eddie-jekyll 0.2.4 → 0.4.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 4b25c316b5227f2d8ba9dcf8bcbf23b3ae3ea9a31916c98c9d301bdf482c535d
-  data.tar.gz: eea298e14f549efa67b7befc5980058ff0f0510f7146a66d62b95e8df943c8b1
+  metadata.gz: 0fb05e7b2291be54f498bc2e2df6c8b75689e2fecd1e62f069b05712d78882f5
+  data.tar.gz: 402f2f4a6569e9117f7acf584ff4d6966c05cf3601ce2909c3960b489d3180fd
 SHA512:
-  metadata.gz: 6e782ffb5bac0ccb716b23f947ae1ac8b349ea883a83017cc98dbecd01e35461ec8f4d6c0dba20578398a2851cdf4beb86942ed5853ad841ac24c165019deb9c
-  data.tar.gz: b6054c718bdf4c9bf3255d79b405e2c185a08b685f303e7deff8009c503ad5b942a22915563cbb71706985e2592511d35236c5ec50c1616cc5f858b1ea095272
+  metadata.gz: 28c2cf42975d69ddaa552acb7ece03fe9a7f390d5eac410b35b49387c3f527f85c2b4a01e288f557b681b2a509e40b073deb6d174e5e6296547a0e9466683652
+  data.tar.gz: a6159ac2597c2f1e294ef466d6d1e663cb0f753728e1fa97047a6992987891d5f998cf85d9b5752f61e27bdcf49ca40f7e28e53eb994b0313aece4f36e98aa17

data/assets/eddie-agent-worker.js

@@ -0,0 +1,536 @@
+// SPDX-License-Identifier: GPL-3.0-only
+// Generated by widget/build.sh from widget/src/eddie-agent-worker.js and widget/src/lib/*.js; edit those instead.
+"use strict";
+const EddieLib = {};
+// SPDX-License-Identifier: GPL-3.0-only
+
+// Agent helpers shared by the widget and the agent worker: model selection,
+// prompts, plan parsing, evidence assembly and answer post-processing.
+// Pure functions; no WebLLM, no DOM.
+
+(function (factory) {
+  const api = factory();
+  if (typeof module === "object" && module && module.exports) {
+    module.exports = api;
+  } else if (typeof EddieLib === "object" && EddieLib) {
+    Object.assign(EddieLib, api);
+  } else {
+    globalThis.EddieLib = Object.assign(globalThis.EddieLib || {}, api);
+  }
+})(function () {
+  "use strict";
+
+  const NOHIT = "The site doesn't cover that.";
+  const NOHIT_RE = /\bthe site (?:doesn['’]t|does not|didn['’]t|did not) cover (?:that|this|it)\.?/gi;
+
+  const AGENT_MODEL_SIZES = {
+    "Qwen3.5-0.8B": 0.4e9,
+    "Qwen3.5-2B": 1.2e9,
+    "Qwen3.5-4B": 2.3e9,
+  };
+  const TWO_GIB = 2 * 1024 * 1024 * 1024;
+
+  const PLAN_SCHEMA = {
+    type: "object",
+    properties: {
+      queries: { type: "array", items: { type: "string" }, minItems: 1, maxItems: 3 },
+    },
+    required: ["queries"],
+  };
+
+  function planPrompt(site) {
+    return `You write search queries for a site search engine. The site is ${site}. Reply with JSON only: {"queries": ["..."]}. Give 1 to 3 different short keyword queries (2 to 5 words each, no punctuation) that a site search engine would match against page text. Each query must be different. Do not answer the question.`;
+  }
+
+  function answerPrompt(site) {
+    return `You answer visitor questions about ${site} using only the numbered sources below the question. Answer the question directly in the first sentence. Write 1 to 3 sentences in your own words; never repeat a source's wording. End each sentence with the numbers of the sources it comes from, like [2] or [1][3]. Never cite a number that is not in the list. Do not add calculations or inferences that are not in the sources. If no source answers the question, your entire reply is: ${NOHIT} Never mix that sentence with an answer. Never use outside knowledge.`;
+  }
+
+  /** Strip the WebLLM variant suffix to get the family name shown to visitors. */
+  function baseModelId(id) {
+    return String(id).replace(/-q\d+f(16|32)_\d+-MLC$/i, "");
+  }
+
+  function agentModelBytes(id) {
+    const base = baseModelId(id);
+    return Object.prototype.hasOwnProperty.call(AGENT_MODEL_SIZES, base) ? AGENT_MODEL_SIZES[base] : null;
+  }
+
+  /**
+   * Choose the WebLLM model id.
+   * opts: { mode: "auto"|"quality"|<id>, maxBufferSize, isMobile, hasF16 }
+   */
+  function selectAgentModel(opts) {
+    const o = opts || {};
+    const mode = (o.mode || "auto").trim();
+    const suffix = o.hasF16 ? "-q4f16_1-MLC" : "-q4f32_1-MLC";
+    let base;
+    if (mode === "auto") {
+      const big = Number(o.maxBufferSize) >= TWO_GIB && !o.isMobile;
+      base = big ? "Qwen3.5-2B" : "Qwen3.5-0.8B";
+    } else if (mode === "quality") {
+      base = "Qwen3.5-2B";
+    } else {
+      return { id: mode, base: baseModelId(mode), sizeBytes: agentModelBytes(mode), explicit: true };
+    }
+    const id = base + suffix;
+    return { id, base, sizeBytes: AGENT_MODEL_SIZES[base], explicit: false };
+  }
+
+  function isMobileDevice(nav) {
+    const n = nav || {};
+    if (n.userAgentData && typeof n.userAgentData.mobile === "boolean") {
+      return n.userAgentData.mobile;
+    }
+    return /Mobi|Android|iPhone|iPad|iPod|Windows Phone/i.test(n.userAgent || "");
+  }
+
+  /** Remove <think>…</think> blocks; a dangling <think> loses only the tag. */
+  function stripThink(text) {
+    if (!text) return "";
+    let out = String(text).replace(/<think>[\s\S]*?<\/think>/g, "");
+    out = out.replace(/<think>/g, "").replace(/<\/think>/g, "");
+    return out.trim();
+  }
+
+  /**
+   * Display text for a partial stream: complete think blocks removed, and
+   * anything after an unclosed <think> hidden until it closes.
+   */
+  function visibleStreamText(partial) {
+    if (!partial) return "";
+    let out = String(partial).replace(/<think>[\s\S]*?<\/think>/g, "");
+    const open = out.indexOf("<think>");
+    if (open >= 0) out = out.substring(0, open);
+    return out.replace(/^\s+/, "");
+  }
+
+  function extractJsonObject(text) {
+    const s = String(text);
+    const start = s.indexOf("{");
+    const end = s.lastIndexOf("}");
+    if (start < 0 || end <= start) return null;
+    try {
+      return JSON.parse(s.substring(start, end + 1));
+    } catch (_) {
+      return null;
+    }
+  }
+
+  function cleanQuery(q) {
+    return String(q)
+      .replace(/\/?no_think/gi, "")
+      .replace(/[\s"'`*]+$/g, "")
+      .replace(/^[\s"'`*\-\d.)]+/g, "")
+      .replace(/\s+/g, " ")
+      .trim();
+  }
+
+  /** Parse the planner reply into 1..3 distinct queries; falls back to the question. */
+  function parsePlan(text, question) {
+    const cleaned = stripThink(text);
+    const obj = extractJsonObject(cleaned);
+    const raw = obj && Array.isArray(obj.queries) ? obj.queries : [];
+    const seen = new Set();
+    const out = [];
+    for (const item of raw) {
+      if (typeof item !== "string") continue;
+      const q = cleanQuery(item);
+      if (q.length < 2 || q.length > 80) continue;
+      const key = q.toLowerCase();
+      if (seen.has(key)) continue;
+      seen.add(key);
+      out.push(q);
+      if (out.length === 3) break;
+    }
+    if (out.length === 0 && question && question.trim()) {
+      out.push(question.trim());
+    }
+    return out;
+  }
+
+  function urlKey(url) {
+    return String(url || "").replace(/#.*$/, "").replace(/\/+$/, "").toLowerCase();
+  }
+
+  /**
+   * Round-robin merge of several result lists, deduplicated by URL, at most
+   * `max` items. Each result keeps its own fields.
+   */
+  function mergeEvidence(lists, max) {
+    const limit = max == null ? 6 : max;
+    const seen = new Set();
+    const out = [];
+    const arrays = (lists || []).map((l) => (Array.isArray(l) ? l : []));
+    const longest = arrays.reduce((n, l) => Math.max(n, l.length), 0);
+    for (let i = 0; i < longest && out.length < limit; i++) {
+      for (const list of arrays) {
+        if (out.length >= limit) break;
+        const r = list[i];
+        if (!r || !r.url) continue;
+        const key = urlKey(r.url);
+        if (seen.has(key)) continue;
+        seen.add(key);
+        out.push(r);
+      }
+    }
+    return out;
+  }
+
+  /** Cut to `max` characters at a word boundary, with an ellipsis. */
+  function truncateText(text, max) {
+    const limit = max == null ? 700 : max;
+    const s = String(text || "").replace(/\s+/g, " ").trim();
+    if (s.length <= limit) return s;
+    const cut = s.lastIndexOf(" ", limit - 1);
+    return s.substring(0, cut > limit * 0.6 ? cut : limit - 1).trim() + "…";
+  }
+
+  /** `[n] title (url)\ntext` blocks joined by blank lines. */
+  function formatEvidence(items, maxChars) {
+    return (items || [])
+      .map((e, i) => `[${i + 1}] ${e.title || e.url} (${e.url})\n${truncateText(e.text || e.snippet || "", maxChars)}`)
+      .join("\n\n");
+  }
+
+  function sourcesPrompt(items, question, maxChars) {
+    return `Sources:\n\n${formatEvidence(items, maxChars)}\n\nQuestion: ${question}`;
+  }
+
+  /**
+   * Post-process a raw model answer against the evidence list.
+   * Returns { answer, citations: [{n, url, title}], nohit }.
+   */
+  function postProcessAnswer(raw, evidence) {
+    const ev = Array.isArray(evidence) ? evidence : [];
+    let text = stripThink(raw);
+    // **[1]** -> [1]
+    text = text.replace(/\*\*\s*((?:\[\s*\d+(?:\s*,\s*\d+)*\s*\]\s*)+)\*\*/g, "$1");
+    // Lines that are only citation markers belong to the previous line.
+    const lines = text.split(/\r?\n/);
+    const merged = [];
+    for (const line of lines) {
+      const t = line.trim();
+      if (t && /^(?:\[\s*\d+(?:\s*,\s*\d+)*\s*\]\s*)+$/.test(t)) {
+        let j = merged.length - 1;
+        while (j >= 0 && merged[j].trim() === "") j--;
+        if (j >= 0) {
+          // Append only markers the previous line does not already carry.
+          const prev = merged[j];
+          const fresh = (t.match(/\[\s*\d+(?:\s*,\s*\d+)*\s*\]/g) || []).filter((m) => !prev.includes(m.replace(/\s+/g, "")));
+          merged.length = j + 1;
+          if (fresh.length) merged[j] = prev.replace(/\s+$/, "") + " " + fresh.join("");
+          continue;
+        }
+      }
+      merged.push(line);
+    }
+    text = merged.join("\n");
+
+    // Drop the fallback sentence when anything else remains.
+    const withoutFallback = text.replace(NOHIT_RE, "").replace(/[ \t]+\n/g, "\n").trim();
+    const residue = withoutFallback
+      .replace(/\[\s*\d+(?:\s*,\s*\d+)*\s*\]/g, "")
+      .replace(/^\s*(?:yes|no)\b/i, "")
+      .replace(/[\s.,;:!?'"*-]+/g, "");
+    const onlyFallback = residue === "";
+    if (onlyFallback) {
+      return { answer: NOHIT, citations: [], nohit: true };
+    }
+    text = withoutFallback;
+
+    // Map [n] citations to evidence; drop out-of-range ones.
+    const cited = [];
+    const seen = new Set();
+    text = text.replace(/\[\s*(\d+(?:\s*,\s*\d+)*)\s*\]/g, (_, body) => {
+      const nums = body.split(",").map((x) => Number(x.trim()));
+      let out = "";
+      for (const n of nums) {
+        if (n >= 1 && n <= ev.length) {
+          out += `[${n}]`;
+          if (!seen.has(n)) {
+            seen.add(n);
+            cited.push(n);
+          }
+        }
+      }
+      return out;
+    });
+    // Collapse duplicate adjacent markers ("[1][1]") and tidy whitespace.
+    text = text.replace(/(\[\d+\])(?:\s*\1)+/g, "$1");
+    text = text.replace(/[ \t]{2,}/g, " ").replace(/ +([.,;:!?])/g, "$1").replace(/\n{3,}/g, "\n\n").trim();
+    const citations = cited
+      .sort((a, b) => a - b)
+      .map((n) => ({ n, url: ev[n - 1].url, title: ev[n - 1].title || ev[n - 1].url }));
+    return { answer: text, citations, nohit: text === "" };
+  }
+
+  return {
+    NOHIT,
+    AGENT_MODEL_SIZES,
+    PLAN_SCHEMA,
+    planPrompt,
+    answerPrompt,
+    sourcesPrompt,
+    baseModelId,
+    agentModelBytes,
+    selectAgentModel,
+    isMobileDevice,
+    stripThink,
+    visibleStreamText,
+    parsePlan,
+    mergeEvidence,
+    truncateText,
+    formatEvidence,
+    postProcessAnswer,
+  };
+});
+
+// SPDX-License-Identifier: GPL-3.0-only
+
+// Eddie agent worker (module worker, created on the first "Ask").
+//
+// Runs WebLLM in the worker; retrieval stays in the widget, which owns the
+// search worker and passes evidence in. widget/build.sh concatenates
+// widget/src/lib/agent.js ahead of this file (EddieLib).
+//
+// Protocol (main thread -> worker):
+//   load  {model}
+//   plan  {requestId, question, site}
+//   ask   {requestId, question, site, evidence: [{title, url, text}]}
+//   abort {requestId?}
+// (worker -> main thread):
+//   progress    {text, progress}
+//   loaded      {model, loadMs}
+//   plan_result {requestId, queries, ms}
+//   token       {requestId, text}
+//   done        {requestId, answer, citations: [{n, url, title}], nohit, usage}
+//   aborted     {requestId}
+//   error       {requestId?, message}
+
+"use strict";
+
+const WEBLLM_URL = "https://esm.run/@mlc-ai/web-llm@0.2.84";
+const EVIDENCE_CHARS = 700;
+
+const lib = EddieLib;
+
+let webllm = null;
+let engine = null;
+let modelId = null;
+let loading = null;
+let active = null; // { requestId, aborted }
+let queue = Promise.resolve();
+
+self.onmessage = function (e) {
+  const msg = e.data || {};
+  switch (msg.type) {
+    case "load":
+      load(msg);
+      break;
+    case "plan":
+      enqueue(() => plan(msg), msg.requestId);
+      break;
+    case "ask":
+      enqueue(() => ask(msg), msg.requestId);
+      break;
+    case "abort":
+      abort(msg);
+      break;
+    default:
+      postError(msg.requestId, `unknown message type ${String(msg.type)}`);
+  }
+};
+
+function enqueue(fn, requestId) {
+  queue = queue
+    .then(() => {
+      console.debug("eddie agent worker: start", requestId);
+      return fn();
+    })
+    .catch((err) => postError(requestId, describe(err)))
+    .then(() => console.debug("eddie agent worker: end", requestId));
+}
+
+async function load(msg) {
+  const model = String(msg.model || "");
+  if (!model) {
+    postError(undefined, "load: model is required");
+    return;
+  }
+  if (engine && modelId === model) {
+    self.postMessage({ type: "loaded", model, loadMs: 0, cached: true });
+    return;
+  }
+  if (loading) {
+    try {
+      await loading;
+    } catch (_) {
+      // fall through and try again
+    }
+    if (engine && modelId === model) {
+      self.postMessage({ type: "loaded", model, loadMs: 0, cached: true });
+      return;
+    }
+  }
+  loading = (async () => {
+    const t0 = performance.now();
+    if (!webllm) {
+      self.postMessage({ type: "progress", text: "Loading the WebLLM runtime…", progress: 0 });
+      webllm = await import(WEBLLM_URL);
+    }
+    if (engine) {
+      try {
+        await engine.unload();
+      } catch (_) {
+        // ignore
+      }
+      engine = null;
+      modelId = null;
+    }
+    const created = await webllm.CreateMLCEngine(model, {
+      initProgressCallback: (p) => {
+        self.postMessage({
+          type: "progress",
+          text: p && p.text ? p.text : "Loading model…",
+          progress: p && typeof p.progress === "number" ? p.progress : null,
+        });
+      },
+    });
+    engine = created;
+    modelId = model;
+    self.postMessage({ type: "loaded", model, loadMs: Math.round(performance.now() - t0) });
+  })();
+  try {
+    await loading;
+  } catch (err) {
+    engine = null;
+    modelId = null;
+    postError(undefined, describe(err));
+  } finally {
+    loading = null;
+  }
+}
+
+function requireEngine() {
+  if (!engine) throw new Error("model not loaded");
+}
+
+async function plan(msg) {
+  requireEngine();
+  const question = String(msg.question || "").trim();
+  const site = String(msg.site || "this website");
+  const t0 = performance.now();
+  const reply = await engine.chat.completions.create({
+    messages: [
+      { role: "system", content: lib.planPrompt(site) },
+      { role: "user", content: question },
+    ],
+    temperature: 0,
+    max_tokens: 100,
+    response_format: { type: "json_object", schema: JSON.stringify(lib.PLAN_SCHEMA) },
+    extra_body: { enable_thinking: false },
+  });
+  const content = reply && reply.choices && reply.choices[0] && reply.choices[0].message ? reply.choices[0].message.content : "";
+  const queries = lib.parsePlan(content, question);
+  self.postMessage({ type: "plan_result", requestId: msg.requestId, queries, ms: Math.round(performance.now() - t0) });
+}
+
+async function ask(msg) {
+  requireEngine();
+  const requestId = msg.requestId;
+  const question = String(msg.question || "").trim();
+  const site = String(msg.site || "this website");
+  const evidence = Array.isArray(msg.evidence) ? msg.evidence.filter((e) => e && e.url) : [];
+  if (evidence.length === 0) {
+    self.postMessage({
+      type: "done",
+      requestId,
+      answer: lib.NOHIT,
+      citations: [],
+      nohit: true,
+      raw: "",
+      usage: { ttftMs: 0, totalMs: 0, tps: null, completionTokens: 0 },
+    });
+    return;
+  }
+  active = { requestId, aborted: false };
+  const t0 = performance.now();
+  let first = 0;
+  let text = "";
+  let usage = null;
+  try {
+    const stream = await engine.chat.completions.create({
+      messages: [
+        { role: "system", content: lib.answerPrompt(site) },
+        { role: "user", content: lib.sourcesPrompt(evidence, question, EVIDENCE_CHARS) },
+      ],
+      stream: true,
+      stream_options: { include_usage: true },
+      temperature: 0,
+      frequency_penalty: 0.5,
+      presence_penalty: 0,
+      max_tokens: 220,
+      extra_body: { enable_thinking: false },
+    });
+    // Never break out of this loop: WebLLM releases its generation lock at
+    // the end of the async generator, and an early exit skips that release,
+    // hanging every later completion. After interruptGenerate() the stream
+    // ends by itself within one decode step; drop the tokens until then.
+    for await (const chunk of stream) {
+      if (active.aborted) continue;
+      const delta = chunk && chunk.choices && chunk.choices[0] && chunk.choices[0].delta ? chunk.choices[0].delta.content : null;
+      if (delta) {
+        if (!first) first = performance.now();
+        text += delta;
+        self.postMessage({ type: "token", requestId, text: delta });
+      }
+      if (chunk && chunk.usage) usage = chunk.usage;
+    }
+  } finally {
+    const wasAborted = active && active.aborted;
+    active = null;
+    if (wasAborted) {
+      self.postMessage({ type: "aborted", requestId });
+      return;
+    }
+  }
+  const processed = lib.postProcessAnswer(text, evidence);
+  const totalMs = Math.round(performance.now() - t0);
+  self.postMessage({
+    type: "done",
+    requestId,
+    answer: processed.answer,
+    citations: processed.citations,
+    nohit: processed.nohit,
+    raw: text,
+    usage: {
+      ttftMs: first ? Math.round(first - t0) : totalMs,
+      totalMs,
+      tps: usage && usage.extra && typeof usage.extra.decode_tokens_per_s === "number" ? Math.round(usage.extra.decode_tokens_per_s) : null,
+      completionTokens: usage ? usage.completion_tokens : null,
+    },
+  });
+}
+
+function abort(msg) {
+  console.debug("eddie agent worker: abort", msg.requestId, active ? active.requestId : null);
+  if (!active) return;
+  if (msg.requestId != null && msg.requestId !== active.requestId) return;
+  active.aborted = true;
+  try {
+    if (engine) engine.interruptGenerate();
+  } catch (err) {
+    console.warn("eddie agent: interrupt failed", err);
+  }
+}
+
+function postError(requestId, message) {
+  self.postMessage({ type: "error", requestId: requestId == null ? undefined : requestId, message });
+}
+
+function describe(err) {
+  if (err == null) return "unknown error";
+  if (typeof err === "string") return err;
+  return err.message || String(err);
+}