Knowledge Base + RAG agent

Connect Notion, Confluence, or Google Drive to an in-room Fluxy agent using `@fluxy-chat/sdk` knowledge-base helpers and Worker agent invoke.

Overview

Piece	Role
`createKnowledgeBase()`	In-memory index (dev) or swap for Vectorize/D1 in production
`addSource()`	Register Notion/Confluence/GDrive connector config
`ingestDocument()`	Chunk + index page content
`buildRagContext()`	Top-k chunks → prompt prefix for the agent
`POST /agents/:id/invoke`	Run the agent with RAG-augmented system prompt

1. Register sources

import { createKnowledgeBase } from "@fluxy-chat/sdk";

const kb = createKnowledgeBase();

kb.addSource({
  type: "notion",
  name: "Product wiki",
  connectionConfig: {
    integrationToken: process.env.NOTION_TOKEN!,
    databaseId: process.env.NOTION_DB_ID!,
  },
  enabled: true,
  syncIntervalMs: 3600_000,
});

kb.addSource({
  type: "confluence",
  name: "Runbooks",
  connectionConfig: {
    baseUrl: "https://acme.atlassian.net/wiki",
    email: process.env.CONFLUENCE_EMAIL!,
    apiToken: process.env.CONFLUENCE_TOKEN!,
  },
  enabled: true,
  syncIntervalMs: 7200_000,
});

Store secrets in Worker secrets / your backend — never in the browser.

2. Sync documents

Pull pages from each provider on a cron or webhook, then ingest:

async function syncNotionPage(sourceId: string, title: string, markdown: string, url: string) {
  kb.ingestDocument(sourceId, {
    title,
    content: markdown,
    url,
    metadata: { provider: "notion" },
  });
}

For Google Drive, export Docs as plain text or Markdown before ingestDocument.

Chunk size defaults to 500 characters; tune per embedding model context window.

3. Build RAG context on each user message

function augmentPrompt(userMessage: string): string {
  const { synthesizedPrompt, results } = kb.buildRagContext({
    text: userMessage,
    maxResults: 5,
    minScore: 0.4,
  });

  if (results.length === 0) return userMessage;

  return `${synthesizedPrompt}\n\n---\nUser question: ${userMessage}`;
}

synthesizedPrompt lists cited chunks with titles and URLs for the model to ground answers.

4. Wire to Fluxy agent invoke

import { FluxyChatClient } from "@fluxy-chat/sdk";

const client = new FluxyChatClient({ baseUrl, userId: "support-bot", token: serviceJwt });

export async function answerWithKb(roomId: string, userMessage: string) {
  const prompt = augmentPrompt(userMessage);

  await client.invokeAgent(agentId, {
    roomId,
    prompt,
    // optional: pass source citations back as a card after reply
  });
}

In the dashboard, bind the agent to the support room under Agents → enable tools only if the KB miss rate is high.

5. Production hardening

Concern	Approach
Embeddings	Replace stub cosine in `createKnowledgeBase()` with Workers AI or external embeddings
Storage	Persist chunks in D1 + Vectorize instead of in-memory Maps
Freshness	Notion webhooks / Confluence polling cron via Worker `scheduled`
PII	Redact emails/phones before ingest; tenant-scoped source lists
Citations	Post a `Card` with doc links when `results.length > 0`

6. Verify locally

pnpm --filter @fluxychat/worker dev
Ingest a sample markdown doc via a small script calling kb.ingestDocument
Ask the agent a question only answerable from that doc
Confirm the reply cites the chunk title/URL