feat: phase 5 of AI pipeline hardening — R42 retrieval & telemetry

- Add dependency-free TF-IDF retrieval (src/lib/retrieval.js) with NL+EN
  stopwords and a WeakMap-cached index.
- Rewrite buildKbContext to ship the top-K relevant topics + verbatim-
  mentioned ids only, filter relations to the included set, and append a
  [kb_hash: <8 hex>] suffix so the ephemeral prompt cache busts when the
  graph changes. Returns { context, retrievedTopics, allTopics }.
- Add LOOKUP_TOPIC_TOOL and drive useChat through callLLM directly with a
  multi-hop tool_result loop capped at 3 hops; preserve Anthropic-provided
  tool_use ids through callLLM so the loop can echo correct tool_use_id.
- Truncate R42 history to the last 12 turns and prepend a single
  "(earlier conversation truncated)" assistant message.
- Set R42 chat defaults: temperature 0.3, maxTokens 2048.
- Add pb_migrations/1780500002_created_llm_calls.js (the best-effort
  logger in callLLM was already wired) and a new Admin → Diagnostics
  view showing the last 100 calls with token usage, cache-hit rate, and
  USD cost from a local Anthropic price table.
- Finalize AI_PIPELINE_HARDENING_PLAN.md: mark Phases 1–5 shipped and
  Phase 6 (eval harness) explicitly out of scope.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/lib/__tests__/retrieval.test.js

@@ -0,0 +1,58 @@
+import { describe, expect, it } from 'vitest';
+import { buildIndex, retrieveTopK, tokenize } from '../retrieval';
+
+const sampleTopics = [
+  { id: 'software-engineer', label: 'Software Engineer', description: 'Bouwt en onderhoudt applicaties; werkt in agile teams.' },
+  { id: 'onboarding-buddy', label: 'Onboarding Buddy', description: 'Begeleidt nieuwe medewerkers in hun eerste weken.' },
+  { id: 'kennisbeheer', label: 'Kennisbeheer', description: 'Het proces van het vastleggen en ontsluiten van organisatiekennis.' },
+  { id: 'wekelijkse-sessie', label: 'Wekelijkse Leersessie', description: 'Microlearning sessie waarin medewerkers wekelijks leren via AI-gegenereerde quizzen.' },
+];
+
+describe('tokenize', () => {
+  it('lowercases and splits on non-alphanumeric', () => {
+    expect(tokenize('Hello, World!')).toEqual(['hello', 'world']);
+  });
+
+  it('drops stopwords and short tokens', () => {
+    expect(tokenize('de software engineer is hier')).toEqual(['software', 'engineer', 'hier']);
+  });
+
+  it('keeps hyphenated identifiers', () => {
+    expect(tokenize('software-engineer onboarding-buddy')).toEqual(['software-engineer', 'onboarding-buddy']);
+  });
+});
+
+describe('buildIndex / retrieveTopK', () => {
+  it('returns empty for empty topics', () => {
+    const idx = buildIndex([]);
+    expect(retrieveTopK(idx, 'anything')).toEqual([]);
+  });
+
+  it('returns empty for empty query', () => {
+    const idx = buildIndex(sampleTopics);
+    expect(retrieveTopK(idx, '')).toEqual([]);
+  });
+
+  it('ranks the most relevant topic first', () => {
+    const idx = buildIndex(sampleTopics);
+    const hits = retrieveTopK(idx, 'wat doet een onboarding buddy?', 2);
+    expect(hits[0].id).toBe('onboarding-buddy');
+  });
+
+  it('matches on description when label does not contain query terms', () => {
+    const idx = buildIndex(sampleTopics);
+    const hits = retrieveTopK(idx, 'microlearning quizzen', 3);
+    expect(hits.map(h => h.id)).toContain('wekelijkse-sessie');
+  });
+
+  it('returns no hits when no terms match', () => {
+    const idx = buildIndex(sampleTopics);
+    expect(retrieveTopK(idx, 'kwantumfysica raketten')).toEqual([]);
+  });
+
+  it('caches the index per topics array reference', () => {
+    const idx1 = buildIndex(sampleTopics);
+    const idx2 = buildIndex(sampleTopics);
+    expect(idx1).toBe(idx2);
+  });
+});

src/lib/db.js

@@ -327,6 +327,17 @@ export async function bulkSetCurriculum(year, weeks) {
   );
 }
 
+// ── LLM Call Telemetry ───────────────────────────────────────────────────────
+
+export async function getRecentLlmCalls(limit = 100) {
+  try {
+    const r = await pb.collection('llm_calls').getList(1, limit, { sort: '-created' });
+    return r.items;
+  } catch {
+    return [];
+  }
+}
+
 // ── Handbook Sync State ───────────────────────────────────────────────────────
 
 export async function getHandbookSyncStates() {

src/lib/llm.js

@@ -236,7 +236,7 @@ function extractToolUses(content) {
   if (!Array.isArray(content)) return [];
   return content
     .filter((b) => b?.type === 'tool_use')
-    .map((b) => ({ name: b.name, input: b.input }));
+    .map((b) => ({ id: b.id, name: b.name, input: b.input }));
 }
 
 function extractText(content) {

src/lib/retrieval.js

@@ -0,0 +1,95 @@
+/**
+ * Lightweight, dependency-free TF-IDF retrieval over the knowledge graph.
+ *
+ * `buildIndex(topics)` tokenises the `label + description` of each topic and
+ * computes document-frequency stats so queries can be scored with TF-IDF in
+ * `retrieveTopK`. The index is cached against the `topics` array reference,
+ * so repeated calls with the same array don't rebuild.
+ *
+ * Tokeniser: lowercase, split on `[^a-zA-Z0-9-]`, drop short tokens and a
+ * small Dutch/English stopword list.
+ */
+
+const STOPWORDS = new Set([
+  // English
+  'a', 'an', 'and', 'are', 'as', 'at', 'be', 'by', 'for', 'from', 'has', 'have',
+  'how', 'i', 'in', 'is', 'it', 'its', 'of', 'on', 'or', 'than', 'that', 'the',
+  'this', 'to', 'was', 'were', 'what', 'when', 'where', 'which', 'who', 'why',
+  'with', 'you', 'your', 'do', 'does', 'did',
+  // Dutch
+  'de', 'het', 'een', 'en', 'of', 'in', 'op', 'aan', 'bij', 'voor', 'naar',
+  'met', 'uit', 'om', 'door', 'over', 'tegen', 'ook', 'er', 'is', 'zijn',
+  'was', 'waren', 'wat', 'wie', 'hoe', 'waar', 'wanneer', 'welke', 'die',
+  'dat', 'deze', 'dit', 'ik', 'jij', 'hij', 'zij', 'we', 'wij', 'jullie',
+  'als', 'dan', 'maar', 'want', 'omdat', 'niet', 'wel', 'heeft', 'hebben',
+  'word', 'wordt', 'worden', 'kan', 'kunnen', 'mag', 'moet', 'moeten',
+  'zal', 'zou', 'zouden', 'al', 'ook', 'nog', 'naar',
+]);
+
+export function tokenize(text) {
+  if (!text) return [];
+  return String(text)
+    .toLowerCase()
+    .split(/[^a-z0-9-]+/i)
+    .filter(t => t.length >= 2 && !STOPWORDS.has(t));
+}
+
+const indexCache = new WeakMap();
+
+export function buildIndex(topics) {
+  if (!Array.isArray(topics) || topics.length === 0) {
+    return { topics: [], docFreq: new Map(), termsByDoc: [], N: 0 };
+  }
+  const cached = indexCache.get(topics);
+  if (cached) return cached;
+
+  const termsByDoc = topics.map(t => {
+    const text = `${t.label || ''} ${t.description || ''}`;
+    const tokens = tokenize(text);
+    const tf = new Map();
+    for (const tk of tokens) tf.set(tk, (tf.get(tk) || 0) + 1);
+    return tf;
+  });
+
+  const docFreq = new Map();
+  for (const tf of termsByDoc) {
+    for (const term of tf.keys()) {
+      docFreq.set(term, (docFreq.get(term) || 0) + 1);
+    }
+  }
+
+  const index = { topics, docFreq, termsByDoc, N: topics.length };
+  indexCache.set(topics, index);
+  return index;
+}
+
+export function retrieveTopK(index, query, k = 10) {
+  if (!index || !index.N || !query) return [];
+  const qTokens = tokenize(query);
+  if (qTokens.length === 0) return [];
+
+  const idf = (term) => {
+    const df = index.docFreq.get(term) || 0;
+    if (df === 0) return 0;
+    return Math.log((index.N + 1) / (df + 1)) + 1;
+  };
+
+  const scores = new Array(index.N);
+  for (let i = 0; i < index.N; i++) {
+    const tf = index.termsByDoc[i];
+    let s = 0;
+    for (const t of qTokens) {
+      const f = tf.get(t);
+      if (!f) continue;
+      s += (1 + Math.log(f)) * idf(t);
+    }
+    scores[i] = s;
+  }
+
+  const ranked = [];
+  for (let i = 0; i < index.N; i++) {
+    if (scores[i] > 0) ranked.push({ i, s: scores[i] });
+  }
+  ranked.sort((a, b) => b.s - a.s);
+  return ranked.slice(0, k).map(r => index.topics[r.i]);
+}