@bootstrapware/answers 0.2.11 → 0.2.13

package/AGENTS.md

@@ -227,18 +227,18 @@ Questions are limited to 1,000 Unicode code points. A longer draft stays in the
 Hosted indexes the main body of each page, split by headings; navigation, headers, footers, and cookie banners are dropped. The model writes a short answer from the best-matching passages and must cite them. It does not use outside knowledge. `list_answers_capabilities` returns `sourceAdvice` with the same rules.
 
 1. Make sure the fact is stated in a page body, not only implied by a homepage card. A pricing page, a support page, and one page per product work well.
-2. Keep each figure next to what it prices (`Hosted is $19.99/month`). An answer whose price or number is not in a cited passage is rejected unless it follows from cited figures (a single-step percent-off or percent-of of a cited amount, rounded to the nearest cent, or a sequential stack total with each line rounded to the nearest cent).
+2. Keep each figure next to what it prices (`Hosted is $19.99/month`). An answer whose price or number is not in a cited passage is rejected unless it follows from cited figures (a single-step percent-off or percent-of of a cited amount, rounded to the nearest cent, or a sequential stack total with each line rounded to the nearest cent). The same number written with a comma, dot, space, or apostrophe as the thousands separator is that figure (`2,500`, `2.500`, and `2 500`). A comma with one or two decimal digits is that decimal (`19,99` and `$19.99`).
 3. After `add_answers_url_source`, wait for the index job, then use the dashboard retrieval check to see which passages a question finds. Ask through the widget with a test key to see the written answer.
 4. For facts you want stated exactly (prices, refund policy, support contact), add a custom Q&A in the dashboard. Custom Q&A is always considered, and a matching entry is answered first. Hosted entries are indexed on save, usually within about a minute; the dashboard knowledge list shows when each is indexed. With BYO, an entry changed through `upsertQa`, `removeQa`, or `setQa` applies from the next question.
 5. When the pages are silent on a yes/no question (certification, compliance, encryption), the answer says the site does not state it and gives the support contact if a source has one. Add the support contact as a custom Q&A so that works. Unanswered with the fallback is the correct outcome when nothing covers the question. Add a URL or a custom Q&A. Do not send the question through MCP.
 
 ## Knowledge, grounding, and provider
 
-Hosted knowledge kinds are website root, sitemap, single public URL, PDF, DOCX, TXT, Markdown, and custom Q&A (a question and its answer). MCP can add a public URL and reindex it. Files, custom Q&A, conversations, and unanswered question text stay on the dashboard.
+Hosted knowledge kinds are website root, sitemap, single public URL, PDF, DOCX, TXT, Markdown, and custom Q&A (a question and its answer). A website or sitemap crawl also fetches `/llms.txt` and `/llms-full.txt` on that host when `robots.txt` allows them. Text and Markdown are indexed. A missing file is ignored and is not an index error. BYO `refresh()` does the same for each origin in `urls` or the sitemap. These agent indexes stay below product pages when both match a question. MCP can add a public URL and reindex it. Files, custom Q&A, conversations, and unanswered question text stay on the dashboard.
 
 Every Hosted source is visitor-readable. Do not upload private runbooks or secrets. Crawl is public `http` and `https` only. Private hosts are rejected. `robots.txt` is honored. A failed crawl leaves the last good index in place. A page that is only navigation chrome is rejected and does not become a citation.
 
-Retrieval combines keyword and embedding search over the active index. When the question names a product whose pages live under a URL path, those pages are preferred. Custom Q&A is always considered, and a matching entry is answered first. The model answers in 1 to 4 sentences in the visitor's language, citing passages it was given; citations are ids from the retrieved index, and the model does not invent URLs. A draft that declines, cites an unretrieved passage, states a figure that is not in a cited passage and does not follow from cited figures by the rules above, or writes code not quoted from a passage is retried once and then returns `unanswered` with the configured fallback. A derived total that depends on billing order is rejected unless the visitor set the order, the answer says order matters, or the answer itemizes list and discounted line items that sum to the total. Off-topic questions and prompt-injection attempts are declined before retrieval. How-to questions get the documented steps and a citation to the docs page; the model does not write new code. There is no model-memory fallback. Earlier turns are used only for clear follow-ups ("what about Chat?"); a standalone question is answered fresh.
+Retrieval combines keyword and embedding search over the active index. When the question names a product whose pages live under a URL path, those pages are preferred, and passages on that path that mention the distinctive terms in the question are reserved for the model. Custom Q&A is always considered, and a matching entry is answered first. The model answers in 1 to 4 sentences in the visitor's language, citing passages it was given; citations are ids from the retrieved index, and the model does not invent URLs. A draft that declines, cites an unretrieved passage, states a figure that is not in a cited passage and does not follow from cited figures by the rules above, or writes code not quoted from a passage is retried once and then returns `unanswered` with the configured fallback. A derived total that depends on billing order is rejected unless the visitor set the order, the answer says order matters, or the answer itemizes list and discounted line items that sum to the total. Off-topic questions and prompt-injection attempts are declined before retrieval. How-to questions get the documented steps and a citation to the docs page; the model does not write new code. There is no model-memory fallback. Earlier turns are used only for clear follow-ups ("what about Chat?"); a standalone question is answered fresh.
 
 Unanswered observations store normalized question text even when conversation history is off, so the dashboard can group them. MCP does not return that text.
 

package/CHANGELOG.md

@@ -1,5 +1,16 @@
 # Changelog
 
+## 0.2.13
+
+- Grounding treats a thousands group as one figure. `2,500`, `2.500`, `2 500` (including a narrow no-break space), and `2'500` are the same number, and `19,99` is the same decimal as `19.99`. A one-digit group no longer matches a different grouped number that happens to end in the same digits (`1,500` is not `2,500`).
+- Citation markers in an answer are removed when the comma inside the brackets is an Arabic comma, an ideographic comma, or a fullwidth comma (`[2، 5]`), including Arabic-Indic digits.
+- A website or sitemap crawl, and BYO `refresh()` for each origin already being fetched, also indexes `/llms.txt` and `/llms-full.txt` when the host publishes them as text or Markdown. A missing file is ignored and is not listed as a skipped page.
+
+## 0.2.12
+
+- Hosted and BYO retrieval now reserve named-path passages that mention the distinctive terms in the question (`stream` on an Answers page), so a generic product-page crowd cannot hide a documented yes/no.
+- The custom Q&A lane no longer treats “does X support…” as a match for a contact-support entry, and embedding-only Q&A needs a closer match than before.
+
 ## 0.2.11
 
 - Retrieval prefers pages under a product path the question named, so another product's similar vocabulary cannot crowd out a documented yes/no (Chat SSE vs an Answers capability page).

package/README.md

@@ -59,7 +59,7 @@ export const POST = (request: Request) => answers.handler(request);
 
 The index is built on the first question and kept in memory. Call `answers.refresh()` after your content changes (it returns `{ pages, chunks, skipped }`; `skipped` lists every page left out and why, including a URL that redirects to a different page, and `onRefresh` receives the same report after the automatic first build), or save `answers.index.toJSON()` and pass it back as `snapshot` to skip re-embedding on restart. Pass `retrieve` to use your own vector store instead.
 
-Answers come only from the indexed passages. A draft that cites something it did not retrieve, or states a price or number that is not in a cited passage and does not follow from cited figures (a single-step percent-off or percent-of of a cited amount, rounded to the nearest cent, or a sequential stack total with each line rounded to the nearest cent), is rejected and the visitor gets the fallback. A stack total that depends on billing order is rejected unless the visitor set the order, the answer says order matters, or the answer itemizes lines that sum to that total. Custom Q&A is always considered, and a matching entry is answered first. When the sources are silent on a yes/no question, the answer says the site does not state it (with the support contact if a source has one). How-to questions cite the docs page; no new code is written.
+Answers come only from the indexed passages. `refresh()` also fetches `/llms.txt` and `/llms-full.txt` for each origin in `urls` or the sitemap when the host publishes them as text or Markdown, and ignores a missing file. A draft that cites something it did not retrieve, or states a price or number that is not in a cited passage and does not follow from cited figures (a single-step percent-off or percent-of of a cited amount, rounded to the nearest cent, or a sequential stack total with each line rounded to the nearest cent), is rejected and the visitor gets the fallback. The same number with a comma, dot, space, or apostrophe as the thousands separator counts as the figure in the source, and a comma decimal such as `19,99` counts as `$19.99`. A stack total that depends on billing order is rejected unless the visitor set the order, the answer says order matters, or the answer itemizes lines that sum to that total. Custom Q&A is always considered, and a matching entry is answered first. When the sources are silent on a yes/no question, the answer says the site does not state it (with the support contact if a source has one). How-to questions cite the docs page; no new code is written.
 
 Custom Q&A can change at runtime without a refresh: call `await answers.upsertQa({ id, question, answer })`, `answers.removeQa(id)`, or `answers.setQa([...])` from your save handler. Only the changed entry is embedded and it answers on the next question. Relevant custom Q&A are always offered to the model, even when many page passages outrank them, and that includes setups that pass their own `retrieve`.
 

package/dist/server.d.ts

@@ -330,6 +330,9 @@ type RefreshReport = {
 };
 /** Same page: host (ignoring www and scheme) and path, ignoring a trailing slash, query, and fragment. */
 declare function samePage(left: string, right: string): boolean;
+/** `/llms.txt` and `/llms-full.txt` on an origin already being indexed. */
+declare function agentIndexUrls(origin: string): string[];
+declare function isAgentIndexUrl(raw: string): boolean;
 /**
  * BYO Answers backend. Runs on your server with your OpenAI key; questions and knowledge stay on
  * your infrastructure. Pair with createByoAdapter in the browser.
@@ -395,6 +398,33 @@ declare function tokenize(text: string): string[];
 declare function firstPathSegment(url?: string | null): string | null;
 /** Path segments the visitor named that also appear as product roots in the candidate URLs. */
 declare function namedPathQuery(query: string, urls: Array<string | null | undefined>): Set<string>;
+/**
+ * Product or subject the visitor named ("Does Answers support…", capitalized "Answers").
+ * Used so the product name is not treated as the distinctive evidence term.
+ */
+declare function subjectTokens(query: string): Set<string>;
+/** Auxiliaries in "does X support/offer Y?" — they match contact-support Q&A, not the capability. */
+declare function capabilityVerbs(query: string): Set<string>;
+/** Query tokens that should pull the passage that actually states the fact (not the product name). */
+declare function distinctiveTokens(query: string, named?: ReadonlySet<string>): string[];
+declare function contentTokenOverlap(text: string, tokens: readonly string[]): number;
+type EvidenceRow = {
+    id: string;
+    url?: string | null;
+    canonical_url?: string | null;
+    title?: string | null;
+    text?: string;
+    excerpt?: string;
+    kind?: string;
+};
+/** How many named-path passages that mention the distinctive terms are reserved for the model. */
+declare const EVIDENCE_LANE = 3;
+/**
+ * Pages under a product path the question named that also mention the distinctive terms
+ * (`stream` on an Answers grounding page). Generic product pages and another product's
+ * similar vocabulary do not qualify.
+ */
+declare function selectNamedEvidence<T extends EvidenceRow>(query: string, rows: readonly T[], lane?: number): T[];
 /**
  * Sort bucket: 0 custom Q&A, 1 pages under a product path the question named, 2 the rest.
  * Empty `named` keeps page order (everyone 1 except Q&A).
@@ -403,7 +433,7 @@ declare function namedPathBucket(url: string | null | undefined, named: Readonly
 /** How many custom Q&A entries are always offered to the model when they look relevant. */
 declare const QA_LANE = 2;
 /** Embedding similarity at which a custom Q&A counts as relevant even with no shared words. */
-declare const QA_LANE_MIN_SIMILARITY = 0.3;
+declare const QA_LANE_MIN_SIMILARITY = 0.55;
 type QaLaneCandidate = {
     id: string;
     title: string;
@@ -421,4 +451,4 @@ declare function withQaLane<T extends {
     id: string;
 }>(ranked: readonly T[], lane: readonly T[], limit: number): T[];
 
-export { type AnswerCitation, type AnswerDeps, type AnswerInput, type AnswerOutput, type AnswerReason, type AnswersBackend, type AnswersBackendOptions, CHUNK_MAX_CHARS, CHUNK_TARGET_CHARS, type ChatMessage, type Chunk, DEFAULT_ANSWER_MODEL, DEFAULT_EMBEDDING_MODEL, type DocumentInput, EMBEDDING_DIMENSIONS, type ExtractedPage, FALLBACK_ANSWER_MODEL, type IndexedEntry, type LlmClient, LlmError, MemoryIndex, type MemoryIndexSnapshot, type OpenAiOptions, type PageSection, type Passage, QA_LANE, QA_LANE_MIN_SIMILARITY, type QaInput, type QaLaneCandidate, type ReasoningEffort, type RefreshReport, type RetrieveQuery, type Retriever, type SkippedPage, type Turn, answerMessages, answerQuestion, authFailure, chunkPage, citationsFor, cosine, createAnswersBackend, createOpenAiClient, decodeCfEmail, decodeEntities, decodeProtectedEmails, extractPage, extractText, figuresIn, firstPathSegment, inferSources, isReasoningModel, matchingQa, namedPathBucket, namedPathQuery, prepareMessages, reciprocalRankFusion, samePage, selectQaLane, sourceBlock, stem, tokenize, verifyAnswer, withQaLane };
+export { type AnswerCitation, type AnswerDeps, type AnswerInput, type AnswerOutput, type AnswerReason, type AnswersBackend, type AnswersBackendOptions, CHUNK_MAX_CHARS, CHUNK_TARGET_CHARS, type ChatMessage, type Chunk, DEFAULT_ANSWER_MODEL, DEFAULT_EMBEDDING_MODEL, type DocumentInput, EMBEDDING_DIMENSIONS, EVIDENCE_LANE, type EvidenceRow, type ExtractedPage, FALLBACK_ANSWER_MODEL, type IndexedEntry, type LlmClient, LlmError, MemoryIndex, type MemoryIndexSnapshot, type OpenAiOptions, type PageSection, type Passage, QA_LANE, QA_LANE_MIN_SIMILARITY, type QaInput, type QaLaneCandidate, type ReasoningEffort, type RefreshReport, type RetrieveQuery, type Retriever, type SkippedPage, type Turn, agentIndexUrls, answerMessages, answerQuestion, authFailure, capabilityVerbs, chunkPage, citationsFor, contentTokenOverlap, cosine, createAnswersBackend, createOpenAiClient, decodeCfEmail, decodeEntities, decodeProtectedEmails, distinctiveTokens, extractPage, extractText, figuresIn, firstPathSegment, inferSources, isAgentIndexUrl, isReasoningModel, matchingQa, namedPathBucket, namedPathQuery, prepareMessages, reciprocalRankFusion, samePage, selectNamedEvidence, selectQaLane, sourceBlock, stem, subjectTokens, tokenize, verifyAnswer, withQaLane };

package/dist/server.js

@@ -504,11 +504,59 @@ function namedPathQuery(query, urls) {
   }
   const named = /* @__PURE__ */ new Set();
   if (!present.size) return named;
-  for (const token of new Set(tokenize(query))) {
+  for (const token of /* @__PURE__ */ new Set([...tokenize(query), ...subjectTokens(query)])) {
     if (present.has(token)) named.add(token);
   }
   return named;
 }
+function subjectTokens(query) {
+  const out = /* @__PURE__ */ new Set();
+  const subject = query.match(/\b(?:does|do|can|is|are)\s+([A-Za-z][\w-]*)\b/i);
+  if (subject) {
+    const [token] = tokenize(subject[1]);
+    if (token) out.add(token);
+  }
+  for (const word of query.match(/\b[A-Z][A-Za-z]{2,}\b/g) ?? []) {
+    const [token] = tokenize(word);
+    if (token) out.add(token);
+  }
+  return out;
+}
+function capabilityVerbs(query) {
+  const out = /* @__PURE__ */ new Set();
+  if (/\b(?:does|do|can)\s+[A-Za-z][\w-]*\s+(?:support|offer|have|include|allow)\b/i.test(query)) {
+    for (const verb of ["support", "offer", "have", "include", "allow"]) out.add(stem(verb));
+  }
+  return out;
+}
+function distinctiveTokens(query, named = /* @__PURE__ */ new Set()) {
+  const skip = /* @__PURE__ */ new Set([...named, ...subjectTokens(query), ...capabilityVerbs(query)]);
+  return [...new Set(tokenize(query))].filter((token) => token.length >= 4 && !skip.has(token));
+}
+function contentTokenOverlap(text, tokens) {
+  if (!tokens.length) return 0;
+  const own = new Set(tokenize(text));
+  let n = 0;
+  for (const token of tokens) if (own.has(token)) n += 1;
+  return n;
+}
+var EVIDENCE_LANE = 3;
+function selectNamedEvidence(query, rows, lane = EVIDENCE_LANE) {
+  if (!rows.length || lane <= 0) return [];
+  const pages = rows.filter((row) => row.kind !== "qa" && row.kind !== "manual");
+  const named = namedPathQuery(
+    query,
+    pages.map((row) => row.url ?? row.canonical_url)
+  );
+  const terms = distinctiveTokens(query, named);
+  if (!terms.length) return [];
+  return pages.map((row) => {
+    const url = row.url ?? row.canonical_url;
+    const overlap = contentTokenOverlap(`${row.title ?? ""} ${row.excerpt ?? row.text ?? ""}`, terms);
+    const bucket = namedPathBucket(url, named);
+    return { row, overlap, bucket };
+  }).filter((item) => item.overlap >= 1 && (!named.size || item.bucket === 1)).sort((a, b) => b.overlap - a.overlap).slice(0, lane).map((item) => item.row);
+}
 function namedPathBucket(url, named, qa = false) {
   if (qa) return 0;
   if (!named.size) return 1;
@@ -516,7 +564,7 @@ function namedPathBucket(url, named, qa = false) {
   return seg && named.has(seg) ? 1 : 2;
 }
 var QA_LANE = 2;
-var QA_LANE_MIN_SIMILARITY = 0.3;
+var QA_LANE_MIN_SIMILARITY = 0.55;
 function selectQaLane(questions, candidates, lane = QA_LANE) {
   if (!candidates.length || lane <= 0) return [];
   const words = (text) => new Set(tokenize(text));
@@ -525,12 +573,18 @@ function selectQaLane(questions, candidates, lane = QA_LANE) {
   for (const t of titles) for (const w of t) df.set(w, (df.get(w) ?? 0) + 1);
   const common = Math.max(2, Math.ceil(candidates.length / 2));
   const asked = new Set(questions.flatMap((q) => [...words(q)]).filter((w) => (df.get(w) ?? 0) < common || candidates.length < 3));
+  const weak = new Set(questions.flatMap((q) => [...capabilityVerbs(q), ...subjectTokens(q)]));
   return candidates.map((c, i) => {
     let shared = 0;
-    for (const w of asked) if (titles[i].has(w)) shared += 1;
+    let strong = 0;
+    for (const w of asked) {
+      if (!titles[i].has(w)) continue;
+      shared += 1;
+      if (!weak.has(w)) strong += 1;
+    }
     const sim = typeof c.similarity === "number" && Number.isFinite(c.similarity) ? c.similarity : 0;
-    return { id: c.id, shared, sim, score: shared + sim * 2 };
-  }).filter((x) => x.shared >= 1 || x.sim >= QA_LANE_MIN_SIMILARITY).sort((a, b) => b.score - a.score).slice(0, lane).map((x) => x.id);
+    return { id: c.id, shared, strong, sim, score: strong + shared * 0.25 + sim * 2 };
+  }).filter((x) => x.strong >= 1 || x.sim >= QA_LANE_MIN_SIMILARITY).sort((a, b) => b.score - a.score).slice(0, lane).map((x) => x.id);
 }
 function withQaLane(ranked, lane, limit) {
   const seen = /* @__PURE__ */ new Set();
@@ -679,14 +733,21 @@ var MemoryIndex = class _MemoryIndex {
     const lexical = this.lexical(`${query.text} ${query.original}`, depth).map(String);
     const vector = query.embedding ? this.vector(query.embedding, depth).map(String) : [];
     const fused = reciprocalRankFusion([lexical, vector], 60, [1, 1.3]);
-    const named = namedPathQuery(`${query.text} ${query.original}`, this.entries.map((e) => e.url));
+    const asked = `${query.text} ${query.original}`;
+    const named = namedPathQuery(asked, this.entries.map((e) => e.url));
+    const terms = distinctiveTokens(asked, named);
     const boosted = fused.map(({ id, score }) => {
       const entry = this.entries[Number(id)];
       const weight = entry.kind === "qa" ? 1.25 : 1;
-      return { id: Number(id), score: score * weight, bucket: namedPathBucket(entry.url, named, entry.kind === "qa") };
-    }).sort((a, b) => a.bucket - b.bucket || b.score - a.score);
+      const overlap = entry.kind === "qa" ? 0 : contentTokenOverlap(`${entry.title} ${entry.content}`, terms);
+      return { id: Number(id), score: score * weight, overlap, bucket: namedPathBucket(entry.url, named, entry.kind === "qa") };
+    }).sort((a, b) => a.bucket - b.bucket || b.overlap - a.overlap || b.score - a.score);
     const ranked = boosted.slice(0, query.limit).map(({ id }) => this.passage(id));
-    return withQaLane(ranked, this.qaLane(query), query.limit);
+    const evidence = selectNamedEvidence(
+      asked,
+      this.entries.map((e, i) => ({ id: String(i), url: e.url, title: e.title, text: e.content, kind: e.kind }))
+    ).map((row) => this.passage(Number(row.id)));
+    return withQaLane(withQaLane(ranked, evidence, query.limit), this.qaLane(query), query.limit);
   }
 };
 function qaEntry(item, fallbackId) {
@@ -916,10 +977,96 @@ ${sourceBlock(passages)}`,
     { role: "user", content: user }
   ];
 }
-var FIGURE = /(?:[$€£]\s?\d[\d,]*(?:\.\d+)?)|(?:\b\d[\d,]*(?:\.\d+)?\s?%)|(?:\b\d+\.\d+\b)|(?:\b\d{2,}[\d,]*\b)/g;
-var normalizeFigure = (f) => f.replace(/[\s,]/g, "").replace(/^[$€£]/, "").replace(/%$/, "").replace(/\.0+$/, "");
+function isGroupSpace(ch) {
+  return ch === "\xA0" || ch === "\u202F" || ch === "\u2007" || ch === "\u2009";
+}
+function stripInt(digits) {
+  return digits.replace(/^0+(?=\d)/, "") || "0";
+}
+function decimalKey(intDigits, frac) {
+  return { key: `${stripInt(intDigits)}.${frac}`.replace(/\.0+$/, ""), twoPlace: frac.length === 2 };
+}
+function canonicalBody(body) {
+  const noSpace = [...body].map((ch) => isGroupSpace(ch) ? " " : ch).join("").replace(/[’']/g, "").replace(/ /g, "");
+  if (!/^\d+(?:[.,]\d+)*$/.test(noSpace)) return null;
+  if (!noSpace.includes(".") && !noSpace.includes(",")) return { key: stripInt(noSpace), twoPlace: false };
+  if (noSpace.includes(".") && noSpace.includes(",")) {
+    const dec = noSpace.lastIndexOf(".") > noSpace.lastIndexOf(",") ? "." : ",";
+    const at = noSpace.lastIndexOf(dec);
+    const intDigits2 = noSpace.slice(0, at).replace(/[.,]/g, "");
+    const frac2 = noSpace.slice(at + 1);
+    if (!/^\d+$/.test(intDigits2) || !/^\d+$/.test(frac2)) return null;
+    return decimalKey(intDigits2, frac2);
+  }
+  const parts = noSpace.split(noSpace.includes(".") ? "." : ",");
+  if (parts.every((part, index) => /^\d+$/.test(part) && (index === 0 || part.length === 3))) {
+    return { key: stripInt(parts.join("")), twoPlace: false };
+  }
+  const frac = parts[parts.length - 1] ?? "";
+  const intDigits = parts.slice(0, -1).join("");
+  if (!/^\d+$/.test(intDigits) || !/^\d+$/.test(frac)) return null;
+  return decimalKey(intDigits, frac);
+}
+function parseFigure(raw) {
+  let body = raw.trim();
+  let kind = "other";
+  if (/^[$€£]/.test(body)) {
+    kind = "money";
+    body = body.slice(1).trimStart();
+  }
+  if (body.endsWith("%")) {
+    kind = "percent";
+    body = body.slice(0, -1).trimEnd();
+  }
+  const parsed = canonicalBody(body);
+  if (!parsed) return null;
+  if (kind === "other" && parsed.twoPlace) kind = "money";
+  if (kind === "other" && /^\d$/.test(parsed.key)) return null;
+  return { key: parsed.key, kind };
+}
+function scanFigures(text) {
+  const out = [];
+  const groupSep = /[ \u00A0\u202F\u2007\u2009'’.,]/;
+  for (let i = 0; i < text.length; i += 1) {
+    const ch = text[i] ?? "";
+    const prev = text[i - 1] ?? "";
+    if (/\d/.test(ch) && /[\dA-Za-z]/.test(prev)) continue;
+    if (!/\d/.test(ch) && !/[$€£]/.test(ch)) continue;
+    let cursor = i;
+    if (/[$€£]/.test(text[cursor] ?? "")) {
+      cursor += 1;
+      if (text[cursor] === " ") cursor += 1;
+    }
+    if (!/\d/.test(text[cursor] ?? "")) continue;
+    while (cursor < text.length && /\d/.test(text[cursor] ?? "")) cursor += 1;
+    while (cursor < text.length && groupSep.test(text[cursor] ?? "")) {
+      const sep = text[cursor] ?? "";
+      let next = cursor + 1;
+      let digits = 0;
+      while (next < text.length && /\d/.test(text[next] ?? "")) {
+        digits += 1;
+        next += 1;
+      }
+      if (!digits) break;
+      const spaceLike = sep === " " || isGroupSpace(sep) || sep === "'" || sep === "\u2019";
+      if (spaceLike && digits !== 3) break;
+      cursor = next;
+    }
+    let end = cursor;
+    if (text[end] === "%") end += 1;
+    else if (text[end] === " " && text[end + 1] === "%") end += 2;
+    const parsed = parseFigure(text.slice(i, end));
+    if (parsed) out.push(parsed);
+    if (end > i) i = end - 1;
+  }
+  return out;
+}
 function figuresIn(text) {
-  return [...new Set((text.match(FIGURE) ?? []).map(normalizeFigure))].filter(Boolean);
+  const out = [];
+  for (const figure of scanFigures(text)) {
+    if (!out.includes(figure.key)) out.push(figure.key);
+  }
+  return out;
 }
 function figureSupported(figure, haystack) {
   const figures = new Set(figuresIn(haystack));
@@ -930,19 +1077,15 @@ function roundCents(n) {
   return Math.round(n * 100) / 100;
 }
 function figureKey(n) {
-  return normalizeFigure(String(roundCents(n)));
+  return String(roundCents(n)).replace(/\.0+$/, "");
 }
 function sourceFigures(text) {
   const out = [];
-  const re = /(?:[$€£]\s?(\d[\d,]*(?:\.\d+)?))|(?:\b(\d[\d,]*(?:\.\d+)?)\s?%)|(?:\b(\d+\.\d+)\b)|(?:\b(\d{2,}[\d,]*)\b)/g;
-  let m;
-  while (m = re.exec(text)) {
-    if (m[1]) out.push({ num: Number(m[1].replace(/,/g, "")), kind: "money" });
-    else if (m[2]) out.push({ num: Number(m[2].replace(/,/g, "")), kind: "percent" });
-    else if (m[3]) out.push({ num: Number(m[3]), kind: /\.\d{2}$/.test(m[3]) ? "money" : "other" });
-    else if (m[4]) out.push({ num: Number(m[4].replace(/,/g, "")), kind: "other" });
+  for (const figure of scanFigures(text)) {
+    const num = Number(figure.key);
+    if (Number.isFinite(num)) out.push({ num, kind: figure.kind });
   }
-  return out.filter((f) => Number.isFinite(f.num));
+  return out;
 }
 function uniqueNums(values) {
   const seen = /* @__PURE__ */ new Set();
@@ -1224,8 +1367,9 @@ function citationsFor(cited, passages, answer, max) {
   }
   return out;
 }
+var CITATION_MARK = /\s*\[[\d\u0660-\u0669\u06F0-\u06F9]+(?:\s*[,،、，]\s*[\d\u0660-\u0669\u06F0-\u06F9]+)*\]/g;
 function tidy(answer) {
-  return answer.replace(/\s*\[\d+(?:,\s*\d+)*\]/g, "").replace(/^#+\s+/gm, "").replace(/[ \t]+\n/g, "\n").replace(/\n{3,}/g, "\n\n").trim();
+  return answer.replace(CITATION_MARK, "").replace(/^#+\s+/gm, "").replace(/[ \t]+\n/g, "\n").replace(/\n{3,}/g, "\n\n").trim();
 }
 function authFailure(error) {
   return error instanceof LlmError && (error.status === 401 || error.status === 403);
@@ -1344,6 +1488,18 @@ function samePage(left, right) {
 }
 var HISTORY_TURNS = 3;
 var SESSIONS_MAX = 5e3;
+var AGENT_INDEX_PATHS = ["/llms.txt", "/llms-full.txt"];
+function agentIndexUrls(origin) {
+  const base = origin.endsWith("/") ? origin : `${origin}/`;
+  return AGENT_INDEX_PATHS.map((path) => new URL(path, base).href);
+}
+function isAgentIndexUrl(raw) {
+  try {
+    return /^\/llms(?:-full)?\.txt$/i.test(new URL(raw).pathname);
+  } catch {
+    return false;
+  }
+}
 function randomHex(bytes) {
   const values = new Uint8Array(bytes);
   globalThis.crypto.getRandomValues(values);
@@ -1395,20 +1551,32 @@ function createAnswersBackend(options) {
   async function refresh() {
     const urls = [...options.urls ?? []];
     if (options.sitemap) urls.push(...await sitemapUrls(fetchImpl, options.sitemap, options.maxPages ?? 500));
+    const listed = [...new Set(urls)].slice(0, options.maxPages ?? 500);
+    const listedSet = new Set(listed);
+    const optional = /* @__PURE__ */ new Set();
+    for (const url of listed) {
+      try {
+        for (const extra of agentIndexUrls(new URL(url).origin)) {
+          if (!listedSet.has(extra)) optional.add(extra);
+        }
+      } catch {
+      }
+    }
     const docs = [...options.documents ?? []];
     const skipped = [];
-    for (const url of [...new Set(urls)].slice(0, options.maxPages ?? 500)) {
-      const response = await fetchImpl(url, { headers: { accept: "text/html,text/plain" } }).catch(() => null);
+    for (const url of [...listed, ...optional]) {
+      const quiet = optional.has(url);
+      const response = await fetchImpl(url, { headers: { accept: "text/html, text/markdown, text/plain" } }).catch(() => null);
       if (!response) {
-        skipped.push({ url, reason: "The page could not be fetched." });
+        if (!quiet) skipped.push({ url, reason: "The page could not be fetched." });
         continue;
       }
       if (!response.ok) {
-        skipped.push({ url, reason: `HTTP ${response.status}.` });
+        if (!quiet) skipped.push({ url, reason: `HTTP ${response.status}.` });
         continue;
       }
       if (response.redirected && response.url && !samePage(response.url, url)) {
-        skipped.push({ url, reason: `Redirects to ${response.url}. List that URL instead if it should be indexed.` });
+        if (!quiet) skipped.push({ url, reason: `Redirects to ${response.url}. List that URL instead if it should be indexed.` });
         continue;
       }
       const type = response.headers.get("content-type") ?? "";
@@ -1499,27 +1667,33 @@ export {
   DEFAULT_ANSWER_MODEL,
   DEFAULT_EMBEDDING_MODEL,
   EMBEDDING_DIMENSIONS,
+  EVIDENCE_LANE,
   FALLBACK_ANSWER_MODEL,
   LlmError,
   MemoryIndex,
   QA_LANE,
   QA_LANE_MIN_SIMILARITY,
+  agentIndexUrls,
   answerMessages,
   answerQuestion,
   authFailure,
+  capabilityVerbs,
   chunkPage,
   citationsFor,
+  contentTokenOverlap,
   cosine,
   createAnswersBackend,
   createOpenAiClient,
   decodeCfEmail,
   decodeEntities,
   decodeProtectedEmails,
+  distinctiveTokens,
   extractPage,
   extractText,
   figuresIn,
   firstPathSegment,
   inferSources,
+  isAgentIndexUrl,
   isReasoningModel,
   matchingQa,
   namedPathBucket,
@@ -1527,9 +1701,11 @@ export {
   prepareMessages,
   reciprocalRankFusion,
   samePage,
+  selectNamedEvidence,
   selectQaLane,
   sourceBlock,
   stem,
+  subjectTokens,
   tokenize,
   verifyAnswer,
   withQaLane