Merge pull request 'feat: phase 2 of AI pipeline hardening — tool-based structured outputs + prompt caching' (#3) from feat/ai-pipeline-hardening-phase-2 into main

Reviewed-on: #3

src/components/admin/KnowledgeGraph.jsx

@@ -2,7 +2,8 @@ import { useCallback, useEffect, useRef, useState } from 'react';
 import * as d3 from 'd3';
 import { Trash2, Edit2, Save, X, RefreshCw, AlertCircle, Plus, Link as LinkIcon } from 'lucide-react';
 import * as db from '../../lib/db';
-import { anthropicApi } from '../../lib/api';
+import { callLLM } from '../../lib/llm';
+import { EMIT_GRAPH_ACTIONS_TOOL } from '../../lib/llmTools';
 import { analyzeHandbookDelta } from '../../lib/extractionPipeline';
 import { getRepoFolder, getFileContent } from '../../lib/githubService';
 import Button from '../ui/Button';
@@ -304,18 +305,18 @@ const KnowledgeGraph = () => {
       const currentTopics = await db.getTopics();
       const currentRelations = await db.getRelations();
 
-      const systemPrompt = `You are a strict Data Quality AI maintaining a Knowledge Graph.
+      const systemPrompt = `You are a strict Data Quality AI maintaining a Knowledge Graph for Respellion.
-Your goal is to evaluate the provided topics and relations, identify duplicates to merge, useless nodes to delete, and new logical relations to add.
+Evaluate the provided topics and relations and emit the actions to take via the emit_graph_actions tool.
 
 Rules:
-1. Identify topics that mean exactly the same thing. Choose one to keep, and one to delete.
+1. Identify topics that mean exactly the same thing. Choose one to keep, one to delete (merges).
-2. Identify topics that are too vague, irrelevant, or malformed to delete.
+2. Identify topics that are too vague, irrelevant, or malformed (deletions).
-3. Identify missing logical relations (depends_on, part_of, related_to) if two topics are conceptually linked but missing a relation.
+3. Identify missing logical relations (depends_on, part_of, related_to, executed_by) between conceptually linked topics (newRelations).
-4. Evaluate the learning_relevance of each topic. If a topic is purely operational/mundane (like a printer guide), mark it as "exclude". If it's low priority, mark "peripheral".
+4. Evaluate learning_relevance. Mark purely operational topics (printer guides, etc.) as "exclude"; low-priority as "peripheral" (relevanceUpdates).
-5. Return ONLY a valid JSON object describing the ACTIONS to take. Do not return the entire graph. Do not wrap in markdown blocks.`;
+
+Do not return the entire graph — only the actions to take.`;
 
       // Send a compact representation to minimize token usage and avoid rate limits.
-      // The AI only needs id, label, type, and relevance to identify duplicates/merges and adjust relevance.
       const compactTopics = currentTopics.map(({ id, label, type, learning_relevance }) => ({ id, label, type, learning_relevance }));
       const compactRelations = currentRelations.map(r => ({
         source: r.source?.id || r.source,
@@ -324,21 +325,20 @@ Rules:
       }));
 
       const userPrompt = `Here is the current knowledge graph:
-${JSON.stringify({ topics: compactTopics, relations: compactRelations })}
+${JSON.stringify({ topics: compactTopics, relations: compactRelations })}`;
 
-Analyze this graph and return ONLY the optimized JSON object with this EXACT structure:
+      const llmResult = await callLLM({
-{
+        task: 'graph.analyze',
-  "merges": [ { "keepId": "id_to_keep", "deleteId": "id_to_delete" } ],
+        tier: 'reasoning',
-  "deletions": [ "id_to_delete_completely" ],
+        system: [{ type: 'text', text: systemPrompt, cache_control: { type: 'ephemeral' } }],
-  "newRelations": [ { "source": "id1", "target": "id2", "type": "depends_on" } ],
+        user: userPrompt,
-  "relevanceUpdates": [ { "id": "topic_id", "learning_relevance": "exclude" } ]
+        tools: [EMIT_GRAPH_ACTIONS_TOOL],
-}`;
+        toolChoice: { type: 'tool', name: EMIT_GRAPH_ACTIONS_TOOL.name },
+        maxTokens: 4096,
+      });
 
-      const responseText = await anthropicApi.generateContent(systemPrompt, userPrompt);
+      const actions = llmResult.toolUses[0]?.input;
-      const jsonMatch = responseText.match(/\{[\s\S]*\}/);
+      if (!actions) throw new Error('Graph analysis did not emit a tool result.');
-      if (!jsonMatch) throw new Error('AI returned invalid format.');
-
-      const actions = JSON.parse(jsonMatch[0]);
 
       let updatedTopics = [...currentTopics];
       let updatedRelations = [...currentRelations];

src/components/chat/prompts.js

@@ -23,10 +23,9 @@ export const STRINGS = {
   openAria: 'Open R42 chatbot',
 };
 
-export function buildSystemPrompt({ userName, isAdmin, kbContext }) {
+const STABLE_PREAMBLE = [
-  return [
   `Je bent R42, de chatbot-avatar van Respellion — een leerplatform voor microlearning, quizzen en kennisontwikkeling.`,
-    `Antwoord altijd in het Nederlands, kort en zakelijk-vriendelijk. Spreek de gebruiker aan met hun voornaam wanneer dat natuurlijk voelt (${userName}).`,
+  `Antwoord altijd in het Nederlands, kort en zakelijk-vriendelijk. Spreek de gebruiker aan met hun voornaam wanneer dat natuurlijk voelt.`,
   ``,
   `JE TAKEN:`,
   `1. Leg onderwerpen uit die in de kennisbasis staan.`,
@@ -34,9 +33,7 @@ export function buildSystemPrompt({ userName, isAdmin, kbContext }) {
   `3. Verwijs bij twijfel terug naar het bronmateriaal of zeg eerlijk dat je het niet weet.`,
   ``,
   `JE KENNIS:`,
-    `Je kennis is beperkt tot de onderstaande Respellion-kennisgraaf. Als een vraag duidelijk buiten dit bereik valt, zeg dat dan eerlijk en stel voor dat de gebruiker de bron toevoegt via Admin → Sources.`,
+  `Je kennis is beperkt tot de Respellion-kennisgraaf die hieronder volgt. Als een vraag duidelijk buiten dit bereik valt, zeg dat dan eerlijk en stel voor dat de gebruiker de bron toevoegt via Admin → Sources.`,
-    ``,
-    kbContext,
   ``,
   `KENNISGRAAF VERFIJNEN:`,
   `Wanneer de gebruiker iets noemt dat duidelijk een nieuw topic, nieuwe relatie, proces of rol is — en dat nog niet in de kennisgraaf staat — gebruik dan de tool "propose_graph_delta" om een voorstel te maken. Verzin niets: stel alleen iets voor als de gebruiker het concreet noemt. Stel maximaal 3 topics en 5 relaties per beurt voor.`,
@@ -45,10 +42,30 @@ export function buildSystemPrompt({ userName, isAdmin, kbContext }) {
   `- Houd antwoorden onder de 4 zinnen tenzij de gebruiker om uitleg vraagt.`,
   `- Geen markdown-headers; gewone Nederlandse tekst.`,
   `- Bij onzekerheid: "Ik weet het niet zeker — controleer dit in het handboek."`,
-    isAdmin
-      ? `\nDe gebruiker is beheerder; voorstellen die de tool genereert worden direct toegepast.`
-      : `\nDe gebruiker is geen beheerder; voorstellen worden in een goedkeuringswachtrij gezet.`,
 ].join('\n');
+
+/**
+ * Build the R42 system prompt as three cacheable blocks:
+ *   1. stable preamble (role, tasks, style)         — cached
+ *   2. KB context (current topics + relations)      — cached (hash-bust comes in Phase 5)
+ *   3. per-turn tail (user name + admin status)     — NOT cached
+ *
+ * Returning an array lets `callLLM` pass it through unchanged so the
+ * Anthropic API caches each block with the 5-minute ephemeral TTL.
+ */
+export function buildSystemPrompt({ userName, isAdmin, kbContext }) {
+  const tail = [
+    `De gebruiker heet ${userName}.`,
+    isAdmin
+      ? `De gebruiker is beheerder; voorstellen die de tool genereert worden direct toegepast.`
+      : `De gebruiker is geen beheerder; voorstellen worden in een goedkeuringswachtrij gezet.`,
+  ].join('\n');
+
+  return [
+    { type: 'text', text: STABLE_PREAMBLE, cache_control: { type: 'ephemeral' } },
+    { type: 'text', text: kbContext, cache_control: { type: 'ephemeral' } },
+    { type: 'text', text: tail },
+  ];
 }
 
 export const PROPOSE_GRAPH_DELTA_TOOL = {

src/lib/__tests__/articlePatches.test.js

@@ -0,0 +1,104 @@
+import { describe, expect, it } from 'vitest';
+import { applyArticlePatches, applyAndValidate } from '../articlePatches';
+
+const article = () => ({
+  title: 'Onboarding',
+  intro: 'Old intro.',
+  sections: [
+    { heading: 'Day one', body: 'First day body, three sentences long. Welcome. Read the handbook.' },
+    { heading: 'Day two', body: 'Second day body. Three sentences. Meet your team.' },
+  ],
+  keyTakeaways: ['Show up', 'Ask questions'],
+});
+
+describe('applyArticlePatches', () => {
+  it('does not mutate the input article', () => {
+    const original = article();
+    const snapshot = JSON.parse(JSON.stringify(original));
+    applyArticlePatches(original, [
+      { name: 'set_intro', input: { intro: 'New intro.' } },
+    ]);
+    expect(original).toEqual(snapshot);
+  });
+
+  it('set_intro replaces the intro', () => {
+    const result = applyArticlePatches(article(), [
+      { name: 'set_intro', input: { intro: 'Punchier intro.' } },
+    ]);
+    expect(result.intro).toBe('Punchier intro.');
+  });
+
+  it('set_section replaces the matching section body (case-insensitive)', () => {
+    const result = applyArticlePatches(article(), [
+      { name: 'set_section', input: { heading: 'DAY ONE', body: 'Rewritten body. With several sentences. Indeed.' } },
+    ]);
+    expect(result.sections[0].body).toMatch(/Rewritten body/);
+    expect(result.sections[1].body).toMatch(/Second day body/);
+  });
+
+  it('add_section position=start prepends a new section', () => {
+    const result = applyArticlePatches(article(), [
+      { name: 'add_section', input: { heading: 'Before', body: 'New intro section. Three sentences. Indeed.', position: 'start' } },
+    ]);
+    expect(result.sections[0].heading).toBe('Before');
+    expect(result.sections).toHaveLength(3);
+  });
+
+  it('add_section position=end appends a new section', () => {
+    const result = applyArticlePatches(article(), [
+      { name: 'add_section', input: { heading: 'After', body: 'Closing section. Three sentences. Indeed.', position: 'end' } },
+    ]);
+    expect(result.sections[2].heading).toBe('After');
+  });
+
+  it('remove_section drops the matching section', () => {
+    const result = applyArticlePatches(article(), [
+      { name: 'remove_section', input: { heading: 'Day one' } },
+    ]);
+    expect(result.sections).toHaveLength(1);
+    expect(result.sections[0].heading).toBe('Day two');
+  });
+
+  it('replace_takeaways swaps the key takeaways', () => {
+    const result = applyArticlePatches(article(), [
+      { name: 'replace_takeaways', input: { items: ['First', 'Second', 'Third'] } },
+    ]);
+    expect(result.keyTakeaways).toEqual(['First', 'Second', 'Third']);
+  });
+
+  it('applies multiple patches in order', () => {
+    const result = applyArticlePatches(article(), [
+      { name: 'set_intro', input: { intro: 'Brand new intro.' } },
+      { name: 'remove_section', input: { heading: 'Day one' } },
+      { name: 'add_section', input: { heading: 'New', body: 'Body of the new section. Three sentences. Yes.', position: 'end' } },
+    ]);
+    expect(result.intro).toBe('Brand new intro.');
+    expect(result.sections.map(s => s.heading)).toEqual(['Day two', 'New']);
+  });
+
+  it('falls back to appending when set_section cannot find a matching heading', () => {
+    const result = applyArticlePatches(article(), [
+      { name: 'set_section', input: { heading: 'Nonexistent', body: 'New body, with three sentences. Yes indeed. Foo.' } },
+    ]);
+    expect(result.sections).toHaveLength(3);
+    expect(result.sections[2].heading).toBe('Nonexistent');
+  });
+});
+
+describe('applyAndValidate', () => {
+  it('returns the patched article when valid', () => {
+    const patched = applyAndValidate(article(), [
+      { name: 'set_intro', input: { intro: 'Tighter intro.' } },
+    ]);
+    expect(patched.intro).toBe('Tighter intro.');
+  });
+
+  it('throws when patches strip the article to invalid', () => {
+    expect(() =>
+      applyAndValidate(article(), [
+        { name: 'remove_section', input: { heading: 'Day one' } },
+        { name: 'remove_section', input: { heading: 'Day two' } },
+      ]),
+    ).toThrow(/invalid article/i);
+  });
+});

src/lib/__tests__/llmTools.test.js

@@ -0,0 +1,44 @@
+import { describe, expect, it } from 'vitest';
+import {
+  EMIT_KNOWLEDGE_GRAPH_TOOL,
+  EMIT_HANDBOOK_DELTA_TOOL,
+  EMIT_LEARNING_ARTICLE_TOOL,
+  EMIT_LEARNING_SLIDES_TOOL,
+  EMIT_LEARNING_INFOGRAPHIC_TOOL,
+  EMIT_LEARNING_ALL_TOOL,
+  EMIT_CUSTOM_TOPIC_TOOL,
+  EMIT_QUIZ_QUESTIONS_TOOL,
+  EMIT_GRAPH_ACTIONS_TOOL,
+  ARTICLE_PATCH_TOOLS,
+} from '../llmTools';
+import { toolSchemaRegistry } from '../llmSchemas';
+
+const allTools = [
+  EMIT_KNOWLEDGE_GRAPH_TOOL,
+  EMIT_HANDBOOK_DELTA_TOOL,
+  EMIT_LEARNING_ARTICLE_TOOL,
+  EMIT_LEARNING_SLIDES_TOOL,
+  EMIT_LEARNING_INFOGRAPHIC_TOOL,
+  EMIT_LEARNING_ALL_TOOL,
+  EMIT_CUSTOM_TOPIC_TOOL,
+  EMIT_QUIZ_QUESTIONS_TOOL,
+  EMIT_GRAPH_ACTIONS_TOOL,
+  ...ARTICLE_PATCH_TOOLS,
+];
+
+describe('llmTools', () => {
+  it('every tool has a name, description, and object input_schema', () => {
+    for (const t of allTools) {
+      expect(typeof t.name).toBe('string');
+      expect(t.name.length).toBeGreaterThan(0);
+      expect(typeof t.description).toBe('string');
+      expect(t.input_schema).toMatchObject({ type: 'object' });
+    }
+  });
+
+  it('every tool has a matching Zod validator in toolSchemaRegistry', () => {
+    for (const t of allTools) {
+      expect(toolSchemaRegistry[t.name]).toBeTruthy();
+    }
+  });
+});

src/lib/articlePatches.js

@@ -0,0 +1,80 @@
+/**
+ * Apply a sequence of patch operations (the tool_use calls returned by
+ * `refineLearningContent`) to an article object, in order. The returned
+ * article is a fresh object — the input is not mutated.
+ *
+ * Recognised tool names mirror `llmTools.js`:
+ *   set_intro, set_section, add_section, remove_section, replace_takeaways.
+ *
+ * Unknown tool names are ignored on purpose; the caller validates the
+ * result against `learningArticleSchema` and rejects the whole turn if
+ * the patches produced an invalid article.
+ */
+
+import { learningArticleSchema } from './llmSchemas';
+
+function matchesHeading(section, heading) {
+  return (section.heading ?? '').trim().toLowerCase() === heading.trim().toLowerCase();
+}
+
+function cloneArticle(article) {
+  return {
+    ...article,
+    sections: article.sections.map((s) => ({ ...s })),
+    keyTakeaways: [...article.keyTakeaways],
+  };
+}
+
+export function applyArticlePatches(article, toolUses) {
+  let next = cloneArticle(article);
+  for (const tu of toolUses) {
+    switch (tu.name) {
+      case 'set_intro':
+        next = { ...next, intro: tu.input.intro };
+        break;
+      case 'set_section': {
+        const idx = next.sections.findIndex((s) => matchesHeading(s, tu.input.heading));
+        if (idx === -1) {
+          // No matching section — fall back to appending so the model's
+          // intent (provide that body) is preserved rather than lost.
+          next.sections = [...next.sections, { heading: tu.input.heading, body: tu.input.body }];
+        } else {
+          next.sections = next.sections.map((s, i) => (i === idx ? { ...s, body: tu.input.body } : s));
+        }
+        break;
+      }
+      case 'add_section': {
+        const newSection = { heading: tu.input.heading, body: tu.input.body };
+        next.sections = tu.input.position === 'start'
+          ? [newSection, ...next.sections]
+          : [...next.sections, newSection];
+        break;
+      }
+      case 'remove_section':
+        next.sections = next.sections.filter((s) => !matchesHeading(s, tu.input.heading));
+        break;
+      case 'replace_takeaways':
+        next = { ...next, keyTakeaways: [...tu.input.items] };
+        break;
+      default:
+        // Unknown patch op — ignore.
+        break;
+    }
+  }
+  return next;
+}
+
+/**
+ * Apply the patches and re-validate against the article schema. Throws
+ * a clear error if the result is invalid.
+ */
+export function applyAndValidate(article, toolUses) {
+  const updated = applyArticlePatches(article, toolUses);
+  const parsed = learningArticleSchema.safeParse({ article: updated });
+  if (!parsed.success) {
+    const err = new Error(`Refinement produced an invalid article: ${parsed.error.message}`);
+    err.cause = parsed.error;
+    throw err;
+  }
+  return parsed.data.article;
+}

src/lib/extractionPipeline.js

@@ -1,85 +1,62 @@
-import { anthropicApi } from './api';
 import * as db from './db';
+import { callLLM } from './llm';
+import { EMIT_KNOWLEDGE_GRAPH_TOOL, EMIT_HANDBOOK_DELTA_TOOL } from './llmTools';
+import { normalizeHandbookResult } from './llmSchemas';
 
-const SYSTEM_PROMPT = `You are an AI knowledge extractor for Respellion, an IT company built on radical transparency.
+const EXTRACTION_SYSTEM_PROMPT = `You are an AI knowledge extractor for Respellion, an IT company built on radical transparency.
-You receive a source text. Your task is to extract all core concepts, roles, and processes from the text, and return them as a structured JSON Knowledge Graph.
+You receive a source text. Extract every distinct concept, role, and process from it and emit them through the emit_knowledge_graph tool.
-Facts should be integrated into the descriptions of the other labels and NOT be extracted as unique topics.
 
 CRITICAL INSTRUCTIONS FOR COMPLETENESS:
-- You must extract EVERY SINGLE distinct role, process, and concept described or mentioned in the source text. 
+- Extract EVERY SINGLE distinct role, process, and concept described or mentioned in the source text.
-- DO NOT summarize, skip, truncate, or omit any items. 
+- DO NOT summarise, skip, truncate, or omit any items.
-- If the document contains 29 roles, your JSON topics array must contain exactly 29 role topics. 
+- If the document contains 29 roles, the topics array must contain exactly 29 role topics.
-- Completeness is of paramount importance. Failing to extract all topics will result in loss of critical company knowledge.
+- Completeness is paramount. Failing to extract all topics loses critical company knowledge.
-- Keep descriptions concise (max 3 sentences) to ensure you have enough output tokens to list everything.
+- Facts should be integrated into the descriptions of other topics — never extracted as standalone topics.
+- Keep descriptions concise (max 3 sentences) so the response fits.
 
-You MUST assign a learning_relevance to each topic:
+Topic IDs are lowercase kebab-case slugs specific to the topic (e.g. "software-engineer", "data-quality-review"). Do not use generic IDs like "role-1" or "concept-2".
-- "core": Fundamental company knowledge.
-- "standard": Normal learning topics.
-- "peripheral": Good to know, but low priority.
-- "exclude": Pure operational reference material (e.g., printer guides, wifi passwords) that should NEVER be tested.
 
-ALWAYS return a valid JSON object in the following format:
+Assign a learning_relevance to every topic:
-{
+- "core": fundamental company knowledge.
-  "topics": [
+- "standard": normal learning topics.
-    {
+- "peripheral": good to know, low priority.
-      "id": "a-unique-lowercase-kebab-case-slug-specific-to-this-topic (e.g., 'software-engineer' or 'data-quality-review'). DO NOT use generic IDs like 'role-1' or 'concept-2'.",
+- "exclude": pure operational reference (printer guides, wifi passwords) that should never be tested.
-      "label": "Topic title",
-      "type": "concept | role | process",
-      "description": "A concise, clear explanation of max 3 sentences.",
-      "learning_relevance": "core | standard | peripheral | exclude"
-    }
-  ],
-  "relations": [
-    {
-      "source": "topic-id-1",
-      "target": "topic-id-2",
-      "type": "related_to | depends_on | part_of | executed_by"
-    }
-  ]
-}
-Return JSON only. No markdown blocks or other text.`;
 
-const HANDBOOK_SYSTEM_PROMPT = `You are analyzing an update to the Respellion Employee Handbook.
+Relation types: related_to | depends_on | part_of | executed_by.
-Your task is to identify changes and extract structural knowledge.
+`;
 
-CRITICAL INSTRUCTION:
+const HANDBOOK_SYSTEM_PROMPT = `You are analysing an update to the Respellion Employee Handbook. Emit the extracted topics and relations through the emit_handbook_delta tool.
-You must explicitly identify and create relations between Roles, Processes, and Concepts.
-Every Process must have a Role attached (who does it).
-Every Concept must have a relation to a Process or Role.
 
-You MUST assign a learning_relevance to each topic:
+CRITICAL INSTRUCTIONS:
-- "core": Fundamental company knowledge.
+- Every process must have a role attached (the role that executes it).
-- "standard": Normal learning topics.
+- Every concept must connect to a process or role.
-- "peripheral": Good to know, but low priority.
+- Mark handbook topics with metadata.source = "github_handbook".
-- "exclude": Pure operational reference material (e.g., printer guides, wifi passwords) that should NEVER be tested.
+- Assign learning_relevance using the same scale as extraction: core | standard | peripheral | exclude.
 
-Return a JSON object:
+Relation types: related_to | depends_on | part_of | executed_by. (Legacy "executes" relations are normalised by the client into executed_by with source/target swapped.)
-{
+`;
-  "topics": [
+
-    { "id": "...", "label": "...", "type": "role | process | concept", "description": "...", "learning_relevance": "standard", "metadata": { "source": "github_handbook" } }
+const cachedSystem = (text) => [{ type: 'text', text, cache_control: { type: 'ephemeral' } }];
-  ],
-  "relations": [
-    { "source": "role-id", "target": "process-id", "type": "executes | related_to | depends_on | part_of", "description": "Brief metadata about this specific relation" }
-  ]
-}
-Return JSON only. No markdown blocks or other text.`;
 
 export async function analyzeHandbookDelta(fileContent, filePath) {
-  const responseText = await anthropicApi.generateContent(HANDBOOK_SYSTEM_PROMPT, `Analyze the following handbook file update (${filePath}):\n\n${fileContent}`);
+  const result = await callLLM({
+    task: 'extract.handbook',
+    tier: 'standard',
+    system: cachedSystem(HANDBOOK_SYSTEM_PROMPT),
+    user: `Analyze the following handbook file update (${filePath}):\n\n${fileContent}`,
+    tools: [EMIT_HANDBOOK_DELTA_TOOL],
+    toolChoice: { type: 'tool', name: EMIT_HANDBOOK_DELTA_TOOL.name },
+    maxTokens: 8192,
+  });
 
-  let extractedData;
+  const raw = result.toolUses[0]?.input;
-  try {
+  if (!raw) throw new Error('Handbook extraction did not emit a tool result.');
-    const jsonMatch = responseText.match(/\{[\s\S]*\}/);
+  const extractedData = normalizeHandbookResult(raw);
-    const jsonStr = jsonMatch ? jsonMatch[0] : responseText;
-    extractedData = JSON.parse(jsonStr);
-  } catch (e) {
-    console.error('[Pipeline] AI returned non-JSON response for handbook delta:', responseText?.substring(0, 500));
-    throw new Error(`AI response was not valid JSON. The model responded with: "${responseText?.substring(0, 120)}..."`, { cause: e });
-  }
 
   await mergeKnowledgeGraph(extractedData);
   return { success: true, data: extractedData };
 }
+
 function chunkText(text, maxChunkSize = 4000) {
   const paragraphs = text.split(/\n+/);
   const chunks = [];
@@ -98,7 +75,6 @@ function chunkText(text, maxChunkSize = 4000) {
 }
 
 export async function processSourceText(textContent, sourceName) {
-  // Deduplicate: skip if a source with the same name was already successfully processed
   const existing = await db.getSources();
   const alreadyDone = existing.find(
     s => s.name === sourceName && s.status === 'completed'
@@ -124,21 +100,18 @@ export async function processSourceText(textContent, sourceName) {
       }
 
       console.log(`[Pipeline] Processing chunk ${i + 1}/${chunks.length} (${chunks[i].length} chars)...`);
-      const responseText = await anthropicApi.generateContent(
+      const result = await callLLM({
-        SYSTEM_PROMPT, 
+        task: 'extract.source',
-        `Analyze this part of the document (${i + 1}/${chunks.length}):\n\n${chunks[i]}`
+        tier: 'standard',
-      );
+        system: cachedSystem(EXTRACTION_SYSTEM_PROMPT),
-      console.log(`[Pipeline] Raw AI response for chunk ${i + 1}:`, responseText);
+        user: `Analyze this part of the document (${i + 1}/${chunks.length}):\n\n${chunks[i]}`,
+        tools: [EMIT_KNOWLEDGE_GRAPH_TOOL],
+        toolChoice: { type: 'tool', name: EMIT_KNOWLEDGE_GRAPH_TOOL.name },
+        maxTokens: 8192,
+      });
 
-      let extractedData;
+      const extractedData = result.toolUses[0]?.input;
-      try {
+      if (!extractedData) throw new Error(`Extraction did not emit a tool result for chunk ${i + 1}.`);
-        const jsonMatch = responseText.match(/\{[\s\S]*\}/);
-        const jsonStr = jsonMatch ? jsonMatch[0] : responseText;
-        extractedData = JSON.parse(jsonStr);
-      } catch (e) {
-        console.error(`[Pipeline] AI returned non-JSON response for chunk ${i + 1}:`, responseText?.substring(0, 500));
-        throw new Error(`AI response for chunk ${i + 1} was not valid JSON.`, { cause: e });
-      }
 
       if (extractedData.topics && Array.isArray(extractedData.topics)) {
         allExtractedTopics.push(...extractedData.topics);
@@ -148,7 +121,6 @@ export async function processSourceText(textContent, sourceName) {
       }
     }
 
-    // Merge everything together
     await mergeKnowledgeGraph({ topics: allExtractedTopics, relations: allExtractedRelations });
     await db.updateSourceStatus(sourceId, 'completed');
 
@@ -169,13 +141,11 @@ async function mergeKnowledgeGraph(newData) {
   if (newData.topics && Array.isArray(newData.topics)) {
     for (const t of newData.topics) {
       if (topicsMap.has(t.id)) {
-        // Upsert: merge new data into existing topic
         const existing = topicsMap.get(t.id);
         topicsMap.set(t.id, {
           ...existing,
           ...t,
-          // Keep existing description if new one is empty, or combine them if needed. Here we prefer the new one.
+          description: t.description || existing.description,
-          description: t.description || existing.description 
         });
       } else {
         topicsMap.set(t.id, t);

src/lib/learningService.js

@@ -1,51 +1,37 @@
-import { anthropicApi } from './api';
 import * as db from './db';
+import { callLLM } from './llm';
+import {
+  EMIT_LEARNING_ARTICLE_TOOL,
+  EMIT_LEARNING_SLIDES_TOOL,
+  EMIT_LEARNING_INFOGRAPHIC_TOOL,
+  EMIT_LEARNING_ALL_TOOL,
+  EMIT_CUSTOM_TOPIC_TOOL,
+  ARTICLE_PATCH_TOOLS,
+} from './llmTools';
+import { applyAndValidate } from './articlePatches';
 import { getCurriculumTopic } from './curriculumService';
 
 const CONTENT_GENERATION_SYSTEM = `You are an expert learning content writer for Respellion, an internal IT company.
 You write training material for employees based on knowledge topics.
 Always write in clear, professional English.
-ALWAYS return valid JSON only — no markdown code blocks, no extra text.`;
 
-const CONTENT_SCHEMA_ARTICLE = `{
+Emit the requested content through the matching tool — do not return prose JSON.`;
-  "article": {
-    "title": "Article title",
-    "intro": "Short intro of 1-2 sentences",
-    "sections": [
-      { "heading": "Section title", "body": "Section text of at least 3 sentences." }
-    ],
-    "keyTakeaways": ["Takeaway 1", "Takeaway 2", "Takeaway 3"]
-  }
-}`;
 
-const CONTENT_SCHEMA_SLIDES = `{
+const cachedSystem = (text) => [{ type: 'text', text, cache_control: { type: 'ephemeral' } }];
-  "slides": [
-    { "title": "Slide title", "bullets": ["Point 1", "Point 2", "Point 3"], "speakerNote": "Speaker note for this slide." }
-  ]
-}`;
 
+const TOOL_BY_TYPE = {
+  article: EMIT_LEARNING_ARTICLE_TOOL,
+  slides: EMIT_LEARNING_SLIDES_TOOL,
+  infographic: EMIT_LEARNING_INFOGRAPHIC_TOOL,
+  all: EMIT_LEARNING_ALL_TOOL,
+};
 
-
+const INSTRUCTIONS_BY_TYPE = {
-const CONTENT_SCHEMA_INFOGRAPHIC = `{
+  article: 'Provide at least 3 article sections and at least 2 key takeaways.',
-  "infographic": {
+  slides: 'Provide at least 4 slides.',
-    "headline": "A short, punchy headline summarizing the topic (max 8 words)",
+  infographic: 'Provide at least 3 stats and 3 steps.',
-    "tagline": "A subtitle of max 15 words",
+  all: 'Provide at least 3 article sections, 4 slides, 3 stats, and 3 steps in the infographic.',
-    "stats": [
+};
-      { "value": "Number or %", "label": "Short description", "icon": "📊" }
-    ],
-    "steps": [
-      { "number": 1, "title": "Step title", "description": "One-sentence description.", "icon": "🔑" }
-    ],
-    "quote": "An inspiring or insightful quote about the topic.",
-    "colorTheme": "teal"
-  }
-}`;
-
-const CONTENT_SCHEMA_ALL = `{
-  "article": ${CONTENT_SCHEMA_ARTICLE.replace(/^\{|\}$/g, '').trim()},
-  "slides": ${CONTENT_SCHEMA_SLIDES.replace(/^\{|\}$/g, '').trim()},
-  "infographic": ${CONTENT_SCHEMA_INFOGRAPHIC.replace(/^\{|\}$/g, '').trim()}
-}`;
 
 /**
  * Get the assigned topic for a given week.
@@ -53,7 +39,6 @@ const CONTENT_SCHEMA_ALL = `{
  * Falls back to hash-based assignment if no curriculum is configured.
  */
 export async function getAssignedTopic(userId, weekNumber) {
-  // Try curriculum first
   try {
     const { topic } = await getCurriculumTopic(weekNumber);
     if (topic && topic.learning_relevance !== 'exclude') return topic;
@@ -61,9 +46,7 @@ export async function getAssignedTopic(userId, weekNumber) {
     console.warn('[Learn] Curriculum lookup failed, falling back to hash:', e.message);
   }
 
-  // Fallback: hash-based assignment (backwards compatible)
   const allTopics = await db.getTopics();
-  // Filter out 'fact' type topics and 'exclude' relevance topics
   const topics = allTopics.filter(t => t.type !== 'fact' && t.learning_relevance !== 'exclude');
   if (!topics || topics.length === 0) return null;
 
@@ -96,29 +79,15 @@ export async function generateLearningContent(topic, force = false, selectedType
   let cached = null;
   if (!force) {
     cached = await db.getContent(topic.id);
-    if (cached) {
+    if (cached && cached[selectedType]) {
-      if (cached[selectedType]) {
       console.log(`[Learn] Cache hit for topic: ${topic.id} (${selectedType})`);
       return cached;
     }
   }
-  }
 
-  let schema = '';
+  const tool = TOOL_BY_TYPE[selectedType];
-  let instructions = '';
+  if (!tool) throw new Error(`Unknown learning content type: ${selectedType}`);
-  if (selectedType === 'all') {
+  const instructions = INSTRUCTIONS_BY_TYPE[selectedType];
-    schema = CONTENT_SCHEMA_ALL;
-    instructions = 'Provide at least 3 article sections, 4 slides, 3 stats, and 3-5 steps in the infographic.';
-  } else if (selectedType === 'article') {
-    schema = CONTENT_SCHEMA_ARTICLE;
-    instructions = 'Provide at least 3 article sections.';
-  } else if (selectedType === 'slides') {
-    schema = CONTENT_SCHEMA_SLIDES;
-    instructions = 'Provide at least 4 slides.';
-  } else if (selectedType === 'infographic') {
-    schema = CONTENT_SCHEMA_INFOGRAPHIC;
-    instructions = 'Provide at least 3 stats, and 3-5 steps in the infographic.';
-  }
 
   const prompt = `Generate a learning module piece for the following topic:
 
@@ -126,20 +95,20 @@ Label: ${topic.label}
 Type: ${topic.type}
 Description: ${topic.description}
 
-Return ONLY a JSON object with the following structure:
-${schema}
-
 ${instructions}`;
 
-  const responseText = await anthropicApi.generateContent(CONTENT_GENERATION_SYSTEM, prompt);
+  const result = await callLLM({
+    task: `learning.${selectedType}`,
+    tier: 'standard',
+    system: cachedSystem(CONTENT_GENERATION_SYSTEM),
+    user: prompt,
+    tools: [tool],
+    toolChoice: { type: 'tool', name: tool.name },
+    maxTokens: 8192,
+  });
 
-  let newContent;
+  const newContent = result.toolUses[0]?.input;
-  try {
+  if (!newContent) throw new Error('AI did not return learning content. Please try again.');
-    const jsonMatch = responseText.match(/\{[\s\S]*\}/);
-    newContent = JSON.parse(jsonMatch ? jsonMatch[0] : responseText);
-  } catch (e) {
-    throw new Error('AI could not generate valid learning content. Please try again.', { cause: e });
-  }
 
   const mergedContent = { ...(cached || {}), ...newContent };
   await db.setContent(topic.id, mergedContent);
@@ -148,28 +117,37 @@ ${instructions}`;
 
 export async function refineLearningContent(topic, refinementInstruction) {
   const existing = await db.getContent(topic.id);
+  if (!existing?.article) {
+    throw new Error('Refinement is currently only supported for the article. Generate an article for this topic first.');
+  }
 
-  const prompt = `You have previously generated the following learning module for the topic "${topic.label}":
+  const prompt = `You have previously generated the following article for the topic "${topic.label}":
 
-${JSON.stringify(existing, null, 2)}
+${JSON.stringify(existing.article, null, 2)}
 
 The admin has requested the following refinement:
 "${refinementInstruction}"
 
-Apply the refinement and return the complete updated JSON object using the same structure. Return ONLY valid JSON.`;
+Apply the refinement by calling one or more of the available patch tools. Make the smallest set of changes that satisfies the instruction — do not rewrite untouched sections.`;
 
-  const responseText = await anthropicApi.generateContent(CONTENT_GENERATION_SYSTEM, prompt);
+  const result = await callLLM({
+    task: 'learning.refine',
+    tier: 'standard',
+    system: cachedSystem(CONTENT_GENERATION_SYSTEM),
+    user: prompt,
+    tools: ARTICLE_PATCH_TOOLS,
+    toolChoice: { type: 'any' },
+    maxTokens: 4096,
+  });
 
-  let content;
+  if (!result.toolUses.length) {
-  try {
+    throw new Error('AI did not propose any changes for that instruction. Try a more specific request.');
-    const jsonMatch = responseText.match(/\{[\s\S]*\}/);
-    content = JSON.parse(jsonMatch ? jsonMatch[0] : responseText);
-  } catch (e) {
-    throw new Error('AI could not process the refinement. Please try a different instruction.', { cause: e });
   }
 
-  await db.setContent(topic.id, content);
+  const patchedArticle = applyAndValidate(existing.article, result.toolUses);
-  return content;
+  const merged = { ...existing, article: patchedArticle };
+  await db.setContent(topic.id, merged);
+  return merged;
 }
 
 export async function deleteCachedContent(topicId) {
@@ -177,30 +155,20 @@ export async function deleteCachedContent(topicId) {
 }
 
 export async function generateCustomTopic(label) {
-  const prompt = `A user wants to learn about "${label}".
+  const result = await callLLM({
-Create a short description (2-3 sentences) and categorize it.
+    task: 'topic.custom',
+    tier: 'standard',
+    system: cachedSystem('You are a knowledge graph AI categorising user-requested topics for the Respellion learning platform.'),
+    user: `A user wants to learn about "${label}". Provide a polished label, type, and 2–3 sentence description via the emit_custom_topic tool.`,
+    tools: [EMIT_CUSTOM_TOPIC_TOOL],
+    toolChoice: { type: 'tool', name: EMIT_CUSTOM_TOPIC_TOOL.name },
+    maxTokens: 1024,
+  });
 
-Return ONLY a JSON object with this structure:
+  const emitted = result.toolUses[0]?.input;
-{
+  if (!emitted) throw new Error('Could not process custom topic. Please try again.');
-  "label": "Polished topic title",
-  "type": "concept", // one of: concept, role, process
-  "description": "Short description"
-}`;
-
-  const responseText = await anthropicApi.generateContent(
-    "You are a knowledge graph AI categorizing topics.",
-    prompt
-  );
-
-  let newTopic;
-  try {
-    const jsonMatch = responseText.match(/\{[\s\S]*\}/);
-    newTopic = JSON.parse(jsonMatch ? jsonMatch[0] : responseText);
-    newTopic.id = 'custom_' + Date.now().toString(36);
-  } catch (e) {
-    throw new Error('Could not process custom topic. Please try again.', { cause: e });
-  }
 
+  const newTopic = { ...emitted, id: 'custom_' + Date.now().toString(36) };
   await db.upsertTopic(newTopic);
   return newTopic;
 }

src/lib/llm.js

@@ -125,7 +125,7 @@ function isChatLikeTask(task) {
   return task === 'legacy.chat' || task.startsWith('chat.') || task.startsWith('r42.');
 }
 
-const SIMULATION_EXTRACTION_PAYLOAD = JSON.stringify({
+const SIMULATION_EXTRACTION_GRAPH = {
   topics: [
     { id: 'radicale-transparantie', label: 'Radicale Transparantie', type: 'concept', description: 'De kernwaarde van Respellion waarbij alle informatie publiek toegankelijk is.', learning_relevance: 'core' },
     { id: 'kennisbeheer', label: 'Kennisbeheer', type: 'process', description: 'Het proces van het vastleggen en ontsluiten van organisatiekennis.', learning_relevance: 'standard' },
@@ -135,33 +135,87 @@ const SIMULATION_EXTRACTION_PAYLOAD = JSON.stringify({
     { source: 'kennisbeheer', target: 'radicale-transparantie', type: 'depends_on' },
     { source: 'wekelijkse-sessie', target: 'kennisbeheer', type: 'part_of' },
   ],
-});
+};
+
+const SIMULATION_EXTRACTION_PAYLOAD = JSON.stringify(SIMULATION_EXTRACTION_GRAPH);
 
 const SIMULATION_CHAT_TEXT =
   'Simulatiemodus staat aan — vraag een beheerder om Simulation Mode uit te zetten in Admin → Settings om met R42 te chatten.';
 
-async function simulatedResponse({ task }) {
+const SIMULATION_ARTICLE = {
-  await new Promise((r) => setTimeout(r, 400));
+  title: 'Voorbeeld leermodule',
-  if (isChatLikeTask(task)) {
+  intro: 'Dit is een simulatie. Schakel Simulation Mode uit om echte content te genereren.',
+  sections: [
+    { heading: 'Wat dit is', body: 'Dit is een placeholder-sectie die alleen verschijnt wanneer simulatiemodus aan staat. Hij illustreert de structuur van het artikel zonder een echte API-aanroep te doen. Dat is handig voor UI-werk.' },
+  ],
+  keyTakeaways: ['Simulatiemodus levert geen echte inhoud.', 'Schakel uit voor productie.'],
+};
+
+const SIMULATION_SLIDE = {
+  title: 'Voorbeeldslide',
+  bullets: ['Eerste punt', 'Tweede punt'],
+  speakerNote: 'Spreker-notitie ter illustratie.',
+};
+
+const SIMULATION_INFOGRAPHIC = {
+  headline: 'Simulatie',
+  tagline: 'Vervang door echte content',
+  stats: [{ value: '100%', label: 'simulatie', icon: '📊' }],
+  steps: [{ number: 1, title: 'Schakel uit', description: 'Zet simulatiemodus uit in Admin → Settings.', icon: '🔧' }],
+  quote: 'Een simulatie vertelt niets nieuws.',
+  colorTheme: 'teal',
+};
+
+const SIMULATION_TOOL_STUBS = {
+  emit_knowledge_graph: SIMULATION_EXTRACTION_GRAPH,
+  emit_handbook_delta: SIMULATION_EXTRACTION_GRAPH,
+  emit_learning_article: { article: SIMULATION_ARTICLE },
+  emit_learning_slides: { slides: [SIMULATION_SLIDE] },
+  emit_learning_infographic: { infographic: SIMULATION_INFOGRAPHIC },
+  emit_learning_all: { article: SIMULATION_ARTICLE, slides: [SIMULATION_SLIDE], infographic: SIMULATION_INFOGRAPHIC },
+  emit_custom_topic: { label: 'Simulatie onderwerp', type: 'concept', description: 'Een placeholder-onderwerp gegenereerd in simulatiemodus.' },
+  emit_quiz_questions: {
+    questions: [
+      {
+        id: 'sim-q1',
+        question: 'Wat doet simulatiemodus?',
+        topicLabel: 'Simulatie',
+        options: ['Echte API-aanroepen', 'Stub-data tonen', 'Niets', 'Crasht de app'],
+        correctIndex: 1,
+        explanation: 'Simulatiemodus retourneert vaste stub-data zonder de API te raken.',
+      },
+    ],
+  },
+  emit_graph_actions: { merges: [], deletions: [], newRelations: [], relevanceUpdates: [] },
+  set_intro: { intro: 'Bijgewerkte intro (simulatie).' },
+};
+
+function stubResponse({ stopReason = 'end_turn', text = '', toolUses = [] }) {
   return {
-      text: SIMULATION_CHAT_TEXT,
+    text,
-      toolUses: [],
+    toolUses,
-      stopReason: 'end_turn',
+    stopReason,
     usage: { input_tokens: 0, output_tokens: 0, cache_creation_input_tokens: 0, cache_read_input_tokens: 0 },
     requestId: null,
     model: 'simulation',
     durationMs: 400,
   };
 }
-  return {
+
-    text: SIMULATION_EXTRACTION_PAYLOAD,
+async function simulatedResponse({ task, toolChoice }) {
-    toolUses: [],
+  await new Promise((r) => setTimeout(r, 400));
-    stopReason: 'end_turn',
+
-    usage: { input_tokens: 0, output_tokens: 0, cache_creation_input_tokens: 0, cache_read_input_tokens: 0 },
+  if (toolChoice?.type === 'tool' && SIMULATION_TOOL_STUBS[toolChoice.name]) {
-    requestId: null,
+    return stubResponse({
-    model: 'simulation',
+      stopReason: 'tool_use',
-    durationMs: 400,
+      toolUses: [{ name: toolChoice.name, input: SIMULATION_TOOL_STUBS[toolChoice.name] }],
-  };
+    });
+  }
+
+  if (isChatLikeTask(task)) {
+    return stubResponse({ text: SIMULATION_CHAT_TEXT });
+  }
+  return stubResponse({ text: SIMULATION_EXTRACTION_PAYLOAD });
 }
 
 function linkSignals(userSignal, timeoutSignal) {
@@ -241,7 +295,7 @@ export async function callLLM(options) {
   if (!task) throw new Error('callLLM requires a `task` label.');
 
   const useSimulation = storage.get('admin:use_simulation') === true;
-  if (useSimulation) return simulatedResponse({ task });
+  if (useSimulation) return simulatedResponse({ task, toolChoice });
 
   const model = resolveModel(tier);
   const messagesPayload = buildMessages({ messages, user });

src/lib/llmSchemas.js

@@ -183,6 +183,31 @@ export const proposeGraphDeltaSchema = z.object({
   relations: z.array(deltaRelationSchema).max(5).optional(),
 });
 
+// ── Article patch operation schemas (Phase 2.4) ──────────────────────────────
+
+export const setIntroPatchSchema = z.object({
+  intro: z.string().min(1),
+});
+
+export const setSectionPatchSchema = z.object({
+  heading: z.string().min(1),
+  body: z.string().min(1),
+});
+
+export const addSectionPatchSchema = z.object({
+  heading: z.string().min(1),
+  body: z.string().min(1),
+  position: z.enum(['start', 'end']),
+});
+
+export const removeSectionPatchSchema = z.object({
+  heading: z.string().min(1),
+});
+
+export const replaceTakeawaysPatchSchema = z.object({
+  items: z.array(z.string().min(1)).min(1),
+});
+
 /**
  * Registry mapping known tool names to their input schemas. `callLLM`
  * consults this when the caller does not pass an explicit `toolSchemas`
@@ -199,4 +224,9 @@ export const toolSchemaRegistry = {
   emit_custom_topic: customTopicSchema,
   emit_graph_actions: graphActionsSchema,
   propose_graph_delta: proposeGraphDeltaSchema,
+  set_intro: setIntroPatchSchema,
+  set_section: setSectionPatchSchema,
+  add_section: addSectionPatchSchema,
+  remove_section: removeSectionPatchSchema,
+  replace_takeaways: replaceTakeawaysPatchSchema,
 };

src/lib/llmTools.js

@@ -0,0 +1,324 @@
+/**
+ * Anthropic tool definitions used by every structured-output flow.
+ *
+ * Each `tool_use` reply the model emits is validated against the matching
+ * Zod schema in `llmSchemas.js` (see `toolSchemaRegistry`). The two stay
+ * in lock-step on purpose — JSON Schema here drives the model, Zod there
+ * defends the application.
+ */
+
+const TOPIC_TYPES = ['concept', 'role', 'process'];
+const LEARNING_RELEVANCE = ['core', 'standard', 'peripheral', 'exclude'];
+const RELATION_TYPES_STRICT = ['related_to', 'depends_on', 'part_of', 'executed_by'];
+const RELATION_TYPES_LOOSE = ['related_to', 'depends_on', 'part_of', 'executed_by', 'executes'];
+
+const extractionTopicSchema = {
+  type: 'object',
+  properties: {
+    id: { type: 'string', description: 'kebab-case slug specific to the topic. Reuse existing IDs when the same concept recurs.' },
+    label: { type: 'string' },
+    type: { type: 'string', enum: TOPIC_TYPES },
+    description: { type: 'string', description: 'Max 3 sentences.' },
+    learning_relevance: { type: 'string', enum: LEARNING_RELEVANCE },
+  },
+  required: ['id', 'label', 'type', 'description', 'learning_relevance'],
+};
+
+const extractionRelationSchema = {
+  type: 'object',
+  properties: {
+    source: { type: 'string', description: 'Topic id.' },
+    target: { type: 'string', description: 'Topic id.' },
+    type: { type: 'string', enum: RELATION_TYPES_STRICT },
+  },
+  required: ['source', 'target', 'type'],
+};
+
+export const EMIT_KNOWLEDGE_GRAPH_TOOL = {
+  name: 'emit_knowledge_graph',
+  description: 'Return the complete knowledge graph extracted from the supplied source text — every distinct role, process and concept as a topic, plus the relations between them.',
+  input_schema: {
+    type: 'object',
+    properties: {
+      topics: { type: 'array', items: extractionTopicSchema },
+      relations: { type: 'array', items: extractionRelationSchema },
+    },
+    required: ['topics', 'relations'],
+  },
+};
+
+const handbookTopicSchema = {
+  type: 'object',
+  properties: {
+    ...extractionTopicSchema.properties,
+    metadata: {
+      type: 'object',
+      properties: { source: { type: 'string' } },
+    },
+  },
+  required: extractionTopicSchema.required,
+};
+
+const handbookRelationSchema = {
+  type: 'object',
+  properties: {
+    source: { type: 'string' },
+    target: { type: 'string' },
+    type: { type: 'string', enum: RELATION_TYPES_LOOSE },
+    description: { type: 'string' },
+  },
+  required: ['source', 'target', 'type'],
+};
+
+export const EMIT_HANDBOOK_DELTA_TOOL = {
+  name: 'emit_handbook_delta',
+  description: 'Return the topics and relations extracted from a handbook file update. Every process must have a role attached; every concept must connect to a process or role.',
+  input_schema: {
+    type: 'object',
+    properties: {
+      topics: { type: 'array', items: handbookTopicSchema },
+      relations: { type: 'array', items: handbookRelationSchema },
+    },
+    required: ['topics', 'relations'],
+  },
+};
+
+const articleSectionSchema = {
+  type: 'object',
+  properties: {
+    heading: { type: 'string' },
+    body: { type: 'string', description: 'At least three sentences.' },
+  },
+  required: ['heading', 'body'],
+};
+
+const articleBodySchema = {
+  type: 'object',
+  properties: {
+    title: { type: 'string' },
+    intro: { type: 'string', description: 'One or two sentences.' },
+    sections: { type: 'array', items: articleSectionSchema, minItems: 1 },
+    keyTakeaways: { type: 'array', items: { type: 'string' }, minItems: 1 },
+  },
+  required: ['title', 'intro', 'sections', 'keyTakeaways'],
+};
+
+const slideSchema = {
+  type: 'object',
+  properties: {
+    title: { type: 'string' },
+    bullets: { type: 'array', items: { type: 'string' }, minItems: 1 },
+    speakerNote: { type: 'string' },
+  },
+  required: ['title', 'bullets', 'speakerNote'],
+};
+
+const infographicStatSchema = {
+  type: 'object',
+  properties: {
+    value: { type: 'string' },
+    label: { type: 'string' },
+    icon: { type: 'string' },
+  },
+  required: ['value', 'label', 'icon'],
+};
+
+const infographicStepSchema = {
+  type: 'object',
+  properties: {
+    number: { type: 'integer', minimum: 1 },
+    title: { type: 'string' },
+    description: { type: 'string' },
+    icon: { type: 'string' },
+  },
+  required: ['number', 'title', 'description', 'icon'],
+};
+
+const infographicBodySchema = {
+  type: 'object',
+  properties: {
+    headline: { type: 'string', description: 'Punchy, max 8 words.' },
+    tagline: { type: 'string', description: 'Max 15 words.' },
+    stats: { type: 'array', items: infographicStatSchema, minItems: 1 },
+    steps: { type: 'array', items: infographicStepSchema, minItems: 1 },
+    quote: { type: 'string' },
+    colorTheme: { type: 'string', description: 'Tailwind colour token (e.g. "teal").' },
+  },
+  required: ['headline', 'tagline', 'stats', 'steps', 'quote', 'colorTheme'],
+};
+
+export const EMIT_LEARNING_ARTICLE_TOOL = {
+  name: 'emit_learning_article',
+  description: 'Return the article body for a learning module. At least three sections.',
+  input_schema: {
+    type: 'object',
+    properties: { article: articleBodySchema },
+    required: ['article'],
+  },
+};
+
+export const EMIT_LEARNING_SLIDES_TOOL = {
+  name: 'emit_learning_slides',
+  description: 'Return the slide deck for a learning module. At least four slides.',
+  input_schema: {
+    type: 'object',
+    properties: { slides: { type: 'array', items: slideSchema, minItems: 1 } },
+    required: ['slides'],
+  },
+};
+
+export const EMIT_LEARNING_INFOGRAPHIC_TOOL = {
+  name: 'emit_learning_infographic',
+  description: 'Return the infographic for a learning module. At least three stats and three steps.',
+  input_schema: {
+    type: 'object',
+    properties: { infographic: infographicBodySchema },
+    required: ['infographic'],
+  },
+};
+
+export const EMIT_LEARNING_ALL_TOOL = {
+  name: 'emit_learning_all',
+  description: 'Return article, slides and infographic for a learning module in one call.',
+  input_schema: {
+    type: 'object',
+    properties: {
+      article: articleBodySchema,
+      slides: { type: 'array', items: slideSchema, minItems: 1 },
+      infographic: infographicBodySchema,
+    },
+    required: ['article', 'slides', 'infographic'],
+  },
+};
+
+export const EMIT_CUSTOM_TOPIC_TOOL = {
+  name: 'emit_custom_topic',
+  description: 'Return a polished label, type and short description for a user-requested topic.',
+  input_schema: {
+    type: 'object',
+    properties: {
+      label: { type: 'string' },
+      type: { type: 'string', enum: TOPIC_TYPES },
+      description: { type: 'string', description: 'Two or three sentences.' },
+    },
+    required: ['label', 'type', 'description'],
+  },
+};
+
+const quizQuestionSchema = {
+  type: 'object',
+  properties: {
+    id: { type: 'string' },
+    question: { type: 'string' },
+    topicLabel: { type: 'string' },
+    options: { type: 'array', items: { type: 'string' }, minItems: 4, maxItems: 4 },
+    correctIndex: { type: 'integer', minimum: 0, maximum: 3 },
+    explanation: { type: 'string', description: 'Why the correct answer is correct (1–2 sentences).' },
+  },
+  required: ['id', 'question', 'topicLabel', 'options', 'correctIndex', 'explanation'],
+};
+
+export const EMIT_QUIZ_QUESTIONS_TOOL = {
+  name: 'emit_quiz_questions',
+  description: 'Return a batch of multiple-choice questions for a topic. Exactly four options each; correctIndex is 0-based.',
+  input_schema: {
+    type: 'object',
+    properties: { questions: { type: 'array', items: quizQuestionSchema, minItems: 1 } },
+    required: ['questions'],
+  },
+};
+
+export const EMIT_GRAPH_ACTIONS_TOOL = {
+  name: 'emit_graph_actions',
+  description: 'Return the actions to take on the knowledge graph: merges, deletions, new relations and relevance updates. Do not return the entire graph.',
+  input_schema: {
+    type: 'object',
+    properties: {
+      merges: {
+        type: 'array',
+        items: {
+          type: 'object',
+          properties: { keepId: { type: 'string' }, deleteId: { type: 'string' } },
+          required: ['keepId', 'deleteId'],
+        },
+      },
+      deletions: { type: 'array', items: { type: 'string' } },
+      newRelations: { type: 'array', items: extractionRelationSchema },
+      relevanceUpdates: {
+        type: 'array',
+        items: {
+          type: 'object',
+          properties: { id: { type: 'string' }, learning_relevance: { type: 'string', enum: LEARNING_RELEVANCE } },
+          required: ['id', 'learning_relevance'],
+        },
+      },
+    },
+  },
+};
+
+// ── Patch tools for refineLearningContent (Phase 2.4) ─────────────────────────
+
+export const SET_INTRO_TOOL = {
+  name: 'set_intro',
+  description: 'Replace the article intro with a new one or two sentences.',
+  input_schema: {
+    type: 'object',
+    properties: { intro: { type: 'string', description: 'New intro text.' } },
+    required: ['intro'],
+  },
+};
+
+export const SET_SECTION_TOOL = {
+  name: 'set_section',
+  description: 'Replace the body of an existing section, matched by its heading (case-insensitive). Use add_section if no section with that heading exists.',
+  input_schema: {
+    type: 'object',
+    properties: {
+      heading: { type: 'string', description: 'Heading of the section to replace.' },
+      body: { type: 'string', description: 'New body for that section, at least three sentences.' },
+    },
+    required: ['heading', 'body'],
+  },
+};
+
+export const ADD_SECTION_TOOL = {
+  name: 'add_section',
+  description: 'Insert a new section into the article at the start or end.',
+  input_schema: {
+    type: 'object',
+    properties: {
+      heading: { type: 'string' },
+      body: { type: 'string', description: 'At least three sentences.' },
+      position: { type: 'string', enum: ['start', 'end'] },
+    },
+    required: ['heading', 'body', 'position'],
+  },
+};
+
+export const REMOVE_SECTION_TOOL = {
+  name: 'remove_section',
+  description: 'Delete a section from the article, matched by its heading (case-insensitive).',
+  input_schema: {
+    type: 'object',
+    properties: { heading: { type: 'string' } },
+    required: ['heading'],
+  },
+};
+
+export const REPLACE_TAKEAWAYS_TOOL = {
+  name: 'replace_takeaways',
+  description: 'Replace the key takeaways list with a new one.',
+  input_schema: {
+    type: 'object',
+    properties: { items: { type: 'array', items: { type: 'string' }, minItems: 1 } },
+    required: ['items'],
+  },
+};
+
+export const ARTICLE_PATCH_TOOLS = [
+  SET_INTRO_TOOL,
+  SET_SECTION_TOOL,
+  ADD_SECTION_TOOL,
+  REMOVE_SECTION_TOOL,
+  REPLACE_TAKEAWAYS_TOOL,
+];

src/lib/testService.js

@@ -1,11 +1,15 @@
-import { anthropicApi } from './api';
 import * as db from './db';
+import { callLLM } from './llm';
+import { EMIT_QUIZ_QUESTIONS_TOOL } from './llmTools';
 import { getCurriculumTopic, getQuarterForWeek } from './curriculumService';
 
 const QUIZ_SYSTEM = `You are a quiz generator for Respellion, an internal IT company learning platform.
 You generate multiple-choice questions to test employee knowledge on specific topics.
 Always write in clear, professional English.
-ALWAYS return valid JSON only — no markdown code blocks, no extra text.`;
+
+Emit questions through the emit_quiz_questions tool. Each question has exactly four options; correctIndex is 0-based; mix difficulty roughly 4 easy / 4 medium / 2 hard.`;
+
+const cachedSystem = (text) => [{ type: 'text', text, cache_control: { type: 'ephemeral' } }];
 
 async function selectTestTopics(userId, weekNumber) {
   const allTopics = await db.getTopics();
@@ -66,45 +70,31 @@ export async function getCachedQuiz(userId, weekNumber) {
 export async function forceGenerateTopicQuestions(topic, count = 10) {
   let bank = await db.getQuizBank(topic.id);
 
-  const prompt = `Generate exactly ${count} multiple-choice quiz questions based on this knowledge topic:
+  const prompt = `Generate exactly ${count} multiple-choice quiz questions for this knowledge topic and emit them via the emit_quiz_questions tool:
 
 Topic: ${topic.label}
 Type: ${topic.type}
 Description: ${topic.description}
 
-Return ONLY a JSON object with this structure:
+Options must be prefixed "A) ", "B) ", "C) ", "D) ". Make questions specific and practical, not trivial.`;
-{
-  "questions": [
-    {
-      "id": "unique-id-string",
-      "question": "The question text",
-      "topicLabel": "${topic.label}",
-      "options": ["A) First option", "B) Second option", "C) Third option", "D) Fourth option"],
-      "correctIndex": 0,
-      "explanation": "A clear 1-2 sentence explanation of why the correct answer is correct."
-    }
-  ]
-}
 
-Rules:
+  const result = await callLLM({
-- Each question must have exactly 4 options.
+    task: 'quiz.generate',
-- correctIndex is 0-based (0=A, 1=B, 2=C, 3=D).
+    tier: 'standard',
-- Mix difficulty: 4 easy, 4 medium, 2 hard.
+    system: cachedSystem(QUIZ_SYSTEM),
-- Make questions specific and practical, not trivial.`;
+    user: prompt,
-
+    tools: [EMIT_QUIZ_QUESTIONS_TOOL],
-  const responseText = await anthropicApi.generateContent(QUIZ_SYSTEM, prompt);
+    toolChoice: { type: 'tool', name: EMIT_QUIZ_QUESTIONS_TOOL.name },
-  let newQuestions;
+    maxTokens: 4096,
-  try {
-    const jsonMatch = responseText.match(/\{[\s\S]*\}/);
-    const parsed = JSON.parse(jsonMatch ? jsonMatch[0] : responseText);
-    newQuestions = parsed.questions || [];
-    newQuestions.forEach(q => {
-      q.id = `${topic.id}-${Math.random().toString(36).substr(2, 9)}`;
   });
-  } catch (e) {
+
-    console.error('Failed to generate questions for topic', topic.label, e);
+  const emitted = result.toolUses[0]?.input;
-    throw new Error(`Could not generate questions for ${topic.label}`, { cause: e });
+  if (!emitted) throw new Error(`Could not generate questions for ${topic.label}`);
-  }
+
+  const newQuestions = (emitted.questions || []).map(q => ({
+    ...q,
+    id: `${topic.id}-${Math.random().toString(36).slice(2, 11)}`,
+  }));
 
   bank = [...bank, ...newQuestions];
   await db.setQuizBank(topic.id, bank);