knowledge-evaluator

A Cloudflare Worker that scores conversation excerpts using Workers AI and routes them based on confidence — automatically promoting high-signal knowledge to Notion and surfacing ambiguous items for human review.

Built for the Notion MCP Challenge.

How it works

Conversation excerpt
        ↓
Workers AI (Llama 3.3 70B) scores 3 signals
        ↓
Score ≥ 0.67 → Knowledge Memory (auto-promoted via Notion REST)
Score 0.33–0.67 → Review Queue (Pending, awaiting human judgment)
Score < 0.33 → Discarded
        ↓
Claude Desktop queries Review Queue via @notionhq/notion-mcp-server
        ↓
Human approves or rejects in Notion

Notion isn't a log — it's the decision surface.

The three signals

Signal	Question
Usage	Is there a concrete technique, command, or pattern being applied?
Validation	Is it confirmed to work?
Specificity	Is it actionable, not just vague advice?

Each signal is binary (0 or 1). Score = signals fired / 3.

Live demo

Endpoint: https://knowledge-evaluator.fpl-test.workers.dev

curl -X POST https://knowledge-evaluator.fpl-test.workers.dev/evaluate \
  -H "Content-Type: application/json" \
  -d '{"text": "Use db.prepare().bind().all() for D1 batch queries.", "source": "dev-chat"}'

Full write-up: DEV.to article
Demo video: YouTube
MCP demo: YouTube

Stack

• Cloudflare Workers — runtime
• Workers AI (Llama 3.3 70B fp8) — evaluator
• Hono — routing
• Notion REST API — writes to Review Queue + Knowledge Memory
• @notionhq/notion-mcp-server — MCP query layer (reads)

Deploy your own

1. Clone and install

git clone https://github.com/dannwaneri/knowledge-evaluator
cd knowledge-evaluator
npm install

2. Create two Notion databases

Review Queue — properties:

• Name (title)
• Status (select: Pending, Approved, Rejected)
• Score (number, percent)
• Signals (multi-select: usage, validation, specificity)
• Raw Context (rich text)
• Source (rich text)
• Created (date)

Knowledge Memory — properties:

• Name (title)
• Score (number, percent)
• Signals (multi-select)
• Source (rich text)
• Provenance (rich text)
• Promoted At (date)

3. Create a Notion integration

Go to notion.so/profile/integrations, create an internal integration, and connect it to both databases.

4. Set secrets

wrangler secret put NOTION_TOKEN
wrangler secret put REVIEW_QUEUE_ID
wrangler secret put KNOWLEDGE_MEMORY_ID

5. Deploy

wrangler deploy

Endpoints

Method	Path	Description
GET	/	Health check
POST	/evaluate	Score a knowledge item and route it
GET	/pending	Query Notion for items awaiting review

POST /evaluate

{
  "text": "Your conversation excerpt here",
  "source": "optional-source-label"
}

Response:

{
  "score": 67,
  "signals": ["usage", "specificity"],
  "summary": "One sentence summary from the model",
  "destination": "review_queue",
  "notion_page_id": "page-id"
}

Bidirectional MCP

The Worker writes to Notion via REST. To read pending items back via MCP, add @notionhq/notion-mcp-server to your Claude Desktop config:

{
  "mcpServers": {
    "notionApi": {
      "command": "npx",
      "args": ["-y", "@notionhq/notion-mcp-server"],
      "env": {
        "NOTION_TOKEN": "your-token"
      }
    }
  }
}

Then ask Claude: "Query my Notion Review Queue and show me all items with Status Pending"

---

Extracted from Foundation — a federated knowledge system built on Cloudflare Workers.