@crestapps/ai-chat-ui 2.0.0-preview.190 → 2.0.0-preview.193

package/dist/chat-markers.js

@@ -0,0 +1,439 @@
+/*
+** NOTE: This file is generated by Gulp and should not be edited directly!
+** Any changes made directly to this file will be overwritten next time its asset group is processed by Gulp.
+*/
+
+function _typeof(o) { "@babel/helpers - typeof"; return _typeof = "function" == typeof Symbol && "symbol" == typeof Symbol.iterator ? function (o) { return typeof o; } : function (o) { return o && "function" == typeof Symbol && o.constructor === Symbol && o !== Symbol.prototype ? "symbol" : typeof o; }, _typeof(o); }
+/*
+ * Markers the host expands on the model's behalf.
+ *
+ * Retrieval hands the model a short label such as [fig:1] rather than a figure's address, because a model
+ * asked to reproduce a long opaque identifier reproduces its shape and varies the digits instead --
+ * producing addresses that look right and resolve to nothing. The model therefore never types an address,
+ * and this turns the label it does type back into the picture.
+ *
+ * The chart tool works the same way from the other end: it hands back a [chart:{...}] marker and asks for it
+ * verbatim, and the host turns that marker into a canvas. Reading a marker is a statement about a string, so
+ * it lives here; drawing one is not, so each surface still does its own drawing.
+ *
+ * Loaded on its own so every chat surface shares one implementation: the two shared chat scripts and the
+ * MVC chat-interaction view, which renders its markdown with its own inline marked setup.
+ */
+window.CoreAIChatMarkers = window.CoreAIChatMarkers || function () {
+  'use strict';
+
+  // Used when a figure has no caption, so the image is still announced as something.
+  var defaultImageAltText = 'Figure';
+
+  /*
+   * Makes a caption safe to sit inside the alt text of ![alt](link).
+   *
+   * Captions really do contain brackets, and an unescaped one closes the alt text early: a caption of
+   * "Figure 2 (revised) [draft]" would end the image after "Figure 2 (revised) " and leave the rest of the
+   * syntax on the page as markup. Parentheses are harmless there, only the brackets are. A caption that spans
+   * lines would end the paragraph the image lives in, so its whitespace is collapsed to single spaces.
+   */
+  function escapeImageAltText(title) {
+    var text = typeof title === 'string' ? title.replace(/\s+/g, ' ').trim() : '';
+    if (!text) {
+      return defaultImageAltText;
+    }
+
+    // The backslash is escaped as well, or a caption ending in one would escape the bracket that closes
+    // the alt text and the image would swallow the rest of the line.
+    return text.replace(/[\\\[\]]/g, '\\$&');
+  }
+
+  /*
+   * Makes a link safe to sit inside the destination of ![alt](link): the few characters that would end the
+   * destination early are percent-encoded, which leaves the URL addressing the same resource. The link is
+   * still run through sanitizeUrl by the image renderer, so this is about parsing, not about safety.
+   */
+  function encodeImageLink(link) {
+    return link.replace(/[\s()<>"\\]/g, function (character) {
+      return '%' + character.charCodeAt(0).toString(16).toUpperCase().padStart(2, '0');
+    });
+  }
+
+  /*
+   * Returns the content with every marker the reference map describes as a servable image replaced by
+   * markdown image syntax, so the ordinary image renderer draws it with its thumbnail, its download button
+   * and its size cap. Pure: nothing here touches the DOM or the message.
+   *
+   * A marker is replaced only when its reference is an image and carries a link. Anything else is left
+   * exactly as written -- a marker the model invented for a figure that was never in the results, or one
+   * whose picture the host cannot serve, reaches the reader as the few characters the model typed rather
+   * than as a broken image.
+   */
+  function expandImageMarkers(content, references) {
+    if (typeof content !== 'string' || !content) {
+      return typeof content === 'string' ? content : '';
+    }
+    if (!references || _typeof(references) !== 'object') {
+      return content;
+    }
+    var expanded = content;
+    var _loop = function _loop() {
+        var _reference$isImage, _reference$link, _reference$title;
+        var marker = _Object$keys[_i];
+        var reference = references[marker];
+        if (!marker || !reference || _typeof(reference) !== 'object') {
+          return 0; // continue
+        }
+        if (((_reference$isImage = reference.isImage) !== null && _reference$isImage !== void 0 ? _reference$isImage : reference.IsImage) !== true) {
+          return 0; // continue
+        }
+        var rawLink = (_reference$link = reference.link) !== null && _reference$link !== void 0 ? _reference$link : reference.Link;
+        var link = typeof rawLink === 'string' ? rawLink.trim() : '';
+        if (!link) {
+          return 0; // continue
+        }
+        var image = "![".concat(escapeImageAltText((_reference$title = reference.title) !== null && _reference$title !== void 0 ? _reference$title : reference.Title), "](").concat(encodeImageLink(link), ")");
+
+        // Every occurrence, and through a replacer function so a '$' in a caption or a link is not read as
+        // a replacement pattern.
+        expanded = expanded.replaceAll(marker, function () {
+          return image;
+        });
+      },
+      _ret;
+    for (var _i = 0, _Object$keys = Object.keys(references); _i < _Object$keys.length; _i++) {
+      _ret = _loop();
+      if (_ret === 0) continue;
+    }
+    return expanded;
+  }
+
+  // The marker the chart tool emits and asks the model to repeat verbatim, braces and all.
+  var chartMarkerPrefix = '[chart:';
+
+  // What a citation the model typed opens with. The reference map is keyed by the whole marker, so this is
+  // also how a rewritten one is put back together.
+  var citationMarkerPrefix = '[doc:';
+
+  /*
+   * How much text may sit between "[doc:" and its "]" before it stops being a citation.
+   *
+   * A list of reference numbers is short: even ten of them, written the long way, is under a hundred
+   * characters. The limit is what keeps the work per marker fixed, and so the whole pass linear -- without
+   * it, an opening bracket whose "]" is thousands of characters away is examined in full, and a paragraph
+   * full of such openings is examined once per opening.
+   */
+  var maxCitationMarkerLength = 512;
+  function isMarkerWhitespace(character) {
+    return character === ' ' || character === '\n' || character === '\r' || character === '\t';
+  }
+
+  /*
+   * Reads the marker that begins at startIndex, or returns null when what begins there is not one.
+   *
+   * The configuration is JSON and JSON nests, so the end of the object is found by counting braces rather
+   * than by a regular expression: an expression that stopped at the first '}' would truncate every chart
+   * whose options carry an object of their own, which is all of them. The count has to know when it is
+   * inside a string as well -- an axis label really does read "Q1 { Q2", and a counter blind to strings
+   * would end the object in the middle of the label and hand the renderer a fragment. Escapes are tracked
+   * for the same reason, so a \" inside a label does not look like the end of it.
+   */
+  function readChartMarkerAt(text, startIndex) {
+    var index = startIndex + chartMarkerPrefix.length;
+    while (index < text.length && isMarkerWhitespace(text[index])) {
+      index++;
+    }
+    if (index >= text.length || text[index] !== '{') {
+      return null;
+    }
+    var jsonStart = index;
+    var depth = 0;
+    var inString = false;
+    var escaped = false;
+    for (; index < text.length; index++) {
+      var character = text[index];
+      if (inString) {
+        if (escaped) {
+          escaped = false;
+        } else if (character === '\\') {
+          escaped = true;
+        } else if (character === '"') {
+          inString = false;
+        }
+        continue;
+      }
+      if (character === '"') {
+        inString = true;
+        continue;
+      }
+      if (character === '{') {
+        depth++;
+        continue;
+      }
+      if (character !== '}') {
+        continue;
+      }
+      depth--;
+      if (depth > 0) {
+        continue;
+      }
+      var jsonEnd = index + 1;
+      var closeIndex = jsonEnd;
+      while (closeIndex < text.length && isMarkerWhitespace(text[closeIndex])) {
+        closeIndex++;
+      }
+
+      // Only whitespace may sit between the object and the bracket that closes the marker. Hunting
+      // further ahead for a ']' would let a marker whose bracket the model dropped reach the one
+      // belonging to something else, and everything in between -- a paragraph of the answer, the next
+      // marker -- would be swallowed into the span the caller removes from the text.
+      if (closeIndex >= text.length || text[closeIndex] !== ']') {
+        return null;
+      }
+      return {
+        startIndex: startIndex,
+        endIndex: closeIndex + 1,
+        json: text.substring(jsonStart, jsonEnd)
+      };
+    }
+    return null;
+  }
+
+  /*
+   * Finds the first [chart:{...}] marker in the content and returns where it starts, where it ends and the
+   * JSON between, or null when there is no marker. Pure: it knows nothing of canvases, element ids or
+   * Chart.js, because the three chat surfaces draw a chart differently and agree only on what a marker is.
+   *
+   * Text that merely looks like a marker is not one, and none of these return a span: "[chart:" with no
+   * object after it, an object whose braces never balance, an object the model never closed with ']'. Each
+   * reaches the reader as the text the model typed, which is what the caller does with a null.
+   */
+  function findChartMarker(text) {
+    if (typeof text !== 'string' || !text) {
+      return null;
+    }
+    var searchFrom = 0;
+    for (;;) {
+      var start = text.indexOf(chartMarkerPrefix, searchFrom);
+      if (start < 0) {
+        return null;
+      }
+      var marker = readChartMarkerAt(text, start);
+      if (marker) {
+        return marker;
+      }
+
+      // Something that opened like a marker and was not one must not hide a real one further on, so the
+      // search carries on past it rather than giving up at the first disappointment.
+      searchFrom = start + 1;
+    }
+  }
+
+  /*
+   * Splits a citation marker the model wrote as one bracket into the separate markers it means.
+   *
+   * A reference is keyed by the literal string the model is asked to type -- "[doc:1]" -- and the renderer
+   * replaces exactly that. Models routinely gather several into one bracket instead: "[doc:1, doc:2]". No
+   * key matches that, so nothing is replaced and the reader is shown the raw marker in the middle of a
+   * sentence, while the citations the model happened to write singly render correctly beside it.
+   *
+   * Rewriting the combined form into the separate markers is what makes the rest of the pipeline see them.
+   * Both spellings are accepted, since a model that writes "[doc:1, doc:2]" will also write "[doc:1, 2]".
+   * A lone "[doc:1]" is left exactly as it is.
+   */
+  function splitCombinedCitations(content) {
+    if (typeof content !== 'string' || !content) {
+      return '';
+    }
+    var result = '';
+    var from = 0;
+
+    // The closing bracket last searched for. A ']' found from an earlier position is still the next one
+    // from here, because there was none in between -- so text like "[doc:x[doc:x[doc:x...]" is walked once
+    // rather than rescanned from every false start, which would make the whole pass quadratic.
+    var knownClose = -1;
+    for (;;) {
+      var start = content.indexOf(citationMarkerPrefix, from);
+      if (start < 0) {
+        return result + content.slice(from);
+      }
+      var end = knownClose >= start ? knownClose : content.indexOf(']', start);
+      knownClose = end;
+      if (end < 0) {
+        // Nothing closes it, so nothing here is a marker. The rest is returned as it was written.
+        return result + content.slice(from);
+      }
+      var numbers = end - start <= maxCitationMarkerLength ? readCitationNumbers(content.slice(start + citationMarkerPrefix.length, end)) : null;
+      if (!numbers) {
+        // Whatever this opened, it was not a list of references. The search carries on from just
+        // after the prefix rather than past the bracket, because the ']' that was found may belong to
+        // a real marker further along -- "[doc:1, and later [doc:3, doc:4]" closes the second one, and
+        // skipping to it would hide it.
+        result += content.slice(from, start + citationMarkerPrefix.length);
+        from = start + citationMarkerPrefix.length;
+        continue;
+      }
+
+      // Rewritten whenever the bracket does not already read as the canonical marker: several numbers
+      // in one bracket, or a single number carrying a locator the model appended. A bare "[doc:1]"
+      // rewrites to itself, so it is passed through untouched rather than rebuilt.
+      var canonical = numbers.map(function (number) {
+        return citationMarkerPrefix + number + ']';
+      }).join('');
+      var original = content.slice(start, end + 1);
+      result += content.slice(from, start);
+      result += canonical === original ? original : canonical;
+      from = end + 1;
+    }
+  }
+
+  /*
+   * Reads the numbers out of the inside of a citation bracket, or returns null when it does not hold a list
+   * of them.
+   *
+   * Deliberately not a regular expression over the whole bracket. The pattern that reads naturally --
+   * optional spaces on both sides of each comma, repeated -- gives the engine several ways to match the same
+   * text, and on a bracket that never closes it explores all of them: CodeQL flagged it as exponential, and
+   * it was. A model writing "[doc:9, 9, 9, 9..." and forgetting the bracket is not a hostile act, and it is
+   * a reader's own tab that stops responding.
+   *
+   * Splitting on the comma and checking each part on its own is linear, and each pattern below is anchored
+   * over a single character class, so neither can backtrack at all.
+   */
+  /*
+   * A locator a model appends to a citation instead of writing the bare marker it was asked for: "pages
+   * 13-15", "page 4", "pp. 3-5", "pg 7". It is not a citation number and the reference map holds nothing
+   * under it, so before this it made the whole marker unreadable and "[doc:1, pages 13-15]" reached the
+   * reader as those characters.
+   *
+   * Dropped rather than rendered. The page a citation points at is already carried by the reference the
+   * number resolves to, which is the one the host looked up rather than the one the model recalled.
+   */
+  var citationLocatorPattern = /^(?:pages?|pp?g?|pp)\.?\s*[\d\s‐-―,&+-]*$/i;
+  function readCitationNumbers(interior) {
+    var parts = interior.split(',');
+    var numbers = [];
+    for (var index = 0; index < parts.length; index++) {
+      // Only the parts after the first may repeat the prefix: "[doc:1, doc:2]" and "[doc:1, 2]" are both
+      // written by models, "[doc:doc:1]" is not.
+      var part = index === 0 ? parts[index].trim() : parts[index].trim().replace(/^doc:\s*/, '');
+      if (/^\d+$/.test(part)) {
+        numbers.push(part);
+        continue;
+      }
+
+      // A locator is allowed only after a number has been read, so "[doc:pages 3]" is still not a
+      // citation -- something that names no document is not one whatever follows it.
+      if (numbers.length > 0 && citationLocatorPattern.test(part)) {
+        continue;
+      }
+      return null;
+    }
+    return numbers.length > 0 ? numbers : null;
+  }
+
+  /*
+   * The identity of a citation as a reader meets it: the line it prints, and where it points.
+   *
+   * Retrieval returns one reference per chunk, so a single article that answered a question through three of
+   * its chunks arrives as three references. Numbering those separately prints "1,2,3" over the sentence and
+   * then lists the same title three times, which tells the reader there are three sources corroborating the
+   * claim when there is one. Two references that would print the same line and lead to the same place are
+   * one citation, and are numbered once.
+   *
+   * Case and surrounding whitespace are not part of the identity, because they are not part of what the
+   * reader sees as different. Anything else -- a different title, a different page, a link to one figure
+   * rather than another -- keeps the citations apart, since that is a distinction the reader can act on.
+   */
+  function citationIdentity(label, link) {
+    var text = typeof label === 'string' ? label.replace(/\s+/g, ' ').trim().toLowerCase() : '';
+    var target = typeof link === 'string' ? link.trim().toLowerCase() : '';
+
+    // A null byte cannot occur in either part, so no pair of values can collide across the boundary:
+    // without it, label "a" + link "bc" and label "ab" + link "c" would be the same identity.
+    return text + '' + target;
+  }
+
+  /*
+   * Collapses a run of identical citation markers into one.
+   *
+   * Once two references share a number, a sentence that cited both ends with that number twice over --
+   * "1,1" -- which reads as a typo. Only an immediately repeated marker is removed: the same source cited
+   * again later in the paragraph is a separate citation of it and keeps its marker.
+   *
+   * The marker is matched with whatever attributes it carries, because the number is what makes two of them
+   * the same citation; a pattern written for a bare <sup> would silently stop collapsing the moment a
+   * marker gained a tooltip.
+   *
+   * What separates the two markers is swallowed with them, because the model writes its own punctuation
+   * between references -- "[ref1], [ref2]" -- and once both carry the same number that comma is left
+   * between a number and itself, printing "1,1". Only a separator that was joining the two citations is
+   * eaten: whitespace, a comma, or the marker the comma rule inserts. Anything else means the second
+   * marker is a fresh citation later in the sentence, and it is left alone.
+   */
+  var citationSeparator = '(?:\\s*(?:,|<sup(?:\\s[^>]*)?>,<\\/sup>)?\\s*)';
+  var repeatedCitationPattern = new RegExp('(<sup(?:\\s[^>]*)?>(\\d+)<\\/sup>)(?:' + citationSeparator + '<sup(?:\\s[^>]*)?>\\2<\\/sup>)+', 'g');
+  function collapseRepeatedCitations(html) {
+    if (typeof html !== 'string' || !html) {
+      return '';
+    }
+    return html.replace(repeatedCitationPattern, '$1');
+  }
+
+  /*
+   * Puts a comma between two citation markers that ended up side by side, so "12" reads as two sources
+   * rather than as the twelfth.
+   *
+   * Only the boundary between the markers is matched, so whatever attributes either one carries are left
+   * untouched. The inserted marker is not itself rescanned, since a replace walks the string it was given.
+   */
+  function separateAdjacentCitations(html) {
+    if (typeof html !== 'string' || !html) {
+      return '';
+    }
+    return html.replace(/<\/sup><sup/g, '</sup><sup>,</sup><sup');
+  }
+
+  // Escapes the few characters that would end an HTML attribute early, so a caption containing a quote
+  // cannot introduce markup of its own.
+  function escapeAttributeValue(text) {
+    return text.replace(/[&<>"']/g, function (character) {
+      switch (character) {
+        case '&':
+          return '&amp;';
+        case '<':
+          return '&lt;';
+        case '>':
+          return '&gt;';
+        case '"':
+          return '&quot;';
+        default:
+          return '&#39;';
+      }
+    });
+  }
+
+  /*
+   * The inline marker for a citation: its number, and the source it stands for as a tooltip.
+   *
+   * A bare number tells the reader only that something was cited, and answers which source it was by
+   * making them look away from the sentence and match a number against a list underneath. Carrying the
+   * label means the answer arrives where the question is asked. The list stays as it is, since a tooltip
+   * is not reachable by keyboard and cannot be the only place the source appears.
+   */
+  function citationMarkerHtml(displayIndex, label) {
+    var text = typeof label === 'string' ? label.replace(/\s+/g, ' ').trim() : '';
+    if (!text) {
+      return '<sup>' + displayIndex + '</sup>';
+    }
+    return '<sup title="' + escapeAttributeValue(text) + '">' + displayIndex + '</sup>';
+  }
+
+  // Deliberately small: pure functions over strings, for the hosts below and for the tests.
+  return {
+    expandImageMarkers: expandImageMarkers,
+    findChartMarker: findChartMarker,
+    citationIdentity: citationIdentity,
+    splitCombinedCitations: splitCombinedCitations,
+    collapseRepeatedCitations: collapseRepeatedCitations,
+    separateAdjacentCitations: separateAdjacentCitations,
+    citationMarkerHtml: citationMarkerHtml
+  };
+}();
+//# sourceMappingURL=chat-markers.js.map

package/dist/chat-markers.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["chat-markers.js"],"names":["window","CoreAIChatMarkers","defaultImageAltText","escapeImageAltText","title","text","replace","trim","encodeImageLink","link","character","charCodeAt","toString","toUpperCase","padStart","expandImageMarkers","content","references","_typeof","expanded","_loop","_reference$isImage","_reference$link","_reference$title","marker","_Object$keys","_i","reference","isImage","IsImage","rawLink","Link","image","concat","Title","replaceAll","_ret","Object","keys","length","chartMarkerPrefix","citationMarkerPrefix","maxCitationMarkerLength","isMarkerWhitespace","readChartMarkerAt","startIndex","index","jsonStart","depth","inString","escaped","jsonEnd","closeIndex","endIndex","json","substring","findChartMarker","searchFrom","start","indexOf","splitCombinedCitations","result","from","knownClose","slice","end","numbers","readCitationNumbers","canonical","map","number","join","original","citationLocatorPattern","interior","parts","split","part","test","push","citationIdentity","label","toLowerCase","target","citationSeparator","repeatedCitationPattern","RegExp","collapseRepeatedCitations","html","separateAdjacentCitations","escapeAttributeValue","citationMarkerHtml","displayIndex"],"mappings":";;;;;;AAAA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACAA,MAAM,CAACC,iBAAiB,GAAGD,MAAM,CAACC,iBAAiB,IAAK,YAAY;EAChE,YAAY;;EAEZ;EACA,IAAMC,mBAAmB,GAAG,QAAQ;;EAEpC;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;EACI,SAASC,kBAAkBA,CAACC,KAAK,EAAE;IAC/B,IAAMC,IAAI,GAAG,OAAOD,KAAK,KAAK,QAAQ,GAAGA,KAAK,CAACE,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAACC,IAAI,CAAC,CAAC,GAAG,EAAE;IAE/E,IAAI,CAACF,IAAI,EAAE;MACP,OAAOH,mBAAmB;IAC9B;;IAEA;IACA;IACA,OAAOG,IAAI,CAACC,OAAO,CAAC,WAAW,EAAE,MAAM,CAAC;EAC5C;;EAEA;AACJ;AACA;AACA;AACA;EACI,SAASE,eAAeA,CAACC,IAAI,EAAE;IAC3B,OAAOA,IAAI,CAACH,OAAO,CAAC,cAAc,EAAE,UAAUI,SAAS,EAAE;MACrD,OAAO,GAAG,GAAGA,SAAS,CAACC,UAAU,CAAC,CAAC,CAAC,CAACC,QAAQ,CAAC,EAAE,CAAC,CAACC,WAAW,CAAC,CAAC,CAACC,QAAQ,CAAC,CAAC,EAAE,GAAG,CAAC;IACpF,CAAC,CAAC;EACN;;EAEA;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACI,SAASC,kBAAkBA,CAACC,OAAO,EAAEC,UAAU,EAAE;IAC7C,IAAI,OAAOD,OAAO,KAAK,QAAQ,IAAI,CAACA,OAAO,EAAE;MACzC,OAAO,OAAOA,OAAO,KAAK,QAAQ,GAAGA,OAAO,GAAG,EAAE;IACrD;IAEA,IAAI,CAACC,UAAU,IAAIC,OAAA,CAAOD,UAAU,MAAK,QAAQ,EAAE;MAC/C,OAAOD,OAAO;IAClB;IAEA,IAAIG,QAAQ,GAAGH,OAAO;IAAC,IAAAI,KAAA,YAAAA,MAAA,EAEuB;QAAA,IAAAC,kBAAA,EAAAC,eAAA,EAAAC,gBAAA;QAAzC,IAAMC,MAAM,GAAAC,YAAA,CAAAC,EAAA;QACb,IAAMC,SAAS,GAAGV,UAAU,CAACO,MAAM,CAAC;QAEpC,IAAI,CAACA,MAAM,IAAI,CAACG,SAAS,IAAIT,OAAA,CAAOS,SAAS,MAAK,QAAQ,EAAE;UAAA;QAE5D;QAEA,IAAI,EAAAN,kBAAA,GAACM,SAAS,CAACC,OAAO,cAAAP,kBAAA,cAAAA,kBAAA,GAAIM,SAAS,CAACE,OAAO,MAAM,IAAI,EAAE;UAAA;QAEvD;QAEA,IAAMC,OAAO,IAAAR,eAAA,GAAGK,SAAS,CAAClB,IAAI,cAAAa,eAAA,cAAAA,eAAA,GAAIK,SAAS,CAACI,IAAI;QAChD,IAAMtB,IAAI,GAAG,OAAOqB,OAAO,KAAK,QAAQ,GAAGA,OAAO,CAACvB,IAAI,CAAC,CAAC,GAAG,EAAE;QAE9D,IAAI,CAACE,IAAI,EAAE;UAAA;QAEX;QAEA,IAAMuB,KAAK,QAAAC,MAAA,CAAQ9B,kBAAkB,EAAAoB,gBAAA,GAACI,SAAS,CAACvB,KAAK,cAAAmB,gBAAA,cAAAA,gBAAA,GAAII,SAAS,CAACO,KAAK,CAAC,QAAAD,MAAA,CAAKzB,eAAe,CAACC,IAAI,CAAC,MAAG;;QAEtG;QACA;QACAU,QAAQ,GAAGA,QAAQ,CAACgB,UAAU,CAACX,MAAM,EAAE,YAAY;UAAE,OAAOQ,KAAK;QAAE,CAAC,CAAC;MACzE,CAAC;MAAAI,IAAA;IAvBD,SAAAV,EAAA,MAAAD,YAAA,GAAqBY,MAAM,CAACC,IAAI,CAACrB,UAAU,CAAC,EAAAS,EAAA,GAAAD,YAAA,CAAAc,MAAA,EAAAb,EAAA;MAAAU,IAAA,GAAAhB,KAAA;MAAA,IAAAgB,IAAA,QAIpC;IAAS;IAqBjB,OAAOjB,QAAQ;EACnB;;EAEA;EACA,IAAMqB,iBAAiB,GAAG,SAAS;;EAEnC;EACA;EACA,IAAMC,oBAAoB,GAAG,OAAO;;EAEpC;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;EACI,IAAMC,uBAAuB,GAAG,GAAG;EAEnC,SAASC,kBAAkBA,CAACjC,SAAS,EAAE;IACnC,OAAOA,SAAS,KAAK,GAAG,IAAIA,SAAS,KAAK,IAAI,IAAIA,SAAS,KAAK,IAAI,IAAIA,SAAS,KAAK,IAAI;EAC9F;;EAEA;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACI,SAASkC,iBAAiBA,CAACvC,IAAI,EAAEwC,UAAU,EAAE;IACzC,IAAIC,KAAK,GAAGD,UAAU,GAAGL,iBAAiB,CAACD,MAAM;IAEjD,OAAOO,KAAK,GAAGzC,IAAI,CAACkC,MAAM,IAAII,kBAAkB,CAACtC,IAAI,CAACyC,KAAK,CAAC,CAAC,EAAE;MAC3DA,KAAK,EAAE;IACX;IAEA,IAAIA,KAAK,IAAIzC,IAAI,CAACkC,MAAM,IAAIlC,IAAI,CAACyC,KAAK,CAAC,KAAK,GAAG,EAAE;MAC7C,OAAO,IAAI;IACf;IAEA,IAAMC,SAAS,GAAGD,KAAK;IACvB,IAAIE,KAAK,GAAG,CAAC;IACb,IAAIC,QAAQ,GAAG,KAAK;IACpB,IAAIC,OAAO,GAAG,KAAK;IAEnB,OAAOJ,KAAK,GAAGzC,IAAI,CAACkC,MAAM,EAAEO,KAAK,EAAE,EAAE;MACjC,IAAMpC,SAAS,GAAGL,IAAI,CAACyC,KAAK,CAAC;MAE7B,IAAIG,QAAQ,EAAE;QACV,IAAIC,OAAO,EAAE;UACTA,OAAO,GAAG,KAAK;QACnB,CAAC,MAAM,IAAIxC,SAAS,KAAK,IAAI,EAAE;UAC3BwC,OAAO,GAAG,IAAI;QAClB,CAAC,MAAM,IAAIxC,SAAS,KAAK,GAAG,EAAE;UAC1BuC,QAAQ,GAAG,KAAK;QACpB;QAEA;MACJ;MAEA,IAAIvC,SAAS,KAAK,GAAG,EAAE;QACnBuC,QAAQ,GAAG,IAAI;QACf;MACJ;MAEA,IAAIvC,SAAS,KAAK,GAAG,EAAE;QACnBsC,KAAK,EAAE;QACP;MACJ;MAEA,IAAItC,SAAS,KAAK,GAAG,EAAE;QACnB;MACJ;MAEAsC,KAAK,EAAE;MAEP,IAAIA,KAAK,GAAG,CAAC,EAAE;QACX;MACJ;MAEA,IAAMG,OAAO,GAAGL,KAAK,GAAG,CAAC;MACzB,IAAIM,UAAU,GAAGD,OAAO;MAExB,OAAOC,UAAU,GAAG/C,IAAI,CAACkC,MAAM,IAAII,kBAAkB,CAACtC,IAAI,CAAC+C,UAAU,CAAC,CAAC,EAAE;QACrEA,UAAU,EAAE;MAChB;;MAEA;MACA;MACA;MACA;MACA,IAAIA,UAAU,IAAI/C,IAAI,CAACkC,MAAM,IAAIlC,IAAI,CAAC+C,UAAU,CAAC,KAAK,GAAG,EAAE;QACvD,OAAO,IAAI;MACf;MAEA,OAAO;QACHP,UAAU,EAAEA,UAAU;QACtBQ,QAAQ,EAAED,UAAU,GAAG,CAAC;QACxBE,IAAI,EAAEjD,IAAI,CAACkD,SAAS,CAACR,SAAS,EAAEI,OAAO;MAC3C,CAAC;IACL;IAEA,OAAO,IAAI;EACf;;EAEA;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACI,SAASK,eAAeA,CAACnD,IAAI,EAAE;IAC3B,IAAI,OAAOA,IAAI,KAAK,QAAQ,IAAI,CAACA,IAAI,EAAE;MACnC,OAAO,IAAI;IACf;IAEA,IAAIoD,UAAU,GAAG,CAAC;IAElB,SAAS;MACL,IAAMC,KAAK,GAAGrD,IAAI,CAACsD,OAAO,CAACnB,iBAAiB,EAAEiB,UAAU,CAAC;MAEzD,IAAIC,KAAK,GAAG,CAAC,EAAE;QACX,OAAO,IAAI;MACf;MAEA,IAAMlC,MAAM,GAAGoB,iBAAiB,CAACvC,IAAI,EAAEqD,KAAK,CAAC;MAE7C,IAAIlC,MAAM,EAAE;QACR,OAAOA,MAAM;MACjB;;MAEA;MACA;MACAiC,UAAU,GAAGC,KAAK,GAAG,CAAC;IAC1B;EACJ;;EAEA;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACI,SAASE,sBAAsBA,CAAC5C,OAAO,EAAE;IACrC,IAAI,OAAOA,OAAO,KAAK,QAAQ,IAAI,CAACA,OAAO,EAAE;MACzC,OAAO,EAAE;IACb;IAEA,IAAI6C,MAAM,GAAG,EAAE;IACf,IAAIC,IAAI,GAAG,CAAC;;IAEZ;IACA;IACA;IACA,IAAIC,UAAU,GAAG,CAAC,CAAC;IAEnB,SAAS;MACL,IAAML,KAAK,GAAG1C,OAAO,CAAC2C,OAAO,CAAClB,oBAAoB,EAAEqB,IAAI,CAAC;MAEzD,IAAIJ,KAAK,GAAG,CAAC,EAAE;QACX,OAAOG,MAAM,GAAG7C,OAAO,CAACgD,KAAK,CAACF,IAAI,CAAC;MACvC;MAEA,IAAMG,GAAG,GAAGF,UAAU,IAAIL,KAAK,GAAGK,UAAU,GAAG/C,OAAO,CAAC2C,OAAO,CAAC,GAAG,EAAED,KAAK,CAAC;MAE1EK,UAAU,GAAGE,GAAG;MAEhB,IAAIA,GAAG,GAAG,CAAC,EAAE;QACT;QACA,OAAOJ,MAAM,GAAG7C,OAAO,CAACgD,KAAK,CAACF,IAAI,CAAC;MACvC;MAEA,IAAMI,OAAO,GAAGD,GAAG,GAAGP,KAAK,IAAIhB,uBAAuB,GAChDyB,mBAAmB,CAACnD,OAAO,CAACgD,KAAK,CAACN,KAAK,GAAGjB,oBAAoB,CAACF,MAAM,EAAE0B,GAAG,CAAC,CAAC,GAC5E,IAAI;MAEV,IAAI,CAACC,OAAO,EAAE;QACV;QACA;QACA;QACA;QACAL,MAAM,IAAI7C,OAAO,CAACgD,KAAK,CAACF,IAAI,EAAEJ,KAAK,GAAGjB,oBAAoB,CAACF,MAAM,CAAC;QAClEuB,IAAI,GAAGJ,KAAK,GAAGjB,oBAAoB,CAACF,MAAM;QAE1C;MACJ;;MAEA;MACA;MACA;MACA,IAAM6B,SAAS,GAAGF,OAAO,CAACG,GAAG,CAAC,UAAUC,MAAM,EAAE;QAAE,OAAO7B,oBAAoB,GAAG6B,MAAM,GAAG,GAAG;MAAE,CAAC,CAAC,CAACC,IAAI,CAAC,EAAE,CAAC;MACzG,IAAMC,QAAQ,GAAGxD,OAAO,CAACgD,KAAK,CAACN,KAAK,EAAEO,GAAG,GAAG,CAAC,CAAC;MAE9CJ,MAAM,IAAI7C,OAAO,CAACgD,KAAK,CAACF,IAAI,EAAEJ,KAAK,CAAC;MACpCG,MAAM,IAAIO,SAAS,KAAKI,QAAQ,GAAGA,QAAQ,GAAGJ,SAAS;MAEvDN,IAAI,GAAGG,GAAG,GAAG,CAAC;IAClB;EACJ;;EAEA;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACI;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACI,IAAMQ,sBAAsB,GAAG,4CAA4C;EAE3E,SAASN,mBAAmBA,CAACO,QAAQ,EAAE;IACnC,IAAMC,KAAK,GAAGD,QAAQ,CAACE,KAAK,CAAC,GAAG,CAAC;IACjC,IAAMV,OAAO,GAAG,EAAE;IAElB,KAAK,IAAIpB,KAAK,GAAG,CAAC,EAAEA,KAAK,GAAG6B,KAAK,CAACpC,MAAM,EAAEO,KAAK,EAAE,EAAE;MAC/C;MACA;MACA,IAAM+B,IAAI,GAAG/B,KAAK,KAAK,CAAC,GAAG6B,KAAK,CAAC7B,KAAK,CAAC,CAACvC,IAAI,CAAC,CAAC,GAAGoE,KAAK,CAAC7B,KAAK,CAAC,CAACvC,IAAI,CAAC,CAAC,CAACD,OAAO,CAAC,UAAU,EAAE,EAAE,CAAC;MAE5F,IAAI,OAAO,CAACwE,IAAI,CAACD,IAAI,CAAC,EAAE;QACpBX,OAAO,CAACa,IAAI,CAACF,IAAI,CAAC;QAElB;MACJ;;MAEA;MACA;MACA,IAAIX,OAAO,CAAC3B,MAAM,GAAG,CAAC,IAAIkC,sBAAsB,CAACK,IAAI,CAACD,IAAI,CAAC,EAAE;QACzD;MACJ;MAEA,OAAO,IAAI;IACf;IAEA,OAAOX,OAAO,CAAC3B,MAAM,GAAG,CAAC,GAAG2B,OAAO,GAAG,IAAI;EAC9C;;EAEA;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACI,SAASc,gBAAgBA,CAACC,KAAK,EAAExE,IAAI,EAAE;IACnC,IAAMJ,IAAI,GAAG,OAAO4E,KAAK,KAAK,QAAQ,GAAGA,KAAK,CAAC3E,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAACC,IAAI,CAAC,CAAC,CAAC2E,WAAW,CAAC,CAAC,GAAG,EAAE;IAC7F,IAAMC,MAAM,GAAG,OAAO1E,IAAI,KAAK,QAAQ,GAAGA,IAAI,CAACF,IAAI,CAAC,CAAC,CAAC2E,WAAW,CAAC,CAAC,GAAG,EAAE;;IAExE;IACA;IACA,OAAO7E,IAAI,GAAG,GAAG,GAAG8E,MAAM;EAC9B;;EAEA;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;EACI,IAAMC,iBAAiB,GAAG,gDAAgD;EAC1E,IAAMC,uBAAuB,GAAG,IAAIC,MAAM,CACtC,uCAAuC,GAAGF,iBAAiB,GAAG,iCAAiC,EAAE,GAAG,CAAC;EAEzG,SAASG,yBAAyBA,CAACC,IAAI,EAAE;IACrC,IAAI,OAAOA,IAAI,KAAK,QAAQ,IAAI,CAACA,IAAI,EAAE;MACnC,OAAO,EAAE;IACb;IAEA,OAAOA,IAAI,CAAClF,OAAO,CAAC+E,uBAAuB,EAAE,IAAI,CAAC;EACtD;;EAEA;AACJ;AACA;AACA;AACA;AACA;AACA;EACI,SAASI,yBAAyBA,CAACD,IAAI,EAAE;IACrC,IAAI,OAAOA,IAAI,KAAK,QAAQ,IAAI,CAACA,IAAI,EAAE;MACnC,OAAO,EAAE;IACb;IAEA,OAAOA,IAAI,CAAClF,OAAO,CAAC,cAAc,EAAE,wBAAwB,CAAC;EACjE;;EAEA;EACA;EACA,SAASoF,oBAAoBA,CAACrF,IAAI,EAAE;IAChC,OAAOA,IAAI,CAACC,OAAO,CAAC,UAAU,EAAE,UAAUI,SAAS,EAAE;MACjD,QAAQA,SAAS;QACb,KAAK,GAAG;UAAE,OAAO,OAAO;QACxB,KAAK,GAAG;UAAE,OAAO,MAAM;QACvB,KAAK,GAAG;UAAE,OAAO,MAAM;QACvB,KAAK,GAAG;UAAE,OAAO,QAAQ;QACzB;UAAS,OAAO,OAAO;MAC3B;IACJ,CAAC,CAAC;EACN;;EAEA;AACJ;AACA;AACA;AACA;AACA;AACA;AACA;EACI,SAASiF,kBAAkBA,CAACC,YAAY,EAAEX,KAAK,EAAE;IAC7C,IAAM5E,IAAI,GAAG,OAAO4E,KAAK,KAAK,QAAQ,GAAGA,KAAK,CAAC3E,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC,CAACC,IAAI,CAAC,CAAC,GAAG,EAAE;IAE/E,IAAI,CAACF,IAAI,EAAE;MACP,OAAO,OAAO,GAAGuF,YAAY,GAAG,QAAQ;IAC5C;IAEA,OAAO,cAAc,GAAGF,oBAAoB,CAACrF,IAAI,CAAC,GAAG,IAAI,GAAGuF,YAAY,GAAG,QAAQ;EACvF;;EAEA;EACA,OAAO;IACH7E,kBAAkB,EAAEA,kBAAkB;IACtCyC,eAAe,EAAEA,eAAe;IAChCwB,gBAAgB,EAAEA,gBAAgB;IAClCpB,sBAAsB,EAAEA,sBAAsB;IAC9C2B,yBAAyB,EAAEA,yBAAyB;IACpDE,yBAAyB,EAAEA,yBAAyB;IACpDE,kBAAkB,EAAEA;EACxB,CAAC;AACL,CAAC,CAAE,CAAC","file":"chat-markers.js","sourcesContent":["/*\n * Markers the host expands on the model's behalf.\n *\n * Retrieval hands the model a short label such as [fig:1] rather than a figure's address, because a model\n * asked to reproduce a long opaque identifier reproduces its shape and varies the digits instead --\n * producing addresses that look right and resolve to nothing. The model therefore never types an address,\n * and this turns the label it does type back into the picture.\n *\n * The chart tool works the same way from the other end: it hands back a [chart:{...}] marker and asks for it\n * verbatim, and the host turns that marker into a canvas. Reading a marker is a statement about a string, so\n * it lives here; drawing one is not, so each surface still does its own drawing.\n *\n * Loaded on its own so every chat surface shares one implementation: the two shared chat scripts and the\n * MVC chat-interaction view, which renders its markdown with its own inline marked setup.\n */\nwindow.CoreAIChatMarkers = window.CoreAIChatMarkers || (function () {\n    'use strict';\n\n    // Used when a figure has no caption, so the image is still announced as something.\n    const defaultImageAltText = 'Figure';\n\n    /*\n     * Makes a caption safe to sit inside the alt text of ![alt](link).\n     *\n     * Captions really do contain brackets, and an unescaped one closes the alt text early: a caption of\n     * \"Figure 2 (revised) [draft]\" would end the image after \"Figure 2 (revised) \" and leave the rest of the\n     * syntax on the page as markup. Parentheses are harmless there, only the brackets are. A caption that spans\n     * lines would end the paragraph the image lives in, so its whitespace is collapsed to single spaces.\n     */\n    function escapeImageAltText(title) {\n        const text = typeof title === 'string' ? title.replace(/\\s+/g, ' ').trim() : '';\n\n        if (!text) {\n            return defaultImageAltText;\n        }\n\n        // The backslash is escaped as well, or a caption ending in one would escape the bracket that closes\n        // the alt text and the image would swallow the rest of the line.\n        return text.replace(/[\\\\\\[\\]]/g, '\\\\$&');\n    }\n\n    /*\n     * Makes a link safe to sit inside the destination of ![alt](link): the few characters that would end the\n     * destination early are percent-encoded, which leaves the URL addressing the same resource. The link is\n     * still run through sanitizeUrl by the image renderer, so this is about parsing, not about safety.\n     */\n    function encodeImageLink(link) {\n        return link.replace(/[\\s()<>\"\\\\]/g, function (character) {\n            return '%' + character.charCodeAt(0).toString(16).toUpperCase().padStart(2, '0');\n        });\n    }\n\n    /*\n     * Returns the content with every marker the reference map describes as a servable image replaced by\n     * markdown image syntax, so the ordinary image renderer draws it with its thumbnail, its download button\n     * and its size cap. Pure: nothing here touches the DOM or the message.\n     *\n     * A marker is replaced only when its reference is an image and carries a link. Anything else is left\n     * exactly as written -- a marker the model invented for a figure that was never in the results, or one\n     * whose picture the host cannot serve, reaches the reader as the few characters the model typed rather\n     * than as a broken image.\n     */\n    function expandImageMarkers(content, references) {\n        if (typeof content !== 'string' || !content) {\n            return typeof content === 'string' ? content : '';\n        }\n\n        if (!references || typeof references !== 'object') {\n            return content;\n        }\n\n        let expanded = content;\n\n        for (const marker of Object.keys(references)) {\n            const reference = references[marker];\n\n            if (!marker || !reference || typeof reference !== 'object') {\n                continue;\n            }\n\n            if ((reference.isImage ?? reference.IsImage) !== true) {\n                continue;\n            }\n\n            const rawLink = reference.link ?? reference.Link;\n            const link = typeof rawLink === 'string' ? rawLink.trim() : '';\n\n            if (!link) {\n                continue;\n            }\n\n            const image = `![${escapeImageAltText(reference.title ?? reference.Title)}](${encodeImageLink(link)})`;\n\n            // Every occurrence, and through a replacer function so a '$' in a caption or a link is not read as\n            // a replacement pattern.\n            expanded = expanded.replaceAll(marker, function () { return image; });\n        }\n\n        return expanded;\n    }\n\n    // The marker the chart tool emits and asks the model to repeat verbatim, braces and all.\n    const chartMarkerPrefix = '[chart:';\n\n    // What a citation the model typed opens with. The reference map is keyed by the whole marker, so this is\n    // also how a rewritten one is put back together.\n    const citationMarkerPrefix = '[doc:';\n\n    /*\n     * How much text may sit between \"[doc:\" and its \"]\" before it stops being a citation.\n     *\n     * A list of reference numbers is short: even ten of them, written the long way, is under a hundred\n     * characters. The limit is what keeps the work per marker fixed, and so the whole pass linear -- without\n     * it, an opening bracket whose \"]\" is thousands of characters away is examined in full, and a paragraph\n     * full of such openings is examined once per opening.\n     */\n    const maxCitationMarkerLength = 512;\n\n    function isMarkerWhitespace(character) {\n        return character === ' ' || character === '\\n' || character === '\\r' || character === '\\t';\n    }\n\n    /*\n     * Reads the marker that begins at startIndex, or returns null when what begins there is not one.\n     *\n     * The configuration is JSON and JSON nests, so the end of the object is found by counting braces rather\n     * than by a regular expression: an expression that stopped at the first '}' would truncate every chart\n     * whose options carry an object of their own, which is all of them. The count has to know when it is\n     * inside a string as well -- an axis label really does read \"Q1 { Q2\", and a counter blind to strings\n     * would end the object in the middle of the label and hand the renderer a fragment. Escapes are tracked\n     * for the same reason, so a \\\" inside a label does not look like the end of it.\n     */\n    function readChartMarkerAt(text, startIndex) {\n        let index = startIndex + chartMarkerPrefix.length;\n\n        while (index < text.length && isMarkerWhitespace(text[index])) {\n            index++;\n        }\n\n        if (index >= text.length || text[index] !== '{') {\n            return null;\n        }\n\n        const jsonStart = index;\n        let depth = 0;\n        let inString = false;\n        let escaped = false;\n\n        for (; index < text.length; index++) {\n            const character = text[index];\n\n            if (inString) {\n                if (escaped) {\n                    escaped = false;\n                } else if (character === '\\\\') {\n                    escaped = true;\n                } else if (character === '\"') {\n                    inString = false;\n                }\n\n                continue;\n            }\n\n            if (character === '\"') {\n                inString = true;\n                continue;\n            }\n\n            if (character === '{') {\n                depth++;\n                continue;\n            }\n\n            if (character !== '}') {\n                continue;\n            }\n\n            depth--;\n\n            if (depth > 0) {\n                continue;\n            }\n\n            const jsonEnd = index + 1;\n            let closeIndex = jsonEnd;\n\n            while (closeIndex < text.length && isMarkerWhitespace(text[closeIndex])) {\n                closeIndex++;\n            }\n\n            // Only whitespace may sit between the object and the bracket that closes the marker. Hunting\n            // further ahead for a ']' would let a marker whose bracket the model dropped reach the one\n            // belonging to something else, and everything in between -- a paragraph of the answer, the next\n            // marker -- would be swallowed into the span the caller removes from the text.\n            if (closeIndex >= text.length || text[closeIndex] !== ']') {\n                return null;\n            }\n\n            return {\n                startIndex: startIndex,\n                endIndex: closeIndex + 1,\n                json: text.substring(jsonStart, jsonEnd)\n            };\n        }\n\n        return null;\n    }\n\n    /*\n     * Finds the first [chart:{...}] marker in the content and returns where it starts, where it ends and the\n     * JSON between, or null when there is no marker. Pure: it knows nothing of canvases, element ids or\n     * Chart.js, because the three chat surfaces draw a chart differently and agree only on what a marker is.\n     *\n     * Text that merely looks like a marker is not one, and none of these return a span: \"[chart:\" with no\n     * object after it, an object whose braces never balance, an object the model never closed with ']'. Each\n     * reaches the reader as the text the model typed, which is what the caller does with a null.\n     */\n    function findChartMarker(text) {\n        if (typeof text !== 'string' || !text) {\n            return null;\n        }\n\n        let searchFrom = 0;\n\n        for (;;) {\n            const start = text.indexOf(chartMarkerPrefix, searchFrom);\n\n            if (start < 0) {\n                return null;\n            }\n\n            const marker = readChartMarkerAt(text, start);\n\n            if (marker) {\n                return marker;\n            }\n\n            // Something that opened like a marker and was not one must not hide a real one further on, so the\n            // search carries on past it rather than giving up at the first disappointment.\n            searchFrom = start + 1;\n        }\n    }\n\n    /*\n     * Splits a citation marker the model wrote as one bracket into the separate markers it means.\n     *\n     * A reference is keyed by the literal string the model is asked to type -- \"[doc:1]\" -- and the renderer\n     * replaces exactly that. Models routinely gather several into one bracket instead: \"[doc:1, doc:2]\". No\n     * key matches that, so nothing is replaced and the reader is shown the raw marker in the middle of a\n     * sentence, while the citations the model happened to write singly render correctly beside it.\n     *\n     * Rewriting the combined form into the separate markers is what makes the rest of the pipeline see them.\n     * Both spellings are accepted, since a model that writes \"[doc:1, doc:2]\" will also write \"[doc:1, 2]\".\n     * A lone \"[doc:1]\" is left exactly as it is.\n     */\n    function splitCombinedCitations(content) {\n        if (typeof content !== 'string' || !content) {\n            return '';\n        }\n\n        let result = '';\n        let from = 0;\n\n        // The closing bracket last searched for. A ']' found from an earlier position is still the next one\n        // from here, because there was none in between -- so text like \"[doc:x[doc:x[doc:x...]\" is walked once\n        // rather than rescanned from every false start, which would make the whole pass quadratic.\n        let knownClose = -1;\n\n        for (;;) {\n            const start = content.indexOf(citationMarkerPrefix, from);\n\n            if (start < 0) {\n                return result + content.slice(from);\n            }\n\n            const end = knownClose >= start ? knownClose : content.indexOf(']', start);\n\n            knownClose = end;\n\n            if (end < 0) {\n                // Nothing closes it, so nothing here is a marker. The rest is returned as it was written.\n                return result + content.slice(from);\n            }\n\n            const numbers = end - start <= maxCitationMarkerLength\n                ? readCitationNumbers(content.slice(start + citationMarkerPrefix.length, end))\n                : null;\n\n            if (!numbers) {\n                // Whatever this opened, it was not a list of references. The search carries on from just\n                // after the prefix rather than past the bracket, because the ']' that was found may belong to\n                // a real marker further along -- \"[doc:1, and later [doc:3, doc:4]\" closes the second one, and\n                // skipping to it would hide it.\n                result += content.slice(from, start + citationMarkerPrefix.length);\n                from = start + citationMarkerPrefix.length;\n\n                continue;\n            }\n\n            // Rewritten whenever the bracket does not already read as the canonical marker: several numbers\n            // in one bracket, or a single number carrying a locator the model appended. A bare \"[doc:1]\"\n            // rewrites to itself, so it is passed through untouched rather than rebuilt.\n            const canonical = numbers.map(function (number) { return citationMarkerPrefix + number + ']'; }).join('');\n            const original = content.slice(start, end + 1);\n\n            result += content.slice(from, start);\n            result += canonical === original ? original : canonical;\n\n            from = end + 1;\n        }\n    }\n\n    /*\n     * Reads the numbers out of the inside of a citation bracket, or returns null when it does not hold a list\n     * of them.\n     *\n     * Deliberately not a regular expression over the whole bracket. The pattern that reads naturally --\n     * optional spaces on both sides of each comma, repeated -- gives the engine several ways to match the same\n     * text, and on a bracket that never closes it explores all of them: CodeQL flagged it as exponential, and\n     * it was. A model writing \"[doc:9, 9, 9, 9...\" and forgetting the bracket is not a hostile act, and it is\n     * a reader's own tab that stops responding.\n     *\n     * Splitting on the comma and checking each part on its own is linear, and each pattern below is anchored\n     * over a single character class, so neither can backtrack at all.\n     */\n    /*\n     * A locator a model appends to a citation instead of writing the bare marker it was asked for: \"pages\n     * 13-15\", \"page 4\", \"pp. 3-5\", \"pg 7\". It is not a citation number and the reference map holds nothing\n     * under it, so before this it made the whole marker unreadable and \"[doc:1, pages 13-15]\" reached the\n     * reader as those characters.\n     *\n     * Dropped rather than rendered. The page a citation points at is already carried by the reference the\n     * number resolves to, which is the one the host looked up rather than the one the model recalled.\n     */\n    const citationLocatorPattern = /^(?:pages?|pp?g?|pp)\\.?\\s*[\\d\\s‐-―,&+-]*$/i;\n\n    function readCitationNumbers(interior) {\n        const parts = interior.split(',');\n        const numbers = [];\n\n        for (let index = 0; index < parts.length; index++) {\n            // Only the parts after the first may repeat the prefix: \"[doc:1, doc:2]\" and \"[doc:1, 2]\" are both\n            // written by models, \"[doc:doc:1]\" is not.\n            const part = index === 0 ? parts[index].trim() : parts[index].trim().replace(/^doc:\\s*/, '');\n\n            if (/^\\d+$/.test(part)) {\n                numbers.push(part);\n\n                continue;\n            }\n\n            // A locator is allowed only after a number has been read, so \"[doc:pages 3]\" is still not a\n            // citation -- something that names no document is not one whatever follows it.\n            if (numbers.length > 0 && citationLocatorPattern.test(part)) {\n                continue;\n            }\n\n            return null;\n        }\n\n        return numbers.length > 0 ? numbers : null;\n    }\n\n    /*\n     * The identity of a citation as a reader meets it: the line it prints, and where it points.\n     *\n     * Retrieval returns one reference per chunk, so a single article that answered a question through three of\n     * its chunks arrives as three references. Numbering those separately prints \"1,2,3\" over the sentence and\n     * then lists the same title three times, which tells the reader there are three sources corroborating the\n     * claim when there is one. Two references that would print the same line and lead to the same place are\n     * one citation, and are numbered once.\n     *\n     * Case and surrounding whitespace are not part of the identity, because they are not part of what the\n     * reader sees as different. Anything else -- a different title, a different page, a link to one figure\n     * rather than another -- keeps the citations apart, since that is a distinction the reader can act on.\n     */\n    function citationIdentity(label, link) {\n        const text = typeof label === 'string' ? label.replace(/\\s+/g, ' ').trim().toLowerCase() : '';\n        const target = typeof link === 'string' ? link.trim().toLowerCase() : '';\n\n        // A null byte cannot occur in either part, so no pair of values can collide across the boundary:\n        // without it, label \"a\" + link \"bc\" and label \"ab\" + link \"c\" would be the same identity.\n        return text + '\u0000' + target;\n    }\n\n    /*\n     * Collapses a run of identical citation markers into one.\n     *\n     * Once two references share a number, a sentence that cited both ends with that number twice over --\n     * \"1,1\" -- which reads as a typo. Only an immediately repeated marker is removed: the same source cited\n     * again later in the paragraph is a separate citation of it and keeps its marker.\n     *\n     * The marker is matched with whatever attributes it carries, because the number is what makes two of them\n     * the same citation; a pattern written for a bare <sup> would silently stop collapsing the moment a\n     * marker gained a tooltip.\n     *\n     * What separates the two markers is swallowed with them, because the model writes its own punctuation\n     * between references -- \"[ref1], [ref2]\" -- and once both carry the same number that comma is left\n     * between a number and itself, printing \"1,1\". Only a separator that was joining the two citations is\n     * eaten: whitespace, a comma, or the marker the comma rule inserts. Anything else means the second\n     * marker is a fresh citation later in the sentence, and it is left alone.\n     */\n    const citationSeparator = '(?:\\\\s*(?:,|<sup(?:\\\\s[^>]*)?>,<\\\\/sup>)?\\\\s*)';\n    const repeatedCitationPattern = new RegExp(\n        '(<sup(?:\\\\s[^>]*)?>(\\\\d+)<\\\\/sup>)(?:' + citationSeparator + '<sup(?:\\\\s[^>]*)?>\\\\2<\\\\/sup>)+', 'g');\n\n    function collapseRepeatedCitations(html) {\n        if (typeof html !== 'string' || !html) {\n            return '';\n        }\n\n        return html.replace(repeatedCitationPattern, '$1');\n    }\n\n    /*\n     * Puts a comma between two citation markers that ended up side by side, so \"12\" reads as two sources\n     * rather than as the twelfth.\n     *\n     * Only the boundary between the markers is matched, so whatever attributes either one carries are left\n     * untouched. The inserted marker is not itself rescanned, since a replace walks the string it was given.\n     */\n    function separateAdjacentCitations(html) {\n        if (typeof html !== 'string' || !html) {\n            return '';\n        }\n\n        return html.replace(/<\\/sup><sup/g, '</sup><sup>,</sup><sup');\n    }\n\n    // Escapes the few characters that would end an HTML attribute early, so a caption containing a quote\n    // cannot introduce markup of its own.\n    function escapeAttributeValue(text) {\n        return text.replace(/[&<>\"']/g, function (character) {\n            switch (character) {\n                case '&': return '&amp;';\n                case '<': return '&lt;';\n                case '>': return '&gt;';\n                case '\"': return '&quot;';\n                default: return '&#39;';\n            }\n        });\n    }\n\n    /*\n     * The inline marker for a citation: its number, and the source it stands for as a tooltip.\n     *\n     * A bare number tells the reader only that something was cited, and answers which source it was by\n     * making them look away from the sentence and match a number against a list underneath. Carrying the\n     * label means the answer arrives where the question is asked. The list stays as it is, since a tooltip\n     * is not reachable by keyboard and cannot be the only place the source appears.\n     */\n    function citationMarkerHtml(displayIndex, label) {\n        const text = typeof label === 'string' ? label.replace(/\\s+/g, ' ').trim() : '';\n\n        if (!text) {\n            return '<sup>' + displayIndex + '</sup>';\n        }\n\n        return '<sup title=\"' + escapeAttributeValue(text) + '\">' + displayIndex + '</sup>';\n    }\n\n    // Deliberately small: pure functions over strings, for the hosts below and for the tests.\n    return {\n        expandImageMarkers: expandImageMarkers,\n        findChartMarker: findChartMarker,\n        citationIdentity: citationIdentity,\n        splitCombinedCitations: splitCombinedCitations,\n        collapseRepeatedCitations: collapseRepeatedCitations,\n        separateAdjacentCitations: separateAdjacentCitations,\n        citationMarkerHtml: citationMarkerHtml\n    };\n})();\n"]}

package/dist/chat-markers.min.js

@@ -0,0 +1,2 @@
+function _typeof(t){return _typeof="function"==typeof Symbol&&"symbol"==typeof Symbol.iterator?function(t){return typeof t}:function(t){return t&&"function"==typeof Symbol&&t.constructor===Symbol&&t!==Symbol.prototype?"symbol":typeof t},_typeof(t)}window.CoreAIChatMarkers=window.CoreAIChatMarkers||function(){"use strict";var t="[chart:",r="[doc:";function e(t){return" "===t||"\n"===t||"\r"===t||"\t"===t}function n(t,r){for(var n=r+7;n<t.length&&e(t[n]);)n++;if(n>=t.length||"{"!==t[n])return null;for(var i=n,o=0,u=!1,s=!1;n<t.length;n++){var f=t[n];if(u)s?s=!1:"\\"===f?s=!0:'"'===f&&(u=!1);else if('"'!==f)if("{"!==f){if("}"===f&&!(--o>0)){for(var a=n+1,l=a;l<t.length&&e(t[l]);)l++;return l>=t.length||"]"!==t[l]?null:{startIndex:r,endIndex:l+1,json:t.substring(i,a)}}}else o++;else u=!0}return null}var i=/^(?:pages?|pp?g?|pp)\.?\s*[\d\s‐-―,&+-]*$/i;function o(t){for(var r=t.split(","),e=[],n=0;n<r.length;n++){var o=0===n?r[n].trim():r[n].trim().replace(/^doc:\s*/,"");if(/^\d+$/.test(o))e.push(o);else if(!(e.length>0&&i.test(o)))return null}return e.length>0?e:null}var u=new RegExp("(<sup(?:\\s[^>]*)?>(\\d+)<\\/sup>)(?:(?:\\s*(?:,|<sup(?:\\s[^>]*)?>,<\\/sup>)?\\s*)<sup(?:\\s[^>]*)?>\\2<\\/sup>)+","g");return{expandImageMarkers:function(t,r){if("string"!=typeof t||!t)return"string"==typeof t?t:"";if(!r||"object"!==_typeof(r))return t;for(var e=t,n=function(){var t,n,u,s=o[i],f=r[s];if(!s||!f||"object"!==_typeof(f))return 0;if(!0!==(null!==(t=f.isImage)&&void 0!==t?t:f.IsImage))return 0;var a=null!==(n=f.link)&&void 0!==n?n:f.Link,l="string"==typeof a?a.trim():"";if(!l)return 0;var p,c,g="![".concat((p=null!==(u=f.title)&&void 0!==u?u:f.Title,c="string"==typeof p?p.replace(/\s+/g," ").trim():"",c?c.replace(/[\\\[\]]/g,"\\$&"):"Figure"),"](").concat(function(t){return t.replace(/[\s()<>"\\]/g,function(t){return"%"+t.charCodeAt(0).toString(16).toUpperCase().padStart(2,"0")})}(l),")");e=e.replaceAll(s,function(){return g})},i=0,o=Object.keys(r);i<o.length;i++)n();return e},findChartMarker:function(r){if("string"!=typeof r||!r)return null;for(var e=0;;){var i=r.indexOf(t,e);if(i<0)return null;var o=n(r,i);if(o)return o;e=i+1}},citationIdentity:function(t,r){return("string"==typeof t?t.replace(/\s+/g," ").trim().toLowerCase():"")+"\0"+("string"==typeof r?r.trim().toLowerCase():"")},splitCombinedCitations:function(t){if("string"!=typeof t||!t)return"";for(var e="",n=0,i=-1;;){var u=t.indexOf(r,n);if(u<0)return e+t.slice(n);var s=i>=u?i:t.indexOf("]",u);if(i=s,s<0)return e+t.slice(n);var f=s-u<=512?o(t.slice(u+5,s)):null;if(f){var a=f.map(function(t){return r+t+"]"}).join(""),l=t.slice(u,s+1);e+=t.slice(n,u),e+=a===l?l:a,n=s+1}else e+=t.slice(n,u+5),n=u+5}},collapseRepeatedCitations:function(t){return"string"==typeof t&&t?t.replace(u,"$1"):""},separateAdjacentCitations:function(t){return"string"==typeof t&&t?t.replace(/<\/sup><sup/g,"</sup><sup>,</sup><sup"):""},citationMarkerHtml:function(t,r){var e="string"==typeof r?r.replace(/\s+/g," ").trim():"";return e?'<sup title="'+function(t){return t.replace(/[&<>"']/g,function(t){switch(t){case"&":return"&amp;";case"<":return"&lt;";case">":return"&gt;";case'"':return"&quot;";default:return"&#39;"}})}(e)+'">'+t+"</sup>":"<sup>"+t+"</sup>"}}}();
+//# sourceMappingURL=chat-markers.min.js.map

package/dist/chat-markers.min.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["chat-markers.js"],"names":["_typeof","o","Symbol","iterator","constructor","prototype","window","CoreAIChatMarkers","chartMarkerPrefix","citationMarkerPrefix","isMarkerWhitespace","character","readChartMarkerAt","text","startIndex","index","length","jsonStart","depth","inString","escaped","jsonEnd","closeIndex","endIndex","json","substring","citationLocatorPattern","readCitationNumbers","interior","parts","split","numbers","part","trim","replace","test","push","repeatedCitationPattern","RegExp","expandImageMarkers","content","references","expanded","_loop","_reference$isImage","_reference$link","_reference$title","marker","_Object$keys","_i","reference","isImage","IsImage","rawLink","link","Link","title","image","concat","Title","charCodeAt","toString","toUpperCase","padStart","encodeImageLink","replaceAll","Object","keys","findChartMarker","searchFrom","start","indexOf","citationIdentity","label","toLowerCase","splitCombinedCitations","result","from","knownClose","slice","end","canonical","map","number","join","original","collapseRepeatedCitations","html","separateAdjacentCitations","citationMarkerHtml","displayIndex","escapeAttributeValue"],"mappings":"AAKA,SAASA,QAAQC,GAAgC,OAAOD,QAAU,mBAAqBE,QAAU,iBAAmBA,OAAOC,SAAW,SAAUF,GAAK,cAAcA,CAAG,EAAI,SAAUA,GAAK,OAAOA,GAAK,mBAAqBC,QAAUD,EAAEG,cAAgBF,QAAUD,IAAMC,OAAOG,UAAY,gBAAkBJ,CAAG,EAAGD,QAAQC,EAAI,CAU7TK,OAAOC,kBAAoBD,OAAOC,mBAAsB,WACpD,aAsFA,IAAMC,EAAoB,UAIpBC,EAAuB,QAY7B,SAASC,EAAmBC,GACxB,MAAqB,MAAdA,GAAmC,OAAdA,GAAoC,OAAdA,GAAoC,OAAdA,CAC5E,CAYA,SAASC,EAAkBC,EAAMC,GAG7B,IAFA,IAAIC,EAAQD,EAAaN,EAElBO,EAAQF,EAAKG,QAAUN,EAAmBG,EAAKE,KAClDA,IAGJ,GAAIA,GAASF,EAAKG,QAA0B,MAAhBH,EAAKE,GAC7B,OAAO,KAQX,IALA,IAAME,EAAYF,EACdG,EAAQ,EACRC,GAAW,EACXC,GAAU,EAEPL,EAAQF,EAAKG,OAAQD,IAAS,CACjC,IAAMJ,EAAYE,EAAKE,GAEvB,GAAII,EACIC,EACAA,GAAU,EACW,OAAdT,EACPS,GAAU,EACW,MAAdT,IACPQ,GAAW,QAMnB,GAAkB,MAAdR,EAKJ,GAAkB,MAAdA,GAKJ,GAAkB,MAAdA,OAIJO,EAEY,GAAZ,CAOA,IAHA,IAAMG,EAAUN,EAAQ,EACpBO,EAAaD,EAEVC,EAAaT,EAAKG,QAAUN,EAAmBG,EAAKS,KACvDA,IAOJ,OAAIA,GAAcT,EAAKG,QAA+B,MAArBH,EAAKS,GAC3B,KAGJ,CACHR,WAAYA,EACZS,SAAUD,EAAa,EACvBE,KAAMX,EAAKY,UAAUR,EAAWI,GApBpC,OAZIH,SALAC,GAAW,CAuCnB,CAEA,OAAO,IACX,CAgIA,IAAMO,EAAyB,6CAE/B,SAASC,EAAoBC,GAIzB,IAHA,IAAMC,EAAQD,EAASE,MAAM,KACvBC,EAAU,GAEPhB,EAAQ,EAAGA,EAAQc,EAAMb,OAAQD,IAAS,CAG/C,IAAMiB,EAAiB,IAAVjB,EAAcc,EAAMd,GAAOkB,OAASJ,EAAMd,GAAOkB,OAAOC,QAAQ,WAAY,IAEzF,GAAI,QAAQC,KAAKH,GACbD,EAAQK,KAAKJ,QAOjB,KAAID,EAAQf,OAAS,GAAKU,EAAuBS,KAAKH,IAItD,OAAO,IACX,CAEA,OAAOD,EAAQf,OAAS,EAAIe,EAAU,IAC1C,CAyCA,IACMM,EAA0B,IAAIC,OAChC,qHAAiG,KA0DrG,MAAO,CACHC,mBAjZJ,SAA4BC,EAASC,GACjC,GAAuB,iBAAZD,IAAyBA,EAChC,MAA0B,iBAAZA,EAAuBA,EAAU,GAGnD,IAAKC,GAAoC,WAAtBzC,QAAOyC,GACtB,OAAOD,EAKX,IAFA,IAAIE,EAAWF,EAAQG,EAAA,WAEuB,IAAAC,EAAAC,EAAAC,EAAnCC,EAAMC,EAAAC,GACPC,EAAYT,EAAWM,GAE7B,IAAKA,IAAWG,GAAkC,WAArBlD,QAAOkD,GAAwB,OAAA,EAI5D,IAAiD,KAA3B,QAAlBN,EAACM,EAAUC,eAAO,IAAAP,EAAAA,EAAIM,EAAUE,SAAmB,OAAA,EAIvD,IAAMC,EAAwB,QAAjBR,EAAGK,EAAUI,YAAI,IAAAT,EAAAA,EAAIK,EAAUK,KACtCD,EAA0B,iBAAZD,EAAuBA,EAAQpB,OAAS,GAE5D,IAAKqB,EAAM,OAAA,EAIX,IA9DoBE,EAClB3C,EA6DI4C,EAAK,KAAAC,QA9DSF,EA8DiC,QAAhBV,EAACI,EAAUM,aAAK,IAAAV,EAAAA,EAAII,EAAUS,MA7DjE9C,EAAwB,iBAAV2C,EAAqBA,EAAMtB,QAAQ,OAAQ,KAAKD,OAAS,GAExEpB,EAMEA,EAAKqB,QAAQ,YAAa,QAnBT,UAwEqD,MAAAwB,OA7CjF,SAAyBJ,GACrB,OAAOA,EAAKpB,QAAQ,eAAgB,SAAUvB,GAC1C,MAAO,IAAMA,EAAUiD,WAAW,GAAGC,SAAS,IAAIC,cAAcC,SAAS,EAAG,IAChF,EACJ,CAyCsFC,CAAgBV,GAAK,KAInGZ,EAAWA,EAASuB,WAAWlB,EAAQ,WAAc,OAAOU,CAAO,EACvE,EAvBAR,EAAA,EAAAD,EAAqBkB,OAAOC,KAAK1B,GAAWQ,EAAAD,EAAAhC,OAAAiC,IAAAN,IAyB5C,OAAOD,CACX,EA6WI0B,gBAvPJ,SAAyBvD,GACrB,GAAoB,iBAATA,IAAsBA,EAC7B,OAAO,KAKX,IAFA,IAAIwD,EAAa,IAER,CACL,IAAMC,EAAQzD,EAAK0D,QAAQ/D,EAAmB6D,GAE9C,GAAIC,EAAQ,EACR,OAAO,KAGX,IAAMvB,EAASnC,EAAkBC,EAAMyD,GAEvC,GAAIvB,EACA,OAAOA,EAKXsB,EAAaC,EAAQ,CACzB,CACJ,EAgOIE,iBAzFJ,SAA0BC,EAAOnB,GAM7B,OAL8B,iBAAVmB,EAAqBA,EAAMvC,QAAQ,OAAQ,KAAKD,OAAOyC,cAAgB,IAK7E,MAJiB,iBAATpB,EAAoBA,EAAKrB,OAAOyC,cAAgB,GAK1E,EAmFIC,uBAnNJ,SAAgCnC,GAC5B,GAAuB,iBAAZA,IAAyBA,EAChC,MAAO,GAWX,IARA,IAAIoC,EAAS,GACTC,EAAO,EAKPC,GAAc,IAET,CACL,IAAMR,EAAQ9B,EAAQ+B,QAAQ9D,EAAsBoE,GAEpD,GAAIP,EAAQ,EACR,OAAOM,EAASpC,EAAQuC,MAAMF,GAGlC,IAAMG,EAAMF,GAAcR,EAAQQ,EAAatC,EAAQ+B,QAAQ,IAAKD,GAIpE,GAFAQ,EAAaE,EAETA,EAAM,EAEN,OAAOJ,EAASpC,EAAQuC,MAAMF,GAGlC,IAAM9C,EAAUiD,EAAMV,GAxKE,IAyKlB3C,EAAoBa,EAAQuC,MAAMT,EAAQ7D,EAA6BuE,IACvE,KAEN,GAAKjD,EAAL,CAcA,IAAMkD,EAAYlD,EAAQmD,IAAI,SAAUC,GAAU,OAAO1E,EAAuB0E,EAAS,GAAK,GAAGC,KAAK,IAChGC,EAAW7C,EAAQuC,MAAMT,EAAOU,EAAM,GAE5CJ,GAAUpC,EAAQuC,MAAMF,EAAMP,GAC9BM,GAAUK,IAAcI,EAAWA,EAAWJ,EAE9CJ,EAAOG,EAAM,CAXb,MAJIJ,GAAUpC,EAAQuC,MAAMF,EAAMP,EAAQ7D,GACtCoE,EAAOP,EAAQ7D,CAevB,CACJ,EA6JI6E,0BA7DJ,SAAmCC,GAC/B,MAAoB,iBAATA,GAAsBA,EAI1BA,EAAKrD,QAAQG,EAAyB,MAHlC,EAIf,EAwDImD,0BA/CJ,SAAmCD,GAC/B,MAAoB,iBAATA,GAAsBA,EAI1BA,EAAKrD,QAAQ,eAAgB,0BAHzB,EAIf,EA0CIuD,mBAlBJ,SAA4BC,EAAcjB,GACtC,IAAM5D,EAAwB,iBAAV4D,EAAqBA,EAAMvC,QAAQ,OAAQ,KAAKD,OAAS,GAE7E,OAAKpB,EAIE,eA3BX,SAA8BA,GAC1B,OAAOA,EAAKqB,QAAQ,WAAY,SAAUvB,GACtC,OAAQA,GACJ,IAAK,IAAK,MAAO,QACjB,IAAK,IAAK,MAAO,OACjB,IAAK,IAAK,MAAO,OACjB,IAAK,IAAK,MAAO,SACjB,QAAS,MAAO,QAExB,EACJ,CAiB4BgF,CAAqB9E,GAAQ,KAAO6E,EAAe,SAHhE,QAAUA,EAAe,QAIxC,EAYJ,CAxcwD","file":"chat-markers.min.js","sourcesContent":["/*\n * Markers the host expands on the model's behalf.\n *\n * Retrieval hands the model a short label such as [fig:1] rather than a figure's address, because a model\n * asked to reproduce a long opaque identifier reproduces its shape and varies the digits instead --\n * producing addresses that look right and resolve to nothing. The model therefore never types an address,\n * and this turns the label it does type back into the picture.\n *\n * The chart tool works the same way from the other end: it hands back a [chart:{...}] marker and asks for it\n * verbatim, and the host turns that marker into a canvas. Reading a marker is a statement about a string, so\n * it lives here; drawing one is not, so each surface still does its own drawing.\n *\n * Loaded on its own so every chat surface shares one implementation: the two shared chat scripts and the\n * MVC chat-interaction view, which renders its markdown with its own inline marked setup.\n */\nwindow.CoreAIChatMarkers = window.CoreAIChatMarkers || (function () {\n    'use strict';\n\n    // Used when a figure has no caption, so the image is still announced as something.\n    const defaultImageAltText = 'Figure';\n\n    /*\n     * Makes a caption safe to sit inside the alt text of ![alt](link).\n     *\n     * Captions really do contain brackets, and an unescaped one closes the alt text early: a caption of\n     * \"Figure 2 (revised) [draft]\" would end the image after \"Figure 2 (revised) \" and leave the rest of the\n     * syntax on the page as markup. Parentheses are harmless there, only the brackets are. A caption that spans\n     * lines would end the paragraph the image lives in, so its whitespace is collapsed to single spaces.\n     */\n    function escapeImageAltText(title) {\n        const text = typeof title === 'string' ? title.replace(/\\s+/g, ' ').trim() : '';\n\n        if (!text) {\n            return defaultImageAltText;\n        }\n\n        // The backslash is escaped as well, or a caption ending in one would escape the bracket that closes\n        // the alt text and the image would swallow the rest of the line.\n        return text.replace(/[\\\\\\[\\]]/g, '\\\\$&');\n    }\n\n    /*\n     * Makes a link safe to sit inside the destination of ![alt](link): the few characters that would end the\n     * destination early are percent-encoded, which leaves the URL addressing the same resource. The link is\n     * still run through sanitizeUrl by the image renderer, so this is about parsing, not about safety.\n     */\n    function encodeImageLink(link) {\n        return link.replace(/[\\s()<>\"\\\\]/g, function (character) {\n            return '%' + character.charCodeAt(0).toString(16).toUpperCase().padStart(2, '0');\n        });\n    }\n\n    /*\n     * Returns the content with every marker the reference map describes as a servable image replaced by\n     * markdown image syntax, so the ordinary image renderer draws it with its thumbnail, its download button\n     * and its size cap. Pure: nothing here touches the DOM or the message.\n     *\n     * A marker is replaced only when its reference is an image and carries a link. Anything else is left\n     * exactly as written -- a marker the model invented for a figure that was never in the results, or one\n     * whose picture the host cannot serve, reaches the reader as the few characters the model typed rather\n     * than as a broken image.\n     */\n    function expandImageMarkers(content, references) {\n        if (typeof content !== 'string' || !content) {\n            return typeof content === 'string' ? content : '';\n        }\n\n        if (!references || typeof references !== 'object') {\n            return content;\n        }\n\n        let expanded = content;\n\n        for (const marker of Object.keys(references)) {\n            const reference = references[marker];\n\n            if (!marker || !reference || typeof reference !== 'object') {\n                continue;\n            }\n\n            if ((reference.isImage ?? reference.IsImage) !== true) {\n                continue;\n            }\n\n            const rawLink = reference.link ?? reference.Link;\n            const link = typeof rawLink === 'string' ? rawLink.trim() : '';\n\n            if (!link) {\n                continue;\n            }\n\n            const image = `![${escapeImageAltText(reference.title ?? reference.Title)}](${encodeImageLink(link)})`;\n\n            // Every occurrence, and through a replacer function so a '$' in a caption or a link is not read as\n            // a replacement pattern.\n            expanded = expanded.replaceAll(marker, function () { return image; });\n        }\n\n        return expanded;\n    }\n\n    // The marker the chart tool emits and asks the model to repeat verbatim, braces and all.\n    const chartMarkerPrefix = '[chart:';\n\n    // What a citation the model typed opens with. The reference map is keyed by the whole marker, so this is\n    // also how a rewritten one is put back together.\n    const citationMarkerPrefix = '[doc:';\n\n    /*\n     * How much text may sit between \"[doc:\" and its \"]\" before it stops being a citation.\n     *\n     * A list of reference numbers is short: even ten of them, written the long way, is under a hundred\n     * characters. The limit is what keeps the work per marker fixed, and so the whole pass linear -- without\n     * it, an opening bracket whose \"]\" is thousands of characters away is examined in full, and a paragraph\n     * full of such openings is examined once per opening.\n     */\n    const maxCitationMarkerLength = 512;\n\n    function isMarkerWhitespace(character) {\n        return character === ' ' || character === '\\n' || character === '\\r' || character === '\\t';\n    }\n\n    /*\n     * Reads the marker that begins at startIndex, or returns null when what begins there is not one.\n     *\n     * The configuration is JSON and JSON nests, so the end of the object is found by counting braces rather\n     * than by a regular expression: an expression that stopped at the first '}' would truncate every chart\n     * whose options carry an object of their own, which is all of them. The count has to know when it is\n     * inside a string as well -- an axis label really does read \"Q1 { Q2\", and a counter blind to strings\n     * would end the object in the middle of the label and hand the renderer a fragment. Escapes are tracked\n     * for the same reason, so a \\\" inside a label does not look like the end of it.\n     */\n    function readChartMarkerAt(text, startIndex) {\n        let index = startIndex + chartMarkerPrefix.length;\n\n        while (index < text.length && isMarkerWhitespace(text[index])) {\n            index++;\n        }\n\n        if (index >= text.length || text[index] !== '{') {\n            return null;\n        }\n\n        const jsonStart = index;\n        let depth = 0;\n        let inString = false;\n        let escaped = false;\n\n        for (; index < text.length; index++) {\n            const character = text[index];\n\n            if (inString) {\n                if (escaped) {\n                    escaped = false;\n                } else if (character === '\\\\') {\n                    escaped = true;\n                } else if (character === '\"') {\n                    inString = false;\n                }\n\n                continue;\n            }\n\n            if (character === '\"') {\n                inString = true;\n                continue;\n            }\n\n            if (character === '{') {\n                depth++;\n                continue;\n            }\n\n            if (character !== '}') {\n                continue;\n            }\n\n            depth--;\n\n            if (depth > 0) {\n                continue;\n            }\n\n            const jsonEnd = index + 1;\n            let closeIndex = jsonEnd;\n\n            while (closeIndex < text.length && isMarkerWhitespace(text[closeIndex])) {\n                closeIndex++;\n            }\n\n            // Only whitespace may sit between the object and the bracket that closes the marker. Hunting\n            // further ahead for a ']' would let a marker whose bracket the model dropped reach the one\n            // belonging to something else, and everything in between -- a paragraph of the answer, the next\n            // marker -- would be swallowed into the span the caller removes from the text.\n            if (closeIndex >= text.length || text[closeIndex] !== ']') {\n                return null;\n            }\n\n            return {\n                startIndex: startIndex,\n                endIndex: closeIndex + 1,\n                json: text.substring(jsonStart, jsonEnd)\n            };\n        }\n\n        return null;\n    }\n\n    /*\n     * Finds the first [chart:{...}] marker in the content and returns where it starts, where it ends and the\n     * JSON between, or null when there is no marker. Pure: it knows nothing of canvases, element ids or\n     * Chart.js, because the three chat surfaces draw a chart differently and agree only on what a marker is.\n     *\n     * Text that merely looks like a marker is not one, and none of these return a span: \"[chart:\" with no\n     * object after it, an object whose braces never balance, an object the model never closed with ']'. Each\n     * reaches the reader as the text the model typed, which is what the caller does with a null.\n     */\n    function findChartMarker(text) {\n        if (typeof text !== 'string' || !text) {\n            return null;\n        }\n\n        let searchFrom = 0;\n\n        for (;;) {\n            const start = text.indexOf(chartMarkerPrefix, searchFrom);\n\n            if (start < 0) {\n                return null;\n            }\n\n            const marker = readChartMarkerAt(text, start);\n\n            if (marker) {\n                return marker;\n            }\n\n            // Something that opened like a marker and was not one must not hide a real one further on, so the\n            // search carries on past it rather than giving up at the first disappointment.\n            searchFrom = start + 1;\n        }\n    }\n\n    /*\n     * Splits a citation marker the model wrote as one bracket into the separate markers it means.\n     *\n     * A reference is keyed by the literal string the model is asked to type -- \"[doc:1]\" -- and the renderer\n     * replaces exactly that. Models routinely gather several into one bracket instead: \"[doc:1, doc:2]\". No\n     * key matches that, so nothing is replaced and the reader is shown the raw marker in the middle of a\n     * sentence, while the citations the model happened to write singly render correctly beside it.\n     *\n     * Rewriting the combined form into the separate markers is what makes the rest of the pipeline see them.\n     * Both spellings are accepted, since a model that writes \"[doc:1, doc:2]\" will also write \"[doc:1, 2]\".\n     * A lone \"[doc:1]\" is left exactly as it is.\n     */\n    function splitCombinedCitations(content) {\n        if (typeof content !== 'string' || !content) {\n            return '';\n        }\n\n        let result = '';\n        let from = 0;\n\n        // The closing bracket last searched for. A ']' found from an earlier position is still the next one\n        // from here, because there was none in between -- so text like \"[doc:x[doc:x[doc:x...]\" is walked once\n        // rather than rescanned from every false start, which would make the whole pass quadratic.\n        let knownClose = -1;\n\n        for (;;) {\n            const start = content.indexOf(citationMarkerPrefix, from);\n\n            if (start < 0) {\n                return result + content.slice(from);\n            }\n\n            const end = knownClose >= start ? knownClose : content.indexOf(']', start);\n\n            knownClose = end;\n\n            if (end < 0) {\n                // Nothing closes it, so nothing here is a marker. The rest is returned as it was written.\n                return result + content.slice(from);\n            }\n\n            const numbers = end - start <= maxCitationMarkerLength\n                ? readCitationNumbers(content.slice(start + citationMarkerPrefix.length, end))\n                : null;\n\n            if (!numbers) {\n                // Whatever this opened, it was not a list of references. The search carries on from just\n                // after the prefix rather than past the bracket, because the ']' that was found may belong to\n                // a real marker further along -- \"[doc:1, and later [doc:3, doc:4]\" closes the second one, and\n                // skipping to it would hide it.\n                result += content.slice(from, start + citationMarkerPrefix.length);\n                from = start + citationMarkerPrefix.length;\n\n                continue;\n            }\n\n            // Rewritten whenever the bracket does not already read as the canonical marker: several numbers\n            // in one bracket, or a single number carrying a locator the model appended. A bare \"[doc:1]\"\n            // rewrites to itself, so it is passed through untouched rather than rebuilt.\n            const canonical = numbers.map(function (number) { return citationMarkerPrefix + number + ']'; }).join('');\n            const original = content.slice(start, end + 1);\n\n            result += content.slice(from, start);\n            result += canonical === original ? original : canonical;\n\n            from = end + 1;\n        }\n    }\n\n    /*\n     * Reads the numbers out of the inside of a citation bracket, or returns null when it does not hold a list\n     * of them.\n     *\n     * Deliberately not a regular expression over the whole bracket. The pattern that reads naturally --\n     * optional spaces on both sides of each comma, repeated -- gives the engine several ways to match the same\n     * text, and on a bracket that never closes it explores all of them: CodeQL flagged it as exponential, and\n     * it was. A model writing \"[doc:9, 9, 9, 9...\" and forgetting the bracket is not a hostile act, and it is\n     * a reader's own tab that stops responding.\n     *\n     * Splitting on the comma and checking each part on its own is linear, and each pattern below is anchored\n     * over a single character class, so neither can backtrack at all.\n     */\n    /*\n     * A locator a model appends to a citation instead of writing the bare marker it was asked for: \"pages\n     * 13-15\", \"page 4\", \"pp. 3-5\", \"pg 7\". It is not a citation number and the reference map holds nothing\n     * under it, so before this it made the whole marker unreadable and \"[doc:1, pages 13-15]\" reached the\n     * reader as those characters.\n     *\n     * Dropped rather than rendered. The page a citation points at is already carried by the reference the\n     * number resolves to, which is the one the host looked up rather than the one the model recalled.\n     */\n    const citationLocatorPattern = /^(?:pages?|pp?g?|pp)\\.?\\s*[\\d\\s‐-―,&+-]*$/i;\n\n    function readCitationNumbers(interior) {\n        const parts = interior.split(',');\n        const numbers = [];\n\n        for (let index = 0; index < parts.length; index++) {\n            // Only the parts after the first may repeat the prefix: \"[doc:1, doc:2]\" and \"[doc:1, 2]\" are both\n            // written by models, \"[doc:doc:1]\" is not.\n            const part = index === 0 ? parts[index].trim() : parts[index].trim().replace(/^doc:\\s*/, '');\n\n            if (/^\\d+$/.test(part)) {\n                numbers.push(part);\n\n                continue;\n            }\n\n            // A locator is allowed only after a number has been read, so \"[doc:pages 3]\" is still not a\n            // citation -- something that names no document is not one whatever follows it.\n            if (numbers.length > 0 && citationLocatorPattern.test(part)) {\n                continue;\n            }\n\n            return null;\n        }\n\n        return numbers.length > 0 ? numbers : null;\n    }\n\n    /*\n     * The identity of a citation as a reader meets it: the line it prints, and where it points.\n     *\n     * Retrieval returns one reference per chunk, so a single article that answered a question through three of\n     * its chunks arrives as three references. Numbering those separately prints \"1,2,3\" over the sentence and\n     * then lists the same title three times, which tells the reader there are three sources corroborating the\n     * claim when there is one. Two references that would print the same line and lead to the same place are\n     * one citation, and are numbered once.\n     *\n     * Case and surrounding whitespace are not part of the identity, because they are not part of what the\n     * reader sees as different. Anything else -- a different title, a different page, a link to one figure\n     * rather than another -- keeps the citations apart, since that is a distinction the reader can act on.\n     */\n    function citationIdentity(label, link) {\n        const text = typeof label === 'string' ? label.replace(/\\s+/g, ' ').trim().toLowerCase() : '';\n        const target = typeof link === 'string' ? link.trim().toLowerCase() : '';\n\n        // A null byte cannot occur in either part, so no pair of values can collide across the boundary:\n        // without it, label \"a\" + link \"bc\" and label \"ab\" + link \"c\" would be the same identity.\n        return text + '\u0000' + target;\n    }\n\n    /*\n     * Collapses a run of identical citation markers into one.\n     *\n     * Once two references share a number, a sentence that cited both ends with that number twice over --\n     * \"1,1\" -- which reads as a typo. Only an immediately repeated marker is removed: the same source cited\n     * again later in the paragraph is a separate citation of it and keeps its marker.\n     *\n     * The marker is matched with whatever attributes it carries, because the number is what makes two of them\n     * the same citation; a pattern written for a bare <sup> would silently stop collapsing the moment a\n     * marker gained a tooltip.\n     *\n     * What separates the two markers is swallowed with them, because the model writes its own punctuation\n     * between references -- \"[ref1], [ref2]\" -- and once both carry the same number that comma is left\n     * between a number and itself, printing \"1,1\". Only a separator that was joining the two citations is\n     * eaten: whitespace, a comma, or the marker the comma rule inserts. Anything else means the second\n     * marker is a fresh citation later in the sentence, and it is left alone.\n     */\n    const citationSeparator = '(?:\\\\s*(?:,|<sup(?:\\\\s[^>]*)?>,<\\\\/sup>)?\\\\s*)';\n    const repeatedCitationPattern = new RegExp(\n        '(<sup(?:\\\\s[^>]*)?>(\\\\d+)<\\\\/sup>)(?:' + citationSeparator + '<sup(?:\\\\s[^>]*)?>\\\\2<\\\\/sup>)+', 'g');\n\n    function collapseRepeatedCitations(html) {\n        if (typeof html !== 'string' || !html) {\n            return '';\n        }\n\n        return html.replace(repeatedCitationPattern, '$1');\n    }\n\n    /*\n     * Puts a comma between two citation markers that ended up side by side, so \"12\" reads as two sources\n     * rather than as the twelfth.\n     *\n     * Only the boundary between the markers is matched, so whatever attributes either one carries are left\n     * untouched. The inserted marker is not itself rescanned, since a replace walks the string it was given.\n     */\n    function separateAdjacentCitations(html) {\n        if (typeof html !== 'string' || !html) {\n            return '';\n        }\n\n        return html.replace(/<\\/sup><sup/g, '</sup><sup>,</sup><sup');\n    }\n\n    // Escapes the few characters that would end an HTML attribute early, so a caption containing a quote\n    // cannot introduce markup of its own.\n    function escapeAttributeValue(text) {\n        return text.replace(/[&<>\"']/g, function (character) {\n            switch (character) {\n                case '&': return '&amp;';\n                case '<': return '&lt;';\n                case '>': return '&gt;';\n                case '\"': return '&quot;';\n                default: return '&#39;';\n            }\n        });\n    }\n\n    /*\n     * The inline marker for a citation: its number, and the source it stands for as a tooltip.\n     *\n     * A bare number tells the reader only that something was cited, and answers which source it was by\n     * making them look away from the sentence and match a number against a list underneath. Carrying the\n     * label means the answer arrives where the question is asked. The list stays as it is, since a tooltip\n     * is not reachable by keyboard and cannot be the only place the source appears.\n     */\n    function citationMarkerHtml(displayIndex, label) {\n        const text = typeof label === 'string' ? label.replace(/\\s+/g, ' ').trim() : '';\n\n        if (!text) {\n            return '<sup>' + displayIndex + '</sup>';\n        }\n\n        return '<sup title=\"' + escapeAttributeValue(text) + '\">' + displayIndex + '</sup>';\n    }\n\n    // Deliberately small: pure functions over strings, for the hosts below and for the tests.\n    return {\n        expandImageMarkers: expandImageMarkers,\n        findChartMarker: findChartMarker,\n        citationIdentity: citationIdentity,\n        splitCombinedCitations: splitCombinedCitations,\n        collapseRepeatedCitations: collapseRepeatedCitations,\n        separateAdjacentCitations: separateAdjacentCitations,\n        citationMarkerHtml: citationMarkerHtml\n    };\n})();\n"]}