@ctcstack/ctcdocs-core 0.17.0 → 0.18.0

package/dist/corpus-structure.d.ts

@@ -23,6 +23,8 @@ export interface CorpusDocument {
     readonly shortId?: string;
     readonly title?: string;
     readonly modified?: string;
+    /** The Google Doc or PDF in Drive, as the page links it. */
+    readonly source?: string;
 }
 export interface CorpusStructure {
     readonly rootFolderId: string | null;

package/dist/corpus-structure.js

@@ -65,6 +65,7 @@ export function parseCorpusStructure(manifest) {
         const shortId = optional('shortId');
         const title = optional('displayTitle');
         const modified = optional('googleModifiedTime');
+        const source = optional('sourceUrl');
         documents.set(id, {
             id,
             parentId: field(raw, 'googleParentId', path, true),
@@ -72,6 +73,7 @@ export function parseCorpusStructure(manifest) {
             ...(shortId ? { shortId } : {}),
             ...(title ? { title } : {}),
             ...(modified ? { modified } : {}),
+            ...(source ? { source } : {}),
             ...(published === '*' ||
                 (Array.isArray(published) &&
                     published.every((group) => typeof group === 'string'))

package/dist/published-markdown.d.ts

@@ -30,4 +30,15 @@ export declare function publishedFileUrl(googleFileId: string, file: string): st
  * titles, which come from Drive names and may contain any of them.
  */
 export declare function inlineMarkdown(value: string): string;
+/**
+ * The page's Markdown without its front matter: its title as a heading and
+ * its body. It is what the MCP server stores, indexes and returns from
+ * `fetch` (ADR-042), so it is also what the sync measures (ADR-043).
+ */
+export declare function publishedMarkdownBody(input: Pick<PublishedMarkdownInput, 'title' | 'ownershipHeader' | 'body' | 'stableSlugs' | 'permanentLinks'>): string;
+/**
+ * The page's Markdown version (ADR-010): front matter, then
+ * `publishedMarkdownBody`. Each front matter value is one JSON string, so no
+ * value spans a line and the first `---` line after the opening one closes it.
+ */
 export declare function serializePublishedMarkdown(input: PublishedMarkdownInput): string;

package/dist/published-markdown.js

@@ -98,8 +98,25 @@ export function inlineMarkdown(value) {
         .trim()
         .replace(/([\\`*_[\]<>#])/gu, '\\$1');
 }
-export function serializePublishedMarkdown(input) {
+/**
+ * The page's Markdown without its front matter: its title as a heading and
+ * its body. It is what the MCP server stores, indexes and returns from
+ * `fetch` (ADR-042), so it is also what the sync measures (ADR-043).
+ */
+export function publishedMarkdownBody(input) {
     const body = rewriteInternalLinks(cleanBody(input.body, input.ownershipHeader), input.stableSlugs, input.permanentLinks);
+    return [
+        `# ${inlineMarkdown(input.title)}`,
+        '',
+        ...(body ? [body, ''] : []),
+    ].join('\n');
+}
+/**
+ * The page's Markdown version (ADR-010): front matter, then
+ * `publishedMarkdownBody`. Each front matter value is one JSON string, so no
+ * value spans a line and the first `---` line after the opening one closes it.
+ */
+export function serializePublishedMarkdown(input) {
     return [
         '---',
         `title: ${yamlString(input.title)}`,
@@ -110,8 +127,6 @@ export function serializePublishedMarkdown(input) {
         `content_hash: ${yamlString(input.contentHash)}`,
         '---',
         '',
-        `# ${inlineMarkdown(input.title)}`,
-        '',
-        ...(body ? [body, ''] : []),
+        publishedMarkdownBody(input),
     ].join('\n');
 }

package/dist/site-configuration.d.ts

@@ -152,13 +152,16 @@ export interface McpSearchConfiguration {
     readonly chunks: number;
     /**
      * The vector similarity below which AI Search drops a chunk, 0 to 1. It
-     * applies to vector similarity alone, so a keyword match far from the
-     * question in meaning is dropped too when it is set high.
+     * drops only chunks vector search alone found, and AI Search ignores it
+     * while reranking is on (ADR-042), so it acts only with reranking off.
      */
     readonly vectorThreshold: number;
     /** Whether a keyword match needs every word of the query, or any. */
     readonly keywordMatch: 'and' | 'or';
-    /** Neighbouring chunks AI Search adds on each side of a match: 0 to 3. */
+    /**
+     * Neighbouring chunks AI Search joins to each side of a match: 0 to 3. A
+     * chunk of AI Search's default size is already longer than a passage.
+     */
     readonly contextChunks: number;
     readonly reranking: {
         readonly enabled: boolean;

package/dist/site-configuration.js

@@ -21,7 +21,7 @@ export const MCP_DEFAULTS = Object.freeze({
         chunks: 50,
         vectorThreshold: 0.2,
         keywordMatch: 'or',
-        contextChunks: 1,
+        contextChunks: 0,
         reranking: Object.freeze({
             enabled: true,
             model: '@cf/baai/bge-reranker-base',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ctcstack/ctcdocs-core",
-  "version": "0.17.0",
+  "version": "0.18.0",
   "description": "Configuration schema, project layout, generated-path allowlist and the published Markdown form shared by the CTCDocs platform packages.",
   "keywords": [
     "ctcdocs",

package/src/corpus-structure.ts

@@ -37,6 +37,8 @@ export interface CorpusDocument {
   readonly shortId?: string;
   readonly title?: string;
   readonly modified?: string;
+  /** The Google Doc or PDF in Drive, as the page links it. */
+  readonly source?: string;
 }
 
 export interface CorpusStructure {
@@ -114,6 +116,7 @@ export function parseCorpusStructure(manifest: unknown): CorpusStructure {
     const shortId = optional('shortId');
     const title = optional('displayTitle');
     const modified = optional('googleModifiedTime');
+    const source = optional('sourceUrl');
     documents.set(id, {
       id,
       parentId: field(raw, 'googleParentId', path, true),
@@ -121,6 +124,7 @@ export function parseCorpusStructure(manifest: unknown): CorpusStructure {
       ...(shortId ? { shortId } : {}),
       ...(title ? { title } : {}),
       ...(modified ? { modified } : {}),
+      ...(source ? { source } : {}),
       ...(published === '*' ||
       (Array.isArray(published) &&
         published.every((group) => typeof group === 'string'))

package/src/published-markdown.ts

@@ -153,14 +153,37 @@ export function inlineMarkdown(value: string): string {
     .replace(/([\\`*_[\]<>#])/gu, '\\$1');
 }
 
-export function serializePublishedMarkdown(
-  input: PublishedMarkdownInput,
+/**
+ * The page's Markdown without its front matter: its title as a heading and
+ * its body. It is what the MCP server stores, indexes and returns from
+ * `fetch` (ADR-042), so it is also what the sync measures (ADR-043).
+ */
+export function publishedMarkdownBody(
+  input: Pick<
+    PublishedMarkdownInput,
+    'title' | 'ownershipHeader' | 'body' | 'stableSlugs' | 'permanentLinks'
+  >,
 ): string {
   const body = rewriteInternalLinks(
     cleanBody(input.body, input.ownershipHeader),
     input.stableSlugs,
     input.permanentLinks,
   );
+  return [
+    `# ${inlineMarkdown(input.title)}`,
+    '',
+    ...(body ? [body, ''] : []),
+  ].join('\n');
+}
+
+/**
+ * The page's Markdown version (ADR-010): front matter, then
+ * `publishedMarkdownBody`. Each front matter value is one JSON string, so no
+ * value spans a line and the first `---` line after the opening one closes it.
+ */
+export function serializePublishedMarkdown(
+  input: PublishedMarkdownInput,
+): string {
   return [
     '---',
     `title: ${yamlString(input.title)}`,
@@ -171,8 +194,6 @@ export function serializePublishedMarkdown(
     `content_hash: ${yamlString(input.contentHash)}`,
     '---',
     '',
-    `# ${inlineMarkdown(input.title)}`,
-    '',
-    ...(body ? [body, ''] : []),
+    publishedMarkdownBody(input),
   ].join('\n');
 }

package/src/site-configuration.ts

@@ -183,13 +183,16 @@ export interface McpSearchConfiguration {
   readonly chunks: number;
   /**
    * The vector similarity below which AI Search drops a chunk, 0 to 1. It
-   * applies to vector similarity alone, so a keyword match far from the
-   * question in meaning is dropped too when it is set high.
+   * drops only chunks vector search alone found, and AI Search ignores it
+   * while reranking is on (ADR-042), so it acts only with reranking off.
    */
   readonly vectorThreshold: number;
   /** Whether a keyword match needs every word of the query, or any. */
   readonly keywordMatch: 'and' | 'or';
-  /** Neighbouring chunks AI Search adds on each side of a match: 0 to 3. */
+  /**
+   * Neighbouring chunks AI Search joins to each side of a match: 0 to 3. A
+   * chunk of AI Search's default size is already longer than a passage.
+   */
   readonly contextChunks: number;
   readonly reranking: {
     readonly enabled: boolean;
@@ -224,7 +227,7 @@ export const MCP_DEFAULTS: Omit<McpConfiguration, 'enabled'> = Object.freeze({
     chunks: 50,
     vectorThreshold: 0.2,
     keywordMatch: 'or',
-    contextChunks: 1,
+    contextChunks: 0,
     reranking: Object.freeze({
       enabled: true,
       model: '@cf/baai/bge-reranker-base',