Unclassified

AI Search + AI Gateway + Workers AI Integration

Date: 2026-06-09 14:20 UTC
Status: ✅ INTEGRATED INTO INTELLIGENCE WORKER
Architecture: v3.6 Consolidated (no new workers)

Architecture Decision

REJECTED: Standalone integratewise-ai-orchestrator worker
ADOPTED: Integrate AI Search RAG into existing intelligence worker

Rationale:

v3.6 architecture already consolidates AI functionality into intelligence
Existing AI router (Workers AI + OpenRouter + AI Gateway) is working
AI Search namespace already created: integratewise-intelligence
Adding RAG is a feature enhancement, not a new service

What's Wired (2026-06-09)

✅ Intelligence Worker Bindings

File: services/intelligence/wrangler.toml

# AI Search for RAG
[[ai_search_namespaces]]
binding = "AI_SEARCH"
namespace = "integratewise-intelligence"

# Workers AI for fast inference
[ai]
binding = "AI"

# OpenRouter API key (free tier)
[[secrets_store_secrets]]
binding = "OPENROUTER_API_KEY"
store_id = "1fd83fc5881e4cd296ae3c2fbe80693c"
secret_name = "OPENROUTER_API_KEY"

# Env vars
[vars]
AI_GATEWAY_ID = "integratewise-ai-dev"

✅ AI Router with RAG Support

File: services/intelligence/src/lib/ai-router.ts

Features:

3-tier fallback: Workers AI → OpenRouter free → OpenRouter fusion
Optional RAG context from AI Search
All requests through AI Gateway (caching + logging)
Source attribution for RAG responses

New Interface:

export interface AIRequest {
  prompt: string;
  systemPrompt?: string;
  maxTokens?: number;
  temperature?: number;
  tier?: "fast" | "balanced" | "critical";
  useRAG?: boolean; // NEW: Enable AI Search RAG
}

export interface AIResponse {
  content: string;
  model: string;
  cached: boolean;
  provider: "workers-ai" | "openrouter";
  cost: number;
  sources?: Array<{ title: string; url?: string; score?: number }>; // NEW: RAG sources
}

RAG Flow:

If useRAG: true, search AI Search namespace with user query
Retrieve top 5 results
Augment prompt with retrieved context
Route through AI tier (fast/balanced/critical)
Return answer + source attribution

✅ Test Endpoint

File: services/intelligence/src/index.ts

New Route: POST /v1/test/rag

Request:

{
  "question": "What integrations does integratewise.ai support?",
  "tier": "fast"
}

Response:

{
  "success": true,
  "data": {
    "content": "IntegrateWise supports...",
    "model": "@cf/meta/llama-3.1-8b-instruct",
    "cached": false,
    "provider": "workers-ai",
    "cost": 0,
    "sources": [
      {
        "title": "Supported Integrations",
        "url": "https://integratewise.ai/integrations",
        "score": 0.92
      }
    ]
  }
}

Deployment

Step 1: Deploy Intelligence Worker

cd /Users/nirmal/Github/integratewise-live
unset CLOUDFLARE_API_TOKEN
wrangler deploy --config services/intelligence/wrangler.toml --env=""

Expected Output:

✅ Worker deployed
✅ AI_SEARCH binding verified
✅ AI binding verified
✅ OPENROUTER_API_KEY binding verified

Step 2: Test RAG Search

# Test RAG-enabled query (fast tier)
curl -X POST https://intelligence.dev.integratewise.ai/v1/test/rag \
  -H "Content-Type: application/json" \
  -d '{
    "question": "What is IntegrateWise and what does it do?",
    "tier": "fast"
  }'

# Test with balanced tier (OpenRouter free)
curl -X POST https://intelligence.dev.integratewise.ai/v1/test/rag \
  -H "Content-Type: application/json" \
  -d '{
    "question": "How does IntegrateWise handle entity resolution?",
    "tier": "balanced"
  }'

# Test with critical tier (OpenRouter fusion)
curl -X POST https://intelligence.dev.integratewise.ai/v1/test/rag \
  -H "Content-Type: application/json" \
  -d '{
    "question": "Compare IntegrateWise to other integration platforms",
    "tier": "critical"
  }'

Expected Behavior:

Query hits AI Search namespace
Retrieves relevant docs from integratewise.ai
Augments prompt with context
Returns answer + sources

Step 3: Verify AI Gateway Logs

Open Cloudflare Dashboard
Navigate to AI Gateway → integratewise-ai-dev
Check logs for:
- Cache hit rate
- Request counts by provider (workers-ai, openrouter)
- Token usage
- Error rates

AI Search Content Status

Current Status

Namespace: integratewise-intelligence
Status: ✅ Created
Content: ⚠️ EMPTY (needs bulk upload)

Next: Bulk Upload Content

Option A: Crawl Website (Recommended)

# Enable Cloudflare AI Search crawler
# 1. Go to Cloudflare Dashboard → AI → AI Crawl Control
# 2. Allow "Cloudflare-AI-Search" bot for integratewise.ai domain
# 3. Trigger initial crawl:

curl -X POST \
  "https://api.cloudflare.com/client/v4/accounts/a1bbbb12a32cdbb68dd170b09fe8b5f3/ai-search/rags/integratewise-intelligence/refresh" \
  -H "Authorization: Bearer ${CLOUDFLARE_API_TOKEN}" \
  -H "Content-Type: application/json"

Option B: Bulk Upload from org_memory (Manual)

# Create script to upload org_memory content
# File: services/intelligence/scripts/upload-to-ai-search.ts

const documents = await fetchOrgMemoryDocs(); // From Supabase

for (const doc of documents) {
  await env.AI_SEARCH.upsert({
    id: doc.id,
    text: doc.content_md,
    metadata: {
      title: doc.title,
      url: doc.url,
      tags: doc.tags,
      topics: doc.topics,
      governance_state: doc.governance_state,
    },
  });
}

Option C: Sync from R2 Bucket

# If org_memory content is already in R2, sync to AI Search
# (Requires AI Search API integration)

What's Missing (Optional)

Hugging Face API Key

Status: Not configured
Priority: 🟢 LOW (optional, OpenRouter covers this)

To Add:

wrangler secrets-store secret create 1fd83fc5881e4cd296ae3c2fbe80693c \
  --name=HUGGINGFACE_API_KEY \
  --scopes=workers

# Add binding to intelligence/wrangler.toml
[[secrets_store_secrets]]
binding = "HF_API_KEY"
store_id = "1fd83fc5881e4cd296ae3c2fbe80693c"
secret_name = "HUGGINGFACE_API_KEY"

Grok API Key

Status: Not configured
Priority: 🟢 LOW (optional, covered by OpenRouter + Workers AI)

To Add:

wrangler secrets-store secret create 1fd83fc5881e4cd296ae3c2fbe80693c \
  --name=GROK_API_KEY \
  --scopes=workers

# Add binding to intelligence/wrangler.toml
[[secrets_store_secrets]]
binding = "GROK_API_KEY"
store_id = "1fd83fc5881e4cd296ae3c2fbe80693c"
secret_name = "GROK_API_KEY"

Note: These are optional fallback providers. Current setup (Workers AI + OpenRouter free) is sufficient for Phase 2.

Usage in Production

Morning Brief with RAG


async function generateMorningBrief(env: Env) {
  const response = await routeAI(
    {
      prompt: "Summarize overnight system changes and pending actions",
      systemPrompt: "You are a system monitor. Use knowledge base context.",
      useRAG: true,
      tier: "fast",
      maxTokens: 500,
    },
    env
  );

  return {
    content: response.content,
    sources: response.sources, // Attribution
    provider: response.provider,
  };
}

Signal Analysis with RAG

async function analyzeSignal(env: Env, signalData: unknown) {
  const response = await routeAI(
    {
      prompt: `Analyze this signal: ${JSON.stringify(signalData)}`,
      systemPrompt: "You are a signal analyzer. Reference knowledge base for context.",
      useRAG: true,
      tier: "balanced",
      maxTokens: 500,
    },
    env
  );

  return {
    analysis: response.content,
    sources: response.sources,
  };
}

Critical Decision with Fusion

async function criticalDecision(env: Env, context: unknown) {
  const response = await routeAI(
    {
      prompt: `Critical decision required: ${JSON.stringify(context)}`,
      systemPrompt: "Multi-model deliberation. Use knowledge base + reasoning.",
      useRAG: true,
      tier: "critical", // OpenRouter fusion
      maxTokens: 1000,
    },
    env
  );

  return {
    recommendation: response.content,
    sources: response.sources,
    model: response.model, // Will show "openrouter/fusion"
  };
}

Cost Analysis

Tier	Provider	Model	Cost	Use Case
Fast	Workers AI	llama-3.1-8b	FREE	Default, morning briefs, quick queries
Balanced	Workers AI	moonshotai/kimi-k3	Paid (per-token)	Signal analysis, medium reasoning, 1M context
Critical	Workers AI	moonshotai/kimi-k3 (`reasoning_effort: 'max'`)	Paid (per-token)	Critical decisions, deep deliberation
Fallback	OpenRouter	openrouter/auto	FREE (tier)	Fallback when Workers AI unavailable

Kimi K3 (moonshotai/kimi-k3): 2.8T parameter frontier model on Workers AI. 1M token context window, always-on reasoning with reasoning_effort parameter. Edge-native — no external API call needed. See Cloudflare docs.

AI Gateway Benefits:

Request caching (reduces API calls)
Rate limiting (prevents abuse)
Unified logs (all providers in one place)
Cost tracking

Success Criteria

Phase 2 Complete When:

✅ Intelligence worker deployed with AI Search binding
✅ AI router supports RAG (useRAG: true)
✅ Test endpoint /v1/test/rag working
✅ AI Gateway logs showing traffic
⏳ AI Search namespace populated with content (bulk upload)
⏳ Production code uses RAG for morning briefs + signal analysis

Next Steps

Immediate (After Deployment)

Deploy intelligence worker:

wrangler deploy --config services/intelligence/wrangler.toml --env=""

Test RAG endpoint:

curl -X POST https://intelligence.dev.integratewise.ai/v1/test/rag \
  -H "Content-Type: application/json" \
  -d '{"question":"What is IntegrateWise?","tier":"fast"}'

Verify AI Gateway logs:
- Check cache hit rate
- Verify requests routing correctly

Phase 3: Content Upload

Option A: Website crawl (simplest)
- Enable AI Search crawler for integratewise.ai
- Trigger initial crawl
- Wait ~1 hour for indexing
Option B: Bulk upload org_memory (more control)
- Write script to fetch org_memory (governance_state=approved)
- Upload to AI Search via API
- Map metadata (title, url, tags, topics)

Phase 4: Production Integration

Update morning brief handler to use RAG
Update signal analysis to use RAG
Add RAG to MCP tools (knowledge search)
Monitor usage + costs

Files Modified

services/intelligence/src/lib/ai-router.ts
- Added useRAG parameter to AIRequest
- Added sources field to AIResponse
- Implemented AI Search integration in routeAI()
- RAG context augmentation before AI call
services/intelligence/src/index.ts
- Added POST /v1/test/rag endpoint
- Integrated routeAI with RAG support
services/intelligence/wrangler.toml
- Already has AI_SEARCH binding ✅
- Already has AI binding ✅
- Already has OPENROUTER_API_KEY binding ✅

Status: READY FOR DEPLOYMENT
Breaking Changes: None
Rollback Available: Yes
Last Updated: 2026-06-09T14:20:00Z