(...o), similarities: o };
   }
   /**
    * Ask a question with JSON response
@@ -466,15 +466,15 @@ ${u}` }), e.tools = [{
    * @returns {Promise<{} | {} | RegExpExecArray | null>}
    */
   async json(r, e, t) {
-    const a = await this.code(r, { ...t, system: [
+    const i = await this.code(r, { ...t, system: [
       t?.system,
       `Only respond using JSON matching this schema:
 \`\`\`json
 ${e}
 \`\`\``
-    ].filter((s) => !!s).join(`
+    ].filter((o) => !!o).join(`
 `) });
-    return a ? _(a, {}) : null;
+    return i ? _(i, {}) : null;
   }
   /**
    * Create a summary of some text
@@ -510,21 +510,21 @@ print(json.dumps(segments))
   pyannote;
   whisperModel;
   async addPunctuation(r, e, t = 150) {
-    const a = (n) => {
+    const i = (n) => {
       if (n = n.toLowerCase().replace(/[^a-z]/g, ""), n.length <= 3) return 1;
-      const c = n.match(/[aeiouy]+/g);
-      let u = c ? c.length : 1;
-      return n.endsWith("e") && u--, Math.max(1, u);
+      const u = n.match(/[aeiouy]+/g);
+      let l = u ? u.length : 1;
+      return n.endsWith("e") && l--, Math.max(1, l);
     };
-    let s = "";
-    return r.transcription.filter((n, c) => {
-      let u = !1;
-      const i = r.transcription[c - 1], m = r.transcription[c + 1];
-      return !n.text && m ? (m.offsets.from = n.offsets.from, m.timestamps.from = n.offsets.from) : n.text && n.text[0] != " " && i && (i.offsets.to = n.offsets.to, i.timestamps.to = n.timestamps.to, i.text += n.text, u = !0), !!n.text && !u;
+    let o = "";
+    return r.transcription.filter((n, u) => {
+      let l = !1;
+      const c = r.transcription[u - 1], m = r.transcription[u + 1];
+      return !n.text && m ? (m.offsets.from = n.offsets.from, m.timestamps.from = n.offsets.from) : n.text && n.text[0] != " " && c && (c.offsets.to = n.offsets.to, c.timestamps.to = n.timestamps.to, c.text += n.text, l = !0), !!n.text && !l;
     }).forEach((n) => {
-      const c = /^[A-Z]/.test(n.text.trim()), u = n.offsets.to - n.offsets.from, m = a(n.text.trim()) * t;
-      c && u > m * 2 && n.text[0] == " " && (s += "."), s += n.text;
-    }), e ? this.ai.language.ask(s, {
+      const u = /^[A-Z]/.test(n.text.trim()), l = n.offsets.to - n.offsets.from, m = i(n.text.trim()) * t;
+      u && l > m * 2 && n.text[0] == " " && (o += "."), o += n.text;
+    }), e ? this.ai.language.ask(o, {
       system: "Remove any misplaced punctuation from the following ASR transcript using the replace tool. Avoid modifying words unless there is an obvious typo",
       temperature: 0.1,
       tools: [{
@@ -534,138 +534,138 @@ print(json.dumps(segments))
           find: { type: "string", description: "Text to find", required: !0 },
           replace: { type: "string", description: "Text to replace", required: !0 }
         },
-        fn: (n) => s = s.replace(n.find, n.replace)
+        fn: (n) => o = o.replace(n.find, n.replace)
       }]
-    }).then(() => s) : s.trim();
+    }).then(() => o) : o.trim();
   }
   async diarizeTranscript(r, e, t) {
-    const a = /* @__PURE__ */ new Map();
-    let s = 0;
+    const i = /* @__PURE__ */ new Map();
+    let o = 0;
     e.forEach((p) => {
-      a.has(p.speaker) || a.set(p.speaker, ++s);
+      i.has(p.speaker) || i.set(p.speaker, ++o);
     });
-    const n = await this.addPunctuation(r, t), c = n.match(/[^.!?]+[.!?]+/g) || [n], u = r.transcription.filter((p) => p.text.trim()), i = c.map((p) => {
+    const n = await this.addPunctuation(r, t), u = n.match(/[^.!?]+[.!?]+/g) || [n], l = r.transcription.filter((p) => p.text.trim()), c = u.map((p) => {
       if (p = p.trim(), !p) return null;
       const f = p.toLowerCase().replace(/[^\w\s]/g, "").split(/\s+/), g = /* @__PURE__ */ new Map();
-      f.forEach((k) => {
-        const x = u.find((y) => k === y.text.trim().toLowerCase().replace(/[^\w]/g, ""));
-        if (!x) return;
-        const $ = x.offsets.from / 1e3, v = e.find((y) => $ >= y.start && $ <= y.end);
+      f.forEach((x) => {
+        const k = l.find((y) => x === y.text.trim().toLowerCase().replace(/[^\w]/g, ""));
+        if (!k) return;
+        const $ = k.offsets.from / 1e3, v = e.find((y) => $ >= y.start && $ <= y.end);
         if (v) {
-          const y = a.get(v.speaker);
+          const y = i.get(v.speaker);
           g.set(y, (g.get(y) || 0) + 1);
         }
       });
       let T = 1, E = 0;
-      return g.forEach((k, x) => {
-        k > E && (E = k, T = x);
+      return g.forEach((x, k) => {
+        x > E && (E = x, T = k);
       }), { speaker: T, text: p };
     }).filter((p) => p !== null), m = [];
-    i.forEach((p) => {
+    c.forEach((p) => {
       const f = m[m.length - 1];
       f && f.speaker === p.speaker ? f.text += " " + p.text : m.push({ ...p });
     });
-    let o = m.map((p) => `[Speaker ${p.speaker}]: ${p.text}`).join(`
+    let s = m.map((p) => `[Speaker ${p.speaker}]: ${p.text}`).join(`
 `).trim();
-    if (!t) return o;
-    let l = this.ai.language.chunk(o, 500, 0);
-    l.length > 4 && (l = [...l.slice(0, 3), l.at(-1)]);
-    const d = await this.ai.language.json(l.join(`
+    if (!t) return s;
+    let a = this.ai.language.chunk(s, 500, 0);
+    a.length > 4 && (a = [...a.slice(0, 3), a.at(-1)]);
+    const d = await this.ai.language.json(a.join(`
 `), '{1: "Detected Name", 2: "Second Name"}', {
       system: "Use the following transcript to identify speakers. Only identify speakers you are positive about, dont mention speakers you are unsure about in your response",
       temperature: 0.1
     });
-    return Object.entries(d).forEach(([p, f]) => o = o.replaceAll(`[Speaker ${p}]`, `[${f}]`)), o;
+    return Object.entries(d).forEach(([p, f]) => s = s.replaceAll(`[Speaker ${p}]`, `[${f}]`)), s;
   }
   runAsr(r, e = {}) {
     let t;
-    const a = new Promise((s, n) => {
-      this.downloadAsrModel(e.model).then((c) => {
+    const i = new Promise((o, n) => {
+      this.downloadAsrModel(e.model).then((u) => {
         if (e.diarization) {
-          let u = M.join(M.dirname(r), "transcript");
-          t = b(
+          let l = M.join(M.dirname(r), "transcript");
+          t = w(
             this.ai.options.whisper,
-            ["-m", c, "-f", r, "-np", "-ml", "1", "-oj", "-of", u],
+            ["-m", u, "-f", r, "-np", "-ml", "1", "-oj", "-of", l],
             { stdio: ["ignore", "ignore", "pipe"] }
-          ), t.on("error", (i) => n(i)), t.on("close", async (i) => {
-            if (i === 0) {
-              u = await w.readFile(u + ".json", "utf-8"), w.rm(u + ".json").catch(() => {
+          ), t.on("error", (c) => n(c)), t.on("close", async (c) => {
+            if (c === 0) {
+              l = await b.readFile(l + ".json", "utf-8"), b.rm(l + ".json").catch(() => {
               });
               try {
-                s(JSON.parse(u));
+                o(JSON.parse(l));
               } catch {
                 n(new Error("Failed to parse whisper JSON"));
               }
             } else
-              n(new Error(`Exit code ${i}`));
+              n(new Error(`Exit code ${c}`));
           });
         } else {
-          let u = "";
-          t = b(this.ai.options.whisper, ["-m", c, "-f", r, "-np", "-nt"]), t.on("error", (i) => n(i)), t.stdout.on("data", (i) => u += i.toString()), t.on("close", async (i) => {
-            i === 0 ? s(u.trim() || null) : n(new Error(`Exit code ${i}`));
+          let l = "";
+          t = w(this.ai.options.whisper, ["-m", u, "-f", r, "-np", "-nt"]), t.on("error", (c) => n(c)), t.stdout.on("data", (c) => l += c.toString()), t.on("close", async (c) => {
+            c === 0 ? o(l.trim() || null) : n(new Error(`Exit code ${c}`));
           });
         }
       });
     });
-    return Object.assign(a, { abort: () => t?.kill("SIGTERM") });
+    return Object.assign(i, { abort: () => t?.kill("SIGTERM") });
   }
   runDiarization(r) {
     let e = !1, t = () => {
       e = !0;
     };
-    const a = (n) => new Promise((c) => {
-      const u = b(n, ["-W", "ignore", "-c", "import pyannote.audio"]);
-      u.on("close", (i) => c(i === 0)), u.on("error", () => c(!1));
-    }), s = Promise.all([
-      a("python"),
-      a("python3")
-    ]).then((async ([n, c]) => {
+    const i = (n) => new Promise((u) => {
+      const l = w(n, ["-W", "ignore", "-c", "import pyannote.audio"]);
+      l.on("close", (c) => u(c === 0)), l.on("error", () => u(!1));
+    }), o = Promise.all([
+      i("python"),
+      i("python3")
+    ]).then((async ([n, u]) => {
       if (e) return;
-      if (!n && !c) throw new Error("Pyannote is not installed: pip install pyannote.audio");
-      const u = c ? "python3" : "python";
-      return new Promise((i, m) => {
+      if (!n && !u) throw new Error("Pyannote is not installed: pip install pyannote.audio");
+      const l = u ? "python3" : "python";
+      return new Promise((c, m) => {
         if (e) return;
-        let o = "";
-        const l = b(u, ["-W", "ignore", "-c", this.pyannote, r]);
-        l.stdout.on("data", (d) => o += d.toString()), l.stderr.on("data", (d) => console.error(d.toString())), l.on("close", (d) => {
+        let s = "";
+        const a = w(l, ["-W", "ignore", "-c", this.pyannote, r]);
+        a.stdout.on("data", (d) => s += d.toString()), a.stderr.on("data", (d) => console.error(d.toString())), a.on("close", (d) => {
           if (d === 0)
             try {
-              i(JSON.parse(o));
+              c(JSON.parse(s));
             } catch {
               m(new Error("Failed to parse diarization output"));
             }
           else
             m(new Error(`Python process exited with code ${d}`));
-        }), l.on("error", m), t = () => l.kill("SIGTERM");
+        }), a.on("error", m), t = () => a.kill("SIGTERM");
       });
     }));
-    return Object.assign(s, { abort: t });
+    return Object.assign(o, { abort: t });
   }
   asr(r, e = {}) {
     if (!this.ai.options.whisper) throw new Error("Whisper not configured");
-    const t = q(V(q(L(), "audio-")), "converted.wav");
+    const t = A(V(A(U(), "audio-")), "converted.wav");
     K(`ffmpeg -i "${r}" -ar 16000 -ac 1 -f wav "${t}"`, { stdio: "ignore" });
-    const a = () => w.rm(P.dirname(t), { recursive: !0, force: !0 }).catch(() => {
+    const i = () => b.rm(q.dirname(t), { recursive: !0, force: !0 }).catch(() => {
     });
     if (!e.diarization) return this.runAsr(t, { model: e.model });
-    const s = this.runAsr(t, { model: e.model, diarization: !0 }), n = this.runDiarization(t);
-    let c = !1, u = () => {
-      c = !0, s.abort(), n.abort(), a();
+    const o = this.runAsr(t, { model: e.model, diarization: !0 }), n = this.runDiarization(t);
+    let u = !1, l = () => {
+      u = !0, o.abort(), n.abort(), i();
     };
-    const i = Promise.allSettled([s, n]).then(async ([m, o]) => {
+    const c = Promise.allSettled([o, n]).then(async ([m, s]) => {
       if (m.status == "rejected") throw new Error(`Whisper.cpp timestamps:
 ` + m.reason);
-      if (o.status == "rejected") throw new Error(`Pyannote:
-` + o.reason);
-      return c || !e.diarization ? m.value : this.diarizeTranscript(m.value, o.value, e.diarization == "llm");
-    }).finally(() => a());
-    return Object.assign(i, { abort: u });
+      if (s.status == "rejected") throw new Error(`Pyannote:
+` + s.reason);
+      return u || !e.diarization ? m.value : this.diarizeTranscript(m.value, s.value, e.diarization == "llm");
+    }).finally(() => i());
+    return Object.assign(c, { abort: l });
   }
   async downloadAsrModel(r = this.whisperModel) {
     if (!this.ai.options.whisper) throw new Error("Whisper not configured");
     r.endsWith(".bin") || (r += ".bin");
-    const e = P.join(this.ai.options.path, r);
-    return await w.stat(e).then(() => !0).catch(() => !1) ? e : this.downloads[r] ? this.downloads[r] : (this.downloads[r] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${r}`).then((t) => t.arrayBuffer()).then((t) => Buffer.from(t)).then(async (t) => (await w.writeFile(e, t), delete this.downloads[r], e)), this.downloads[r]);
+    const e = q.join(this.ai.options.path, r);
+    return await b.stat(e).then(() => !0).catch(() => !1) ? e : this.downloads[r] ? this.downloads[r] : (this.downloads[r] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${r}`).then((t) => t.arrayBuffer()).then((t) => Buffer.from(t)).then(async (t) => (await b.writeFile(e, t), delete this.downloads[r], e)), this.downloads[r]);
   }
 }
 class te {
@@ -679,15 +679,15 @@ class te {
    */
   ocr(r) {
     let e;
-    const t = new Promise(async (a) => {
+    const t = new Promise(async (i) => {
       e = await Y(this.ai.options.ocr || "eng", 2, { cachePath: this.ai.options.path });
-      const { data: s } = await e.recognize(r);
-      await e.terminate(), a(s.text.trim() || null);
+      const { data: o } = await e.recognize(r);
+      await e.terminate(), i(o.text.trim() || null);
     });
     return Object.assign(t, { abort: () => e?.terminate() });
   }
 }
-class we {
+class be {
   constructor(r) {
     this.options = r, r.path || (r.path = O.tmpdir()), process.env.TRANSFORMERS_CACHE = r.path, this.audio = new ee(this), this.language = new X(this), this.vision = new te(this);
   }
@@ -703,17 +703,17 @@ const re = () => O.platform() == "win32" ? "cmd" : j`echo $SHELL`?.split("/").po
   description: "Use the command line interface, returns any output",
   args: { command: { type: "string", description: "Command to run", required: !0 } },
   fn: (h) => j`${h.command}`
-}, be = {
+}, we = {
   name: "get_datetime",
   description: "Get local date / time",
   args: {},
   fn: async () => (/* @__PURE__ */ new Date()).toString()
-}, ke = {
+}, xe = {
   name: "get_datetime_utc",
   description: "Get current UTC date / time",
   args: {},
   fn: async () => (/* @__PURE__ */ new Date()).toUTCString()
-}, xe = {
+}, ke = {
   name: "exec",
   description: "Run code/scripts",
   args: {
@@ -745,7 +745,7 @@ const re = () => O.platform() == "win32" ? "cmd" : j`echo $SHELL`?.split("/").po
     headers: { type: "object", description: "HTTP headers to send", default: {} },
     body: { type: "object", description: "HTTP body to send" }
   },
-  fn: (h) => new N({ url: h.url, headers: h.headers }).request({ method: h.method || "GET", body: h.body })
+  fn: (h) => new C({ url: h.url, headers: h.headers }).request({ method: h.method || "GET", body: h.body })
 }, se = {
   name: "exec_javascript",
   description: "Execute commonjs javascript",
@@ -753,7 +753,7 @@ const re = () => O.platform() == "win32" ? "cmd" : j`echo $SHELL`?.split("/").po
     code: { type: "string", description: "CommonJS javascript", required: !0 }
   },
   fn: async (h) => {
-    const r = C(null), e = await J({ console: r }, h.code, !0).catch((t) => r.output.error.push(t));
+    const r = N(null), e = await J({ console: r }, h.code, !0).catch((t) => r.output.error.push(t));
     return { ...r.output, return: e, stdout: void 0, stderr: void 0 };
   }
 }, oe = {
@@ -765,30 +765,78 @@ const re = () => O.platform() == "win32" ? "cmd" : j`echo $SHELL`?.split("/").po
   fn: async (h) => ({ result: j`python -c "${h.code}"` })
 }, Se = {
   name: "read_webpage",
-  description: "Extract clean, structured content from a webpage. Use after web_search to read specific URLs",
+  description: "Extract clean, structured content from a webpage or convert media/documents to accessible formats",
   args: {
     url: { type: "string", description: "URL to extract content from", required: !0 },
-    focus: { type: "string", description: 'Optional: What aspect to focus on (e.g., "pricing", "features", "contact info")' }
+    mimeRegex: { type: "string", description: 'Optional: Regex pattern to filter MIME types (e.g., "^image/", "text/", "application/pdf")' },
+    maxSize: { type: "number", description: "Optional: Max file size in bytes for binary content (default: 10MB)" }
   },
   fn: async (h) => {
-    const r = await fetch(h.url, { headers: { "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64)" } }).then((n) => n.text()).catch((n) => {
-      throw new Error(`Failed to fetch: ${n.message}`);
-    }), e = Z.load(r);
-    e('script, style, nav, footer, header, aside, iframe, noscript, [role="navigation"], [role="banner"], .ad, .ads, .cookie, .popup').remove();
-    const t = {
-      title: e('meta[property="og:title"]').attr("content") || e("title").text() || "",
-      description: e('meta[name="description"]').attr("content") || e('meta[property="og:description"]').attr("content") || ""
+    const e = await fetch(h.url, {
+      headers: {
+        "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36",
+        Accept: "text/html,application/xhtml+xml,application/xml;q=0.9,image/webp,*/*;q=0.8",
+        "Accept-Language": "en-US,en;q=0.5"
+      },
+      redirect: "follow"
+    }).catch((s) => {
+      throw new Error(`Failed to fetch: ${s.message}`);
+    }), t = e.headers.get("content-type") || "", i = t.split(";")[0].trim().toLowerCase();
+    if (t.match(/charset=([^;]+)/)?.[1], h.mimeRegex && !new RegExp(h.mimeRegex, "i").test(i))
+      return { url: h.url, error: "MIME type rejected", mimeType: i, filter: h.mimeRegex };
+    if (i.startsWith("image/") || i.startsWith("audio/") || i.startsWith("video/")) {
+      const s = await e.arrayBuffer();
+      if (s.byteLength > 10485760)
+        return { url: h.url, type: "media", mimeType: i, error: "File too large", size: s.byteLength, maxSize: 10485760 };
+      const a = Buffer.from(s).toString("base64");
+      return { url: h.url, type: "media", mimeType: i, dataUrl: `data:${i};base64,${a}`, size: s.byteLength };
+    }
+    if (i.match(/^(text\/(plain|csv|xml)|application\/(json|xml|csv|x-yaml))/) || h.url.match(/\.(txt|json|xml|csv|yaml|yml|md)$/i)) {
+      const s = await e.text();
+      return { url: h.url, type: "text", mimeType: i, content: s.slice(0, 1e5) };
+    }
+    if (i === "application/pdf" || i.startsWith("application/") && !i.includes("html")) {
+      const s = await e.arrayBuffer();
+      if (s.byteLength > 10485760)
+        return { url: h.url, type: "binary", mimeType: i, error: "File too large", size: s.byteLength, maxSize: 10485760 };
+      const a = Buffer.from(s).toString("base64");
+      return { url: h.url, type: "binary", mimeType: i, dataUrl: `data:${i};base64,${a}`, size: s.byteLength };
+    }
+    const o = await e.text(), n = Z.load(o);
+    n('script, style, nav, footer, header, aside, iframe, noscript, svg, [role="navigation"], [role="banner"], [role="complementary"], .ad, .ads, .advertisement, .cookie, .popup, .modal, .sidebar, .related, .comments, .social-share').remove();
+    const u = {
+      title: n('meta[property="og:title"]').attr("content") || n("title").text() || "",
+      description: n('meta[name="description"]').attr("content") || n('meta[property="og:description"]').attr("content") || "",
+      author: n('meta[name="author"]').attr("content") || "",
+      published: n('meta[property="article:published_time"]').attr("content") || n("time").attr("datetime") || "",
+      image: n('meta[property="og:image"]').attr("content") || ""
     };
-    let a = "";
-    const s = ["article", "main", '[role="main"]', ".content", ".post", ".entry", "body"];
-    for (const n of s) {
-      const c = e(n).first();
-      if (c.length && c.text().trim().length > 200) {
-        a = c.text();
+    let l = "";
+    const c = ["article", "main", '[role="main"]', ".content", ".post-content", ".entry-content", ".article-content", "body"];
+    for (const s of c) {
+      const a = n(s).first();
+      if (a.length && a.text().trim().length > 200) {
+        l = a.text();
         break;
       }
     }
-    return a || (a = e("body").text()), a = a.replace(/\s+/g, " ").trim().slice(0, 8e3), { url: h.url, title: t.title.trim(), description: t.description.trim(), content: a, focus: h.focus };
+    l || (l = n("body").text()), l = l.replace(/\n\s*\n\s*\n/g, `
+
+`).replace(/[ \t]+/g, " ").trim().slice(0, 5e4);
+    let m = [];
+    return l.length < 500 && (n("a[href]").each((s, a) => {
+      const d = n(a).attr("href"), p = n(a).text().trim();
+      d && p && !d.startsWith("#") && m.push({ text: p, href: d });
+    }), m = m.slice(0, 50)), {
+      url: h.url,
+      type: "html",
+      title: u.title.trim(),
+      description: u.description.trim(),
+      author: u.author.trim(),
+      published: u.published,
+      content: l,
+      links: m.length ? m : void 0
+    };
   }
 }, je = {
   name: "web_search",
@@ -800,28 +848,28 @@ const re = () => O.platform() == "win32" ? "cmd" : j`echo $SHELL`?.split("/").po
   fn: async (h) => {
     const r = await fetch(`https://html.duckduckgo.com/html/?q=${encodeURIComponent(h.query)}`, {
       headers: { "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64)", "Accept-Language": "en-US,en;q=0.9" }
-    }).then((s) => s.text());
+    }).then((o) => o.text());
     let e, t = /<a .*?href="(.+?)".+?<\/a>/g;
-    const a = new D();
+    const i = new F();
     for (; (e = t.exec(r)) !== null; ) {
-      let s = /uddg=(.+)&amp?/.exec(decodeURIComponent(e[1]))?.[1];
-      if (s && (s = decodeURIComponent(s)), s && a.add(s), a.size >= (h.length || 5)) break;
+      let o = /uddg=(.+)&amp?/.exec(decodeURIComponent(e[1]))?.[1];
+      if (o && (o = decodeURIComponent(o)), o && i.add(o), i.size >= (h.length || 5)) break;
     }
-    return a;
+    return i;
   }
 };
 export {
-  we as Ai,
+  be as Ai,
   Q as Anthropic,
   ee as Audio,
   ne as CliTool,
-  be as DateTimeTool,
-  ke as DateTimeUTCTool,
-  xe as ExecTool,
+  we as DateTimeTool,
+  xe as DateTimeUTCTool,
+  ke as ExecTool,
   _e as FetchTool,
   se as JSTool,
-  z as LLMProvider,
-  A as OpenAi,
+  L as LLMProvider,
+  P as OpenAi,
   oe as PythonTool,
   Se as ReadWebpageTool,
   te as Vision,