@ztimson/ai-utils 0.7.6 → 0.7.8

package/dist/index.mjs

@@ -1,109 +1,112 @@
 import * as $ from "node:os";
-import { objectMap as k, JSONAttemptParse as g, findByProp as _, JSONSanitize as b, clean as M, Http as E, consoleInterceptor as P, fn as A, ASet as v } from "@ztimson/utils";
-import { Anthropic as O } from "@anthropic-ai/sdk";
+import { tmpdir as P } from "node:os";
+import { objectMap as j, JSONAttemptParse as w, findByProp as T, JSONSanitize as b, clean as M, Http as O, consoleInterceptor as A, fn as q, ASet as v } from "@ztimson/utils";
+import { Anthropic as R } from "@anthropic-ai/sdk";
 import { OpenAI as U } from "openai";
-import { Worker as x } from "worker_threads";
-import { fileURLToPath as j } from "url";
-import { join as S, dirname as T } from "path";
-import { canDiarization as R } from "./asr.mjs";
-import { createWorker as L } from "tesseract.js";
+import { fileURLToPath as z } from "url";
+import { join as N, dirname as L } from "path";
+import { spawn as g, execSync as C } from "node:child_process";
+import { mkdtempSync as J, rmSync as W } from "node:fs";
+import k from "node:fs/promises";
+import _, { join as x } from "node:path";
+import { createWorker as D } from "tesseract.js";
 import "./embedder.mjs";
-import * as N from "cheerio";
-import { $ as D, $Sync as C } from "@ztimson/node-utils";
-class q {
+import * as I from "cheerio";
+import { $ as F, $Sync as H } from "@ztimson/node-utils";
+class E {
 }
-class W extends q {
-  constructor(s, e, t) {
-    super(), this.ai = s, this.apiToken = e, this.model = t, this.client = new O({ apiKey: e });
+class G extends E {
+  constructor(r, e, t) {
+    super(), this.ai = r, this.apiToken = e, this.model = t, this.client = new R({ apiKey: e });
   }
   client;
-  toStandard(s) {
+  toStandard(r) {
     const e = Date.now(), t = [];
-    for (let c of s)
-      if (typeof c.content == "string")
-        t.push({ timestamp: e, ...c });
+    for (let l of r)
+      if (typeof l.content == "string")
+        t.push({ timestamp: e, ...l });
       else {
-        const r = c.content?.filter((n) => n.type == "text").map((n) => n.text).join(`
+        const n = l.content?.filter((s) => s.type == "text").map((s) => s.text).join(`
 
 `);
-        r && t.push({ timestamp: e, role: c.role, content: r }), c.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 l = t.findLast((o) => o.id == n.tool_use_id);
-            l && (l[n.is_error ? "error" : "content"] = n.content);
+        n && t.push({ timestamp: e, role: l.role, content: n }), l.content.forEach((s) => {
+          if (s.type == "tool_use")
+            t.push({ timestamp: e, role: "tool", id: s.id, name: s.name, args: s.input, content: void 0 });
+          else if (s.type == "tool_result") {
+            const o = t.findLast((c) => c.id == s.tool_use_id);
+            o && (o[s.is_error ? "error" : "content"] = s.content);
           }
         });
       }
     return t;
   }
-  fromStandard(s) {
-    for (let e = 0; e < s.length; e++)
-      if (s[e].role == "tool") {
-        const t = s[e];
-        s.splice(
+  fromStandard(r) {
+    for (let e = 0; e < r.length; e++)
+      if (r[e].role == "tool") {
+        const t = r[e];
+        r.splice(
           e,
           1,
           { role: "assistant", content: [{ type: "tool_use", id: t.id, name: t.name, input: t.args }] },
           { role: "user", content: [{ type: "tool_result", tool_use_id: t.id, is_error: !!t.error, content: t.error || t.content }] }
         ), e++;
       }
-    return s.map(({ timestamp: e, ...t }) => t);
+    return r.map(({ timestamp: e, ...t }) => t);
   }
-  ask(s, e = {}) {
+  ask(r, e = {}) {
     const t = new AbortController();
-    return Object.assign(new Promise(async (c) => {
-      let r = this.fromStandard([...e.history || [], { role: "user", content: s, timestamp: Date.now() }]);
-      const n = e.tools || this.ai.options.llm?.tools || [], l = {
+    return Object.assign(new Promise(async (l) => {
+      let n = this.fromStandard([...e.history || [], { role: "user", content: r, timestamp: Date.now() }]);
+      const s = e.tools || this.ai.options.llm?.tools || [], o = {
         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 || "",
         temperature: e.temperature || this.ai.options.llm?.temperature || 0.7,
-        tools: n.map((m) => ({
+        tools: s.map((m) => ({
           name: m.name,
           description: m.description,
           input_schema: {
             type: "object",
-            properties: m.args ? k(m.args, (i, d) => ({ ...d, required: void 0 })) : {},
+            properties: m.args ? j(m.args, (i, d) => ({ ...d, required: void 0 })) : {},
             required: m.args ? Object.entries(m.args).filter((i) => i[1].required).map((i) => i[0]) : []
           },
           fn: void 0
         })),
-        messages: r,
+        messages: n,
         stream: !!e.stream
       };
-      let o, a = !0;
+      let c, a = !0;
       do {
-        if (o = await this.client.messages.create(l).catch((i) => {
+        if (c = await this.client.messages.create(o).catch((i) => {
           throw i.message += `
 
 Messages:
-${JSON.stringify(r, null, 2)}`, i;
+${JSON.stringify(n, null, 2)}`, i;
         }), e.stream) {
           a ? a = !1 : e.stream({ text: `
 
-` }), o.content = [];
-          for await (const i of o) {
+` }), c.content = [];
+          for await (const i of c) {
             if (t.signal.aborted) break;
             if (i.type === "content_block_start")
-              i.content_block.type === "text" ? o.content.push({ type: "text", text: "" }) : i.content_block.type === "tool_use" && o.content.push({ type: "tool_use", id: i.content_block.id, name: i.content_block.name, input: "" });
+              i.content_block.type === "text" ? c.content.push({ type: "text", text: "" }) : i.content_block.type === "tool_use" && c.content.push({ type: "tool_use", id: i.content_block.id, name: i.content_block.name, input: "" });
             else if (i.type === "content_block_delta")
               if (i.delta.type === "text_delta") {
                 const d = i.delta.text;
-                o.content.at(-1).text += d, e.stream({ text: d });
-              } else i.delta.type === "input_json_delta" && (o.content.at(-1).input += i.delta.partial_json);
+                c.content.at(-1).text += d, e.stream({ text: d });
+              } else i.delta.type === "input_json_delta" && (c.content.at(-1).input += i.delta.partial_json);
             else if (i.type === "content_block_stop") {
-              const d = o.content.at(-1);
-              d.input != null && (d.input = d.input ? g(d.input, {}) : {});
+              const d = c.content.at(-1);
+              d.input != null && (d.input = d.input ? w(d.input, {}) : {});
             } else if (i.type === "message_stop")
               break;
           }
         }
-        const m = o.content.filter((i) => i.type === "tool_use");
+        const m = c.content.filter((i) => i.type === "tool_use");
         if (m.length && !t.signal.aborted) {
-          r.push({ role: "assistant", content: o.content });
+          n.push({ role: "assistant", content: c.content });
           const i = await Promise.all(m.map(async (d) => {
-            const p = n.find(_("name", d.name));
+            const p = s.find(T("name", d.name));
             if (e.stream && e.stream({ tool: d.name }), !p) return { tool_use_id: d.id, is_error: !0, content: "Tool not found" };
             try {
               const u = await p.fn(d.input, e?.stream, this.ai);
@@ -112,45 +115,45 @@ ${JSON.stringify(r, null, 2)}`, i;
               return { type: "tool_result", tool_use_id: d.id, is_error: !0, content: u?.message || u?.toString() || "Unknown" };
             }
           }));
-          r.push({ role: "user", content: i }), l.messages = r;
+          n.push({ role: "user", content: i }), o.messages = n;
         }
-      } while (!t.signal.aborted && o.content.some((m) => m.type === "tool_use"));
-      r.push({ role: "assistant", content: o.content.filter((m) => m.type == "text").map((m) => m.text).join(`
+      } while (!t.signal.aborted && c.content.some((m) => m.type === "tool_use"));
+      n.push({ role: "assistant", content: c.content.filter((m) => m.type == "text").map((m) => m.text).join(`
 
-`) }), r = this.toStandard(r), e.stream && e.stream({ done: !0 }), e.history && e.history.splice(0, e.history.length, ...r), c(r.at(-1)?.content);
+`) }), n = this.toStandard(n), e.stream && e.stream({ done: !0 }), e.history && e.history.splice(0, e.history.length, ...n), l(n.at(-1)?.content);
     }), { abort: () => t.abort() });
   }
 }
-class w extends q {
-  constructor(s, e, t, c) {
-    super(), this.ai = s, this.host = e, this.token = t, this.model = c, this.client = new U(M({
+class S extends E {
+  constructor(r, e, t, l) {
+    super(), this.ai = r, this.host = e, this.token = t, this.model = l, this.client = new U(M({
       baseURL: e,
       apiKey: t
     }));
   }
   client;
-  toStandard(s) {
-    for (let e = 0; e < s.length; e++) {
-      const t = s[e];
+  toStandard(r) {
+    for (let e = 0; e < r.length; e++) {
+      const t = r[e];
       if (t.role === "assistant" && t.tool_calls) {
-        const c = t.tool_calls.map((r) => ({
+        const l = t.tool_calls.map((n) => ({
           role: "tool",
-          id: r.id,
-          name: r.function.name,
-          args: g(r.function.arguments, {}),
+          id: n.id,
+          name: n.function.name,
+          args: w(n.function.arguments, {}),
           timestamp: t.timestamp
         }));
-        s.splice(e, 1, ...c), e += c.length - 1;
+        r.splice(e, 1, ...l), e += l.length - 1;
       } else if (t.role === "tool" && t.content) {
-        const c = s.find((r) => t.tool_call_id == r.id);
-        c && (t.content.includes('"error":') ? c.error = t.content : c.content = t.content), s.splice(e, 1), e--;
+        const l = r.find((n) => t.tool_call_id == n.id);
+        l && (t.content.includes('"error":') ? l.error = t.content : l.content = t.content), r.splice(e, 1), e--;
       }
-      s[e]?.timestamp || (s[e].timestamp = Date.now());
+      r[e]?.timestamp || (r[e].timestamp = Date.now());
     }
-    return s;
+    return r;
   }
-  fromStandard(s) {
-    return s.reduce((e, t) => {
+  fromStandard(r) {
+    return r.reduce((e, t) => {
       if (t.role === "tool")
         e.push({
           role: "assistant",
@@ -164,31 +167,31 @@ class w extends q {
           content: t.error || t.content
         });
       else {
-        const { timestamp: c, ...r } = t;
-        e.push(r);
+        const { timestamp: l, ...n } = t;
+        e.push(n);
       }
       return e;
     }, []);
   }
-  ask(s, e = {}) {
+  ask(r, e = {}) {
     const t = new AbortController();
-    return Object.assign(new Promise(async (c, r) => {
+    return Object.assign(new Promise(async (l, n) => {
       e.system && e.history?.[0]?.role != "system" && e.history?.splice(0, 0, { role: "system", content: e.system, timestamp: Date.now() });
-      let n = this.fromStandard([...e.history || [], { role: "user", content: s, timestamp: Date.now() }]);
-      const l = e.tools || this.ai.options.llm?.tools || [], o = {
+      let s = this.fromStandard([...e.history || [], { role: "user", content: r, timestamp: Date.now() }]);
+      const o = e.tools || this.ai.options.llm?.tools || [], c = {
         model: e.model || this.model,
-        messages: n,
+        messages: s,
         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: l.map((i) => ({
+        tools: o.map((i) => ({
           type: "function",
           function: {
             name: i.name,
             description: i.description,
             parameters: {
               type: "object",
-              properties: i.args ? k(i.args, (d, p) => ({ ...p, required: void 0 })) : {},
+              properties: i.args ? j(i.args, (d, p) => ({ ...p, required: void 0 })) : {},
               required: i.args ? Object.entries(i.args).filter((d) => d[1].required).map((d) => d[0]) : []
             }
           }
@@ -196,11 +199,11 @@ class w extends q {
       };
       let a, m = !0;
       do {
-        if (a = await this.client.chat.completions.create(o).catch((d) => {
+        if (a = await this.client.chat.completions.create(c).catch((d) => {
           throw d.message += `
 
 Messages:
-${JSON.stringify(n, null, 2)}`, d;
+${JSON.stringify(s, null, 2)}`, d;
         }), e.stream) {
           m ? m = !1 : e.stream({ text: `
 
@@ -212,28 +215,28 @@ ${JSON.stringify(n, null, 2)}`, d;
         }
         const i = a.choices[0].message.tool_calls || [];
         if (i.length && !t.signal.aborted) {
-          n.push(a.choices[0].message);
+          s.push(a.choices[0].message);
           const d = await Promise.all(i.map(async (p) => {
-            const u = l?.find(_("name", p.function.name));
+            const u = o?.find(T("name", p.function.name));
             if (e.stream && e.stream({ tool: p.function.name }), !u) return { role: "tool", tool_call_id: p.id, content: '{"error": "Tool not found"}' };
             try {
-              const f = g(p.function.arguments, {}), y = await u.fn(f, e.stream, this.ai);
+              const f = w(p.function.arguments, {}), y = await u.fn(f, e.stream, this.ai);
               return { role: "tool", tool_call_id: p.id, content: b(y) };
             } catch (f) {
               return { role: "tool", tool_call_id: p.id, content: b({ error: f?.message || f?.toString() || "Unknown" }) };
             }
           }));
-          n.push(...d), o.messages = n;
+          s.push(...d), c.messages = s;
         }
       } while (!t.signal.aborted && a.choices?.[0]?.message?.tool_calls?.length);
-      n.push({ role: "assistant", content: a.choices[0].message.content || "" }), n = this.toStandard(n), e.stream && e.stream({ done: !0 }), e.history && e.history.splice(0, e.history.length, ...n), c(n.at(-1)?.content);
+      s.push({ role: "assistant", content: a.choices[0].message.content || "" }), s = this.toStandard(s), e.stream && e.stream({ done: !0 }), e.history && e.history.splice(0, e.history.length, ...s), l(s.at(-1)?.content);
     }), { abort: () => t.abort() });
   }
 }
-class z {
-  constructor(s) {
-    this.ai = s, s.options.llm?.models && Object.entries(s.options.llm.models).forEach(([e, t]) => {
-      this.defaultModel || (this.defaultModel = e), t.proto == "anthropic" ? this.models[e] = new W(this.ai, t.token, e) : t.proto == "ollama" ? this.models[e] = new w(this.ai, t.host, "not-needed", e) : t.proto == "openai" && (this.models[e] = new w(this.ai, t.host || null, t.token, e));
+class B {
+  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 G(this.ai, t.token, e) : t.proto == "ollama" ? this.models[e] = new S(this.ai, t.host, "not-needed", e) : t.proto == "openai" && (this.models[e] = new S(this.ai, t.host || null, t.token, e));
     });
   }
   defaultModel;
@@ -244,25 +247,25 @@ class z {
    * @param {LLMRequest} options Configuration options and chat history
    * @returns {{abort: () => void, response: Promise<string>}} Function to abort response and chat history
    */
-  ask(s, e = {}) {
+  ask(r, e = {}) {
     const t = e.model || this.defaultModel;
     if (!this.models[t]) throw new Error(`Model does not exist: ${t}`);
-    let c = () => {
+    let l = () => {
     };
-    return Object.assign(new Promise(async (r) => {
+    return Object.assign(new Promise(async (n) => {
       if (e.history || (e.history = []), e.memory) {
         e.system = (e.system || "") + `
 You have passive persistent memory never make any mention of your memory capabilities and what you can/cannot remember
 `;
-        const l = async (a, m, i = 50) => {
+        const o = async (a, m, i = 50) => {
           const [d, p] = await Promise.all([
             m ? this.embedding(m) : Promise.resolve(null),
             a ? this.embedding(a) : Promise.resolve(null)
           ]);
           return (e.memory || []).map((u) => ({ ...u, score: d ? this.cosineSimilarity(u.embeddings[0], d[0].embedding) : 1 })).filter((u) => u.score >= 0.8).map((u) => ({ ...u, score: p ? this.cosineSimilarity(u.embeddings[1], p[0].embedding) : u.score })).filter((u) => u.score >= 0.2).toSorted((u, f) => u.score - f.score).slice(0, i);
-        }, o = await l(s);
-        o.length && e.history.push({ role: "assistant", content: `Things I remembered:
-` + o.map((a) => `${a.owner}: ${a.fact}`).join(`
+        }, c = await o(r);
+        c.length && e.history.push({ role: "assistant", content: `Things I remembered:
+` + c.map((a) => `${a.owner}: ${a.fact}`).join(`
 `) }), e.tools = [...e.tools || [], {
           name: "read_memory",
           description: "Check your long-term memory for more information",
@@ -273,30 +276,30 @@ You have passive persistent memory never make any mention of your memory capabil
           },
           fn: (a) => {
             if (!a.subject && !a.query) throw new Error("Either a subject or query argument is required");
-            return l(a.query, a.subject, a.limit || 5);
+            return o(a.query, a.subject, a.limit || 5);
           }
         }];
       }
-      const n = await this.models[t].ask(s, e);
+      const s = await this.models[t].ask(r, e);
       if (e.memory) {
-        const l = e.history?.findIndex((o) => o.role == "assistant" && o.content.startsWith("Things I remembered:"));
-        l != null && l >= 0 && e.history?.splice(l, 1);
+        const o = e.history?.findIndex((c) => c.role == "assistant" && c.content.startsWith("Things I remembered:"));
+        o != null && o >= 0 && e.history?.splice(o, 1);
       }
       if (e.compress || e.memory) {
-        let l = null;
+        let o = null;
         if (e.compress)
-          l = await this.ai.language.compressHistory(e.history, e.compress.max, e.compress.min, e), e.history.splice(0, e.history.length, ...l.history);
+          o = await this.ai.language.compressHistory(e.history, e.compress.max, e.compress.min, e), e.history.splice(0, e.history.length, ...o.history);
         else {
-          const o = e.history?.findLastIndex((a) => a.role == "user") ?? -1;
-          l = await this.ai.language.compressHistory(o != -1 ? e.history.slice(o) : e.history, 0, 0, e);
+          const c = e.history?.findLastIndex((a) => a.role == "user") ?? -1;
+          o = await this.ai.language.compressHistory(c != -1 ? e.history.slice(c) : e.history, 0, 0, e);
         }
         if (e.memory) {
-          const o = e.memory.filter((a) => !l.memory.some((m) => this.cosineSimilarity(a.embeddings[1], m.embeddings[1]) > 0.8)).concat(l.memory);
-          e.memory.splice(0, e.memory.length, ...o);
+          const c = e.memory.filter((a) => !o.memory.some((m) => this.cosineSimilarity(a.embeddings[1], m.embeddings[1]) > 0.8)).concat(o.memory);
+          e.memory.splice(0, e.memory.length, ...c);
         }
       }
-      return r(n);
-    }), { abort: c });
+      return n(s);
+    }), { abort: l });
   }
   /**
    * Compress chat history to reduce context size
@@ -306,24 +309,24 @@ You have passive persistent memory never make any mention of your memory capabil
    * @param {LLMRequest} options LLM options
    * @returns {Promise<LLMMessage[]>} New chat history will summary at index 0
    */
-  async compressHistory(s, e, t, c) {
-    if (this.estimateTokens(s) < e) return { history: s, memory: [] };
-    let r = 0, n = 0;
-    for (let u of s.toReversed())
-      if (n += this.estimateTokens(u.content), n < t) r++;
+  async compressHistory(r, e, t, l) {
+    if (this.estimateTokens(r) < e) return { history: r, memory: [] };
+    let n = 0, s = 0;
+    for (let u of r.toReversed())
+      if (s += this.estimateTokens(u.content), s < t) n++;
       else break;
-    if (s.length <= r) return { history: s, memory: [] };
-    const l = s[0].role == "system" ? s[0] : null, o = r == 0 ? [] : s.slice(-r), a = (r == 0 ? s : s.slice(0, -r)).filter((u) => u.role === "assistant" || u.role === "user"), m = await this.json(a.map((u) => `${u.role}: ${u.content}`).join(`
+    if (r.length <= n) return { history: r, memory: [] };
+    const o = r[0].role == "system" ? r[0] : null, c = n == 0 ? [] : r.slice(-n), a = (n == 0 ? r : r.slice(0, -n)).filter((u) => u.role === "assistant" || u.role === "user"), m = await this.json(a.map((u) => `${u.role}: ${u.content}`).join(`
 
 `), "{summary: string, facts: [[subject, fact]]}", {
       system: "Create the smallest summary possible, no more than 500 tokens. Create a list of NEW facts (split by subject [pro]noun and fact) about what you learned from this conversation that you didn't already know or get from a tool call or system prompt. Focus only on new information about people, topics, or facts. Avoid generating facts about the AI.",
-      model: c?.model,
-      temperature: c?.temperature || 0.3
+      model: l?.model,
+      temperature: l?.temperature || 0.3
     }), i = /* @__PURE__ */ new Date(), d = await Promise.all((m?.facts || [])?.map(async ([u, f]) => {
       const y = await Promise.all([this.embedding(u), this.embedding(`${u}: ${f}`)]);
       return { owner: u, fact: f, embeddings: [y[0][0].embedding, y[1][0].embedding], timestamp: i };
-    })), p = [{ role: "assistant", content: `Conversation Summary: ${m?.summary}`, timestamp: Date.now() }, ...o];
-    return l && p.splice(0, 0, l), { history: p, memory: d };
+    })), p = [{ role: "assistant", content: `Conversation Summary: ${m?.summary}`, timestamp: Date.now() }, ...c];
+    return o && p.splice(0, 0, o), { history: p, memory: d };
   }
   /**
    * Compare the difference between embeddings (calculates the angle between two vectors)
@@ -331,13 +334,13 @@ You have passive persistent memory never make any mention of your memory capabil
    * @param {number[]} v2 Second embedding / vector for comparison
    * @returns {number} Similarity values 0-1: 0 = unique, 1 = identical
    */
-  cosineSimilarity(s, e) {
-    if (s.length !== e.length) throw new Error("Vectors must be same length");
-    let t = 0, c = 0, r = 0;
-    for (let l = 0; l < s.length; l++)
-      t += s[l] * e[l], c += s[l] * s[l], r += e[l] * e[l];
-    const n = Math.sqrt(c) * Math.sqrt(r);
-    return n === 0 ? 0 : t / n;
+  cosineSimilarity(r, e) {
+    if (r.length !== e.length) throw new Error("Vectors must be same length");
+    let t = 0, l = 0, n = 0;
+    for (let o = 0; o < r.length; o++)
+      t += r[o] * e[o], l += r[o] * r[o], n += e[o] * e[o];
+    const s = Math.sqrt(l) * Math.sqrt(n);
+    return s === 0 ? 0 : t / s;
   }
   /**
    * Chunk text into parts for AI digestion
@@ -346,26 +349,26 @@ You have passive persistent memory never make any mention of your memory capabil
    * @param {number} overlapTokens Includes previous X tokens to provide continuity to AI (In addition to max tokens)
    * @returns {string[]} Chunked strings
    */
-  chunk(s, e = 500, t = 50) {
-    const c = (o, a = "") => o ? Object.entries(o).flatMap(([m, i]) => {
+  chunk(r, e = 500, t = 50) {
+    const l = (c, a = "") => c ? Object.entries(c).flatMap(([m, i]) => {
       const d = a ? `${a}${isNaN(+m) ? `.${m}` : `[${m}]`}` : m;
-      return typeof i == "object" && !Array.isArray(i) ? c(i, d) : `${d}: ${Array.isArray(i) ? i.join(", ") : i}`;
-    }) : [], n = (typeof s == "object" ? c(s) : s.split(`
-`)).flatMap((o) => [...o.split(/\s+/).filter(Boolean), `
-`]), l = [];
-    for (let o = 0; o < n.length; ) {
-      let a = "", m = o;
-      for (; m < n.length; ) {
-        const d = a + (a ? " " : "") + n[m];
+      return typeof i == "object" && !Array.isArray(i) ? l(i, d) : `${d}: ${Array.isArray(i) ? i.join(", ") : i}`;
+    }) : [], s = (typeof r == "object" ? l(r) : r.split(`
+`)).flatMap((c) => [...c.split(/\s+/).filter(Boolean), `
+`]), o = [];
+    for (let c = 0; c < s.length; ) {
+      let a = "", m = c;
+      for (; m < s.length; ) {
+        const d = a + (a ? " " : "") + s[m];
         if (this.estimateTokens(d.replace(/\s*\n\s*/g, `
 `)) > e && a) break;
         a = d, m++;
       }
       const i = a.replace(/\s*\n\s*/g, `
 `).trim();
-      i && l.push(i), o = Math.max(m - t, m === o ? o + 1 : m);
+      i && o.push(i), c = Math.max(m - t, m === c ? c + 1 : m);
     }
-    return l;
+    return o;
   }
   /**
    * Create a vector representation of a string
@@ -373,31 +376,48 @@ You have passive persistent memory never make any mention of your memory capabil
    * @param {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
    * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
    */
-  async embedding(s, e = {}) {
-    let { maxTokens: t = 500, overlapTokens: c = 50 } = e;
-    const r = (o) => new Promise((a, m) => {
-      const i = new x(S(T(j(import.meta.url)), "embedder.js")), d = ({ embedding: u }) => {
-        i.terminate(), a(u);
-      }, p = (u) => {
-        i.terminate(), m(u);
-      };
-      i.on("message", d), i.on("error", p), i.on("exit", (u) => {
-        u !== 0 && m(new Error(`Worker exited with code ${u}`));
-      }), i.postMessage({ text: o, model: this.ai.options?.embedder || "bge-small-en-v1.5", modelDir: this.ai.options.path });
-    }), n = this.chunk(s, t, c), l = [];
-    for (let o = 0; o < n.length; o++) {
-      const a = n[o], m = await r(a);
-      l.push({ index: o, embedding: m, text: a, tokens: this.estimateTokens(a) });
-    }
-    return l;
+  embedding(r, e = {}) {
+    let { maxTokens: t = 500, overlapTokens: l = 50 } = e, n = !1;
+    const s = () => {
+      n = !0;
+    }, o = (a) => new Promise((m, i) => {
+      if (n) return i(new Error("Aborted"));
+      const d = [
+        N(L(z(import.meta.url)), "embedder.js"),
+        this.ai.options.path,
+        this.ai.options?.embedder || "bge-small-en-v1.5"
+      ], p = g("node", d, { stdio: ["pipe", "pipe", "ignore"] });
+      p.stdin.write(a), p.stdin.end();
+      let u = "";
+      p.stdout.on("data", (f) => u += f.toString()), p.on("close", (f) => {
+        if (n) return i(new Error("Aborted"));
+        if (f === 0)
+          try {
+            const y = JSON.parse(u);
+            m(y.embedding);
+          } catch {
+            i(new Error("Failed to parse embedding output"));
+          }
+        else
+          i(new Error(`Embedder process exited with code ${f}`));
+      }), p.on("error", i);
+    }), c = (async () => {
+      const a = this.chunk(r, t, l), m = [];
+      for (let i = 0; i < a.length && !n; i++) {
+        const d = a[i], p = await o(d);
+        m.push({ index: i, embedding: p, text: d, tokens: this.estimateTokens(d) });
+      }
+      return m;
+    })();
+    return Object.assign(c, { abort: s });
   }
   /**
    * Estimate variable as tokens
    * @param history Object to size
    * @returns {number} Rough token count
    */
-  estimateTokens(s) {
-    const e = JSON.stringify(s);
+  estimateTokens(r) {
+    const e = JSON.stringify(r);
     return Math.ceil(e.length / 4 * 1.2);
   }
   /**
@@ -406,10 +426,10 @@ You have passive persistent memory never make any mention of your memory capabil
    * @param {string} searchTerms Multiple search terms to check against target
    * @returns {{avg: number, max: number, similarities: number[]}} Similarity values 0-1: 0 = unique, 1 = identical
    */
-  fuzzyMatch(s, ...e) {
+  fuzzyMatch(r, ...e) {
     if (e.length < 2) throw new Error("Requires at least 2 strings to compare");
-    const t = (n, l = 10) => n.toLowerCase().split("").map((o, a) => o.charCodeAt(0) * (a + 1) % l / l).slice(0, l), c = t(s), r = e.map((n) => t(n)).map((n) => this.cosineSimilarity(c, n));
-    return { avg: r.reduce((n, l) => n + l, 0) / r.length, max: Math.max(...r), similarities: r };
+    const t = (s, o = 10) => s.toLowerCase().split("").map((c, a) => c.charCodeAt(0) * (a + 1) % o / o).slice(0, o), l = t(r), n = e.map((s) => t(s)).map((s) => this.cosineSimilarity(l, s));
+    return { avg: n.reduce((s, o) => s + o, 0) / n.length, max: Math.max(...n), similarities: n };
   }
   /**
    * Ask a question with JSON response
@@ -418,15 +438,15 @@ You have passive persistent memory never make any mention of your memory capabil
    * @param {LLMRequest} options Configuration options and chat history
    * @returns {Promise<{} | {} | RegExpExecArray | null>}
    */
-  async json(s, e, t) {
-    let c = await this.ask(s, { ...t, system: (t?.system ? `${t.system}
+  async json(r, e, t) {
+    let l = await this.ask(r, { ...t, system: (t?.system ? `${t.system}
 ` : "") + `Only respond using a JSON code block matching this schema:
 \`\`\`json
 ${e}
 \`\`\`` });
-    if (!c) return {};
-    const r = /```(?:.+)?\s*([\s\S]*?)```/.exec(c), n = r ? r[1].trim() : c;
-    return g(n, {});
+    if (!l) return {};
+    const n = /```(?:.+)?\s*([\s\S]*?)```/.exec(l), s = n ? n[1].trim() : l;
+    return w(s, {});
   }
   /**
    * Create a summary of some text
@@ -435,34 +455,112 @@ ${e}
    * @param options LLM request options
    * @returns {Promise<string>} Summary
    */
-  summarize(s, e, t) {
-    return this.ask(s, { system: `Generate a brief summary <= ${e} tokens. Output nothing else`, temperature: 0.3, ...t });
+  summarize(r, e, t) {
+    return this.ask(r, { system: `Generate a brief summary <= ${e} tokens. Output nothing else`, temperature: 0.3, ...t });
   }
 }
-class I {
-  constructor(s) {
-    this.ai = s;
+class K {
+  constructor(r) {
+    this.ai = r, r.options.whisper && (this.whisperModel = r.options.asr?.endsWith(".bin") ? r.options.asr : r.options.asr + ".bin", this.downloadAsrModel()), this.pyannote = `
+import sys
+import json
+import os
+from pyannote.audio import Pipeline
+
+os.environ['TORCH_HOME'] = r"${r.options.path}"
+pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token="${r.options.hfToken}")
+output = pipeline(sys.argv[1])
+
+segments = []
+for turn, speaker in output.speaker_diarization:
+    segments.append({"start": turn.start, "end": turn.end, "speaker": speaker})
+
+print(json.dumps(segments))
+`;
+  }
+  downloads = {};
+  pyannote;
+  whisperModel;
+  runAsr(r, e = {}) {
+    let t;
+    const l = new Promise((n, s) => {
+      this.downloadAsrModel(e.model).then((o) => {
+        let c = "";
+        const a = [e.diarization ? "-owts" : "-nt", "-m", o, "-f", r];
+        t = g(this.ai.options.whisper, a, { stdio: ["ignore", "pipe", "ignore"] }), t.on("error", (m) => s(m)), t.stdout.on("data", (m) => c += m.toString()), t.on("close", (m) => {
+          if (m === 0)
+            if (e.diarization)
+              try {
+                n(JSON.parse(c));
+              } catch {
+                s(new Error("Failed to parse whisper JSON"));
+              }
+            else
+              n(c.trim() || null);
+          else
+            s(new Error(`Exit code ${m}`));
+        });
+      });
+    });
+    return Object.assign(l, { abort: () => t?.kill("SIGTERM") });
   }
-  asr(s, e = {}) {
-    const { model: t = this.ai.options.asr || "whisper-base", speaker: c = !1 } = e;
-    let r = !1;
-    const n = () => {
-      r = !0;
+  runDiarization(r) {
+    let e = !1, t = () => {
+      e = !0;
     };
-    let l = new Promise((o, a) => {
-      const m = new x(S(T(j(import.meta.url)), "asr.js")), i = ({ text: p, warning: u, error: f }) => {
-        setTimeout(() => m.terminate(), 1e3), !r && (f ? a(new Error(f)) : (u && console.warn(u), o(p)));
-      }, d = (p) => {
-        setTimeout(() => m.terminate(), 1e3), r || a(p);
-      };
-      m.on("message", i), m.on("error", d), m.on("exit", (p) => {
-        p !== 0 && !r && a(new Error(`Worker exited with code ${p}`));
-      }), m.postMessage({ file: s, model: t, speaker: c, modelDir: this.ai.options.path, token: this.ai.options.hfToken });
+    const l = (s) => new Promise((o) => {
+      const c = g(s, ["-c", "import pyannote.audio"]);
+      c.on("close", (a) => o(a === 0)), c.on("error", () => o(!1));
+    }), n = Promise.all([
+      l("python"),
+      l("python3")
+    ]).then((async ([s, o]) => {
+      if (e) return;
+      if (!s && !o) throw new Error("Pyannote is not installed: pip install pyannote.audio");
+      const c = o ? "python3" : "python";
+      let a = null;
+      return new Promise((m, i) => {
+        if (a = x(J(x(P(), "audio-")), "converted.wav"), C(`ffmpeg -i "${r}" -ar 16000 -ac 1 -f wav "${a}"`, { stdio: "ignore" }), e) return;
+        let d = "";
+        const p = g(c, ["-c", this.pyannote, a]);
+        p.stdout.on("data", (u) => d += u.toString()), p.stderr.on("data", (u) => console.error(u.toString())), p.on("close", (u) => {
+          if (u === 0)
+            try {
+              m(JSON.parse(d));
+            } catch {
+              i(new Error("Failed to parse diarization output"));
+            }
+          else
+            i(new Error(`Python process exited with code ${u}`));
+        }), p.on("error", i), t = () => p.kill("SIGTERM");
+      }).finally(() => {
+        a && W(_.dirname(a), { recursive: !0, force: !0 });
+      });
+    }));
+    return Object.assign(n, { abort: t });
+  }
+  combineSpeakerTranscript(r, e) {
+    const t = /* @__PURE__ */ new Map();
+    let l = 0;
+    e.forEach((c) => {
+      t.has(c.speaker) || t.set(c.speaker, ++l);
     });
-    if (e.speaker == "id") {
-      if (!this.ai.language.defaultModel) throw new Error("Configure an LLM for advanced ASR speaker detection");
-      l = l.then(async (o) => {
-        if (!o) return o;
+    const n = [];
+    let s = -1, o = "";
+    return r.transcription.forEach((c) => {
+      const a = c.offsets.from / 1e3, m = e.find((d) => a >= d.start && a <= d.end), i = m ? t.get(m.speaker) : 1;
+      i !== s ? (o && n.push(`[Speaker ${s}]: ${o.trim()}`), s = i, o = c.text) : o += " " + c.text;
+    }), o && n.push(`[Speaker ${s}]: ${o.trim()}`), n.join(`
+`);
+  }
+  asr(r, e = {}) {
+    if (!this.ai.options.whisper) throw new Error("Whisper not configured");
+    const t = this.runAsr(r, { model: e.model, diarization: !!e.diarization }), l = e.diarization ? this.runDiarization(r) : Promise.resolve(null), n = () => {
+      t.abort(), l?.abort?.();
+    }, s = Promise.all([t, l]).then(async ([o, c]) => {
+      if (!e.diarization) return o;
+      if (o = this.combineSpeakerTranscript(o, c), e.diarization === "id") {
+        if (!this.ai.language.defaultModel) throw new Error("Configure an LLM for advanced ASR speaker detection");
         let a = this.ai.language.chunk(o, 500, 0);
         a.length > 4 && (a = [...a.slice(0, 3), a.at(-1)]);
         const m = await this.ai.language.json(a.join(`
@@ -470,37 +568,41 @@ class I {
           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(m).forEach(([i, d]) => {
-          o = o.replaceAll(`[Speaker ${i}]`, `[${d}]`);
-        }), o;
-      });
-    }
-    return Object.assign(l, { abort: n });
+        Object.entries(m).forEach(([i, d]) => o = o.replaceAll(`[Speaker ${i}]`, `[${d}]`));
+      }
+      return o;
+    });
+    return Object.assign(s, { abort: n });
+  }
+  async downloadAsrModel(r = this.whisperModel) {
+    if (!this.ai.options.whisper) throw new Error("Whisper not configured");
+    r.endsWith(".bin") || (r += ".bin");
+    const e = _.join(this.ai.options.path, r);
+    return await k.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 k.writeFile(e, t), delete this.downloads[r], e)), this.downloads[r]);
   }
-  canDiarization = () => R().then((s) => !!s);
 }
-class J {
-  constructor(s) {
-    this.ai = s;
+class V {
+  constructor(r) {
+    this.ai = r;
   }
   /**
    * Convert image to text using Optical Character Recognition
    * @param {string} path Path to image
    * @returns {AbortablePromise<string | null>} Promise of extracted text with abort method
    */
-  ocr(s) {
+  ocr(r) {
     let e;
-    const t = new Promise(async (c) => {
-      e = await L(this.ai.options.ocr || "eng", 2, { cachePath: this.ai.options.path });
-      const { data: r } = await e.recognize(s);
-      await e.terminate(), c(r.text.trim() || null);
+    const t = new Promise(async (l) => {
+      e = await D(this.ai.options.ocr || "eng", 2, { cachePath: this.ai.options.path });
+      const { data: n } = await e.recognize(r);
+      await e.terminate(), l(n.text.trim() || null);
     });
     return Object.assign(t, { abort: () => e?.terminate() });
   }
 }
-class re {
-  constructor(s) {
-    this.options = s, s.path || (s.path = $.tmpdir()), process.env.TRANSFORMERS_CACHE = s.path, this.audio = new I(this), this.language = new z(this), this.vision = new J(this);
+class ue {
+  constructor(r) {
+    this.options = r, r.path || (r.path = $.tmpdir()), process.env.TRANSFORMERS_CACHE = r.path, this.audio = new K(this), this.language = new B(this), this.vision = new V(this);
   }
   /** Audio processing AI */
   audio;
@@ -509,38 +611,38 @@ class re {
   /** Vision processing AI */
   vision;
 }
-const H = {
+const Y = {
   name: "cli",
   description: "Use the command line interface, returns any output",
   args: { command: { type: "string", description: "Command to run", required: !0 } },
-  fn: (h) => D`${h.command}`
-}, ne = {
+  fn: (h) => F`${h.command}`
+}, pe = {
   name: "get_datetime",
   description: "Get current UTC date / time",
   args: {},
   fn: async () => (/* @__PURE__ */ new Date()).toUTCString()
-}, oe = {
+}, he = {
   name: "exec",
   description: "Run code/scripts",
   args: {
     language: { type: "string", description: "Execution language", enum: ["cli", "node", "python"], required: !0 },
     code: { type: "string", description: "Code to execute", required: !0 }
   },
-  fn: async (h, s, e) => {
+  fn: async (h, r, e) => {
     try {
       switch (h.type) {
         case "bash":
-          return await H.fn({ command: h.code }, s, e);
+          return await Y.fn({ command: h.code }, r, e);
         case "node":
-          return await F.fn({ code: h.code }, s, e);
+          return await Q.fn({ code: h.code }, r, e);
         case "python":
-          return await G.fn({ code: h.code }, s, e);
+          return await X.fn({ code: h.code }, r, e);
       }
     } catch (t) {
       return { error: t?.message || t.toString() };
     }
   }
-}, ie = {
+}, fe = {
   name: "fetch",
   description: "Make HTTP request to URL",
   args: {
@@ -549,25 +651,25 @@ const H = {
     headers: { type: "object", description: "HTTP headers to send", default: {} },
     body: { type: "object", description: "HTTP body to send" }
   },
-  fn: (h) => new E({ url: h.url, headers: h.headers }).request({ method: h.method || "GET", body: h.body })
-}, F = {
+  fn: (h) => new O({ url: h.url, headers: h.headers }).request({ method: h.method || "GET", body: h.body })
+}, Q = {
   name: "exec_javascript",
   description: "Execute commonjs javascript",
   args: {
     code: { type: "string", description: "CommonJS javascript", required: !0 }
   },
   fn: async (h) => {
-    const s = P(null), e = await A({ console: s }, h.code, !0).catch((t) => s.output.error.push(t));
-    return { ...s.output, return: e, stdout: void 0, stderr: void 0 };
+    const r = A(null), e = await q({ console: r }, h.code, !0).catch((t) => r.output.error.push(t));
+    return { ...r.output, return: e, stdout: void 0, stderr: void 0 };
   }
-}, G = {
+}, X = {
   name: "exec_javascript",
   description: "Execute commonjs javascript",
   args: {
     code: { type: "string", description: "CommonJS javascript", required: !0 }
   },
-  fn: async (h) => ({ result: C`python -c "${h.code}"` })
-}, ae = {
+  fn: async (h) => ({ result: H`python -c "${h.code}"` })
+}, ye = {
   name: "read_webpage",
   description: "Extract clean, structured content from a webpage. Use after web_search to read specific URLs",
   args: {
@@ -575,26 +677,26 @@ const H = {
     focus: { type: "string", description: 'Optional: What aspect to focus on (e.g., "pricing", "features", "contact info")' }
   },
   fn: async (h) => {
-    const s = 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 = N.load(s);
+    const r = await fetch(h.url, { headers: { "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64)" } }).then((s) => s.text()).catch((s) => {
+      throw new Error(`Failed to fetch: ${s.message}`);
+    }), e = I.load(r);
     e('script, style, nav, footer, header, aside, iframe, noscript, [role="navigation"], [role="banner"], .ad, .ads, .cookie, .popup').remove();
     const t = {
       title: e('meta[property="og:title"]').attr("content") || e("title").text() || "",
       description: e('meta[name="description"]').attr("content") || e('meta[property="og:description"]').attr("content") || ""
     };
-    let c = "";
-    const r = ["article", "main", '[role="main"]', ".content", ".post", ".entry", "body"];
-    for (const n of r) {
-      const l = e(n).first();
-      if (l.length && l.text().trim().length > 200) {
-        c = l.text();
+    let l = "";
+    const n = ["article", "main", '[role="main"]', ".content", ".post", ".entry", "body"];
+    for (const s of n) {
+      const o = e(s).first();
+      if (o.length && o.text().trim().length > 200) {
+        l = o.text();
         break;
       }
     }
-    return c || (c = e("body").text()), c = c.replace(/\s+/g, " ").trim().slice(0, 8e3), { url: h.url, title: t.title.trim(), description: t.description.trim(), content: c, focus: h.focus };
+    return l || (l = e("body").text()), l = l.replace(/\s+/g, " ").trim().slice(0, 8e3), { url: h.url, title: t.title.trim(), description: t.description.trim(), content: l, focus: h.focus };
   }
-}, ce = {
+}, ge = {
   name: "web_search",
   description: "Use duckduckgo (anonymous) to find find relevant online resources. Returns a list of URLs that works great with the `read_webpage` tool",
   args: {
@@ -602,33 +704,32 @@ const H = {
     length: { type: "string", description: "Number of results to return", default: 5 }
   },
   fn: async (h) => {
-    const s = await fetch(`https://html.duckduckgo.com/html/?q=${encodeURIComponent(h.query)}`, {
+    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((r) => r.text());
+    }).then((n) => n.text());
     let e, t = /<a .*?href="(.+?)".+?<\/a>/g;
-    const c = new v();
-    for (; (e = t.exec(s)) !== null; ) {
-      let r = /uddg=(.+)&amp?/.exec(decodeURIComponent(e[1]))?.[1];
-      if (r && (r = decodeURIComponent(r)), r && c.add(r), c.size >= (h.length || 5)) break;
+    const l = new v();
+    for (; (e = t.exec(r)) !== null; ) {
+      let n = /uddg=(.+)&amp?/.exec(decodeURIComponent(e[1]))?.[1];
+      if (n && (n = decodeURIComponent(n)), n && l.add(n), l.size >= (h.length || 5)) break;
     }
-    return c;
+    return l;
   }
 };
 export {
-  re as Ai,
-  W as Anthropic,
-  I as Audio,
-  H as CliTool,
-  ne as DateTimeTool,
-  oe as ExecTool,
-  ie as FetchTool,
-  F as JSTool,
-  q as LLMProvider,
-  w as OpenAi,
-  G as PythonTool,
-  ae as ReadWebpageTool,
-  J as Vision,
-  ce as WebSearchTool,
-  R as canDiarization
+  ue as Ai,
+  G as Anthropic,
+  K as Audio,
+  Y as CliTool,
+  pe as DateTimeTool,
+  he as ExecTool,
+  fe as FetchTool,
+  Q as JSTool,
+  E as LLMProvider,
+  S as OpenAi,
+  X as PythonTool,
+  ye as ReadWebpageTool,
+  V as Vision,
+  ge as WebSearchTool
 };
 //# sourceMappingURL=index.mjs.map