rve/learning-platform
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(r42): improve KB grounding accuracy and add clear-history

Browse Source 
R42 was missing knowledge-graph information (e.g. pension questions)
because retrieval and context-building dropped relevant facts:

- retrieval: exact-token TF-IDF could not match Dutch compound words,
  so a "pensioen" query scored 0 against "pensioenregeling" /
  "partnerpensioen" and never retrieved them. Add a compound-word
  fallback (shared >=6-char stem or containment, 0.4x weight) alongside
  exact matching.
- rag: deep article content was only injected for verbatim-mentioned
  topics; retrieved topics contributed just a 200-char description.
  Inject ~1000 chars of content for up to 5 topics (mentions first,
  then top-ranked retrieved) and widen the description snippet to 320.
- prompts: add a NAUWKEURIGHEID block (use all relevant facts, call
  lookup_topic before giving up) and relax the 4-sentence cap for
  detail/list answers so complete facts aren't summarised away.

Also add a clear-history control: a trash button in the chat header
(confirm dialog) wipes chat🧵{userId} and reseeds the greeting
via clearThread() in useChat.

Tests: compound-word matching + rag deep-content injection. Spec updated.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
RaymondVerhoef
2026-07-13 14:25:08 +02:00
parent e73700763d
commit 85452f66a7
9 changed files with 239 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
src/components/chat/rag.js
View File

@@ -2,6 +2,13 @@ import * as db from '../../lib/db';
import { buildIndex, retrieveTopK } from '../../lib/retrieval';
const TOP_K = 10;
// How many topics get their full article body injected (not just the short
// description). Verbatim-mentioned topics come first, then the highest-ranked
// retrieved ones, so a query that never names a topic exactly still gets rich
// content for what it matched.
const DEEP_CONTENT_LIMIT = 5;
const DEEP_SNIPPET_CHARS = 1000;
const DESC_SNIPPET_CHARS = 320;
async function sha256Hex(input) {
const enc = new TextEncoder().encode(input);
@@ -71,7 +78,7 @@ export async function buildKbContext(userMessage = '') {
const included = [...includedById.values()];
const topicLines = included.map(t => {
const desc = (t.description || '').replace(/\s+/g, ' ').trim().slice(0, 200);
const desc = (t.description || '').replace(/\s+/g, ' ').trim().slice(0, DESC_SNIPPET_CHARS);
return `- ${t.id} (${t.type || 'concept'}) "${t.label}": ${desc}`;
});
@@ -85,19 +92,30 @@ export async function buildKbContext(userMessage = '') {
}
}
const mentionedDeepContent = [];
for (const id of mentionedIds) {
const t = includedById.get(id);
if (!t) continue;
const content = await db.getContent(t.id).catch(() => null);
if (!content) continue;
let raw;
if (typeof content === 'string') raw = content;
else if (content.article) raw = typeof content.article === 'string' ? content.article : JSON.stringify(content.article);
else raw = JSON.stringify(content);
const snippet = raw.replace(/\s+/g, ' ').trim().slice(0, 1200);
mentionedDeepContent.push(`### ${t.label}\n${snippet}`);
// Pick which topics get their full article body: verbatim mentions first,
// then the highest-ranked retrieved topics, capped at DEEP_CONTENT_LIMIT.
const deepIds = [];
for (const id of mentionedIds) deepIds.push(id);
for (const t of retrieved) {
if (deepIds.length >= DEEP_CONTENT_LIMIT) break;
if (!mentionedIds.has(t.id)) deepIds.push(t.id);
}
const deepBlocks = await Promise.all(
deepIds.slice(0, DEEP_CONTENT_LIMIT).map(async (id) => {
const t = includedById.get(id);
if (!t) return null;
const content = await db.getContent(id).catch(() => null);
if (!content) return null;
let raw;
if (typeof content === 'string') raw = content;
else if (content.article) raw = typeof content.article === 'string' ? content.article : JSON.stringify(content.article);
else raw = JSON.stringify(content);
const snippet = raw.replace(/\s+/g, ' ').trim().slice(0, DEEP_SNIPPET_CHARS);
if (!snippet) return null;
return `### ${t.label}\n${snippet}`;
}),
);
const mentionedDeepContent = deepBlocks.filter(Boolean);
const context = [
`KENNISGRAAF — RELEVANTE TOPICS (top ${included.length} van ${allTopics.length}):`,
@@ -106,7 +124,7 @@ export async function buildKbContext(userMessage = '') {
`KENNISGRAAF — RELATIES (binnen deze selectie):`,
relLines.length ? relLines.join('\n') : '(geen relaties binnen deze selectie)',
mentionedDeepContent.length
? `\n\nDIEPERE INHOUD (voor genoemde topics):\n${mentionedDeepContent.join('\n\n')}`
? `\n\nDIEPERE INHOUD (volledige leerinhoud van de meest relevante topics — gebruik álle feiten hieruit die de vraag beantwoorden):\n${mentionedDeepContent.join('\n\n')}`
: '',
``,
`Als de relevante context hierboven te beperkt is, gebruik dan de tool "lookup_topic" om de volledige beschrijving en eventuele leerinhoud van een specifiek topic op te halen.`,