Pipeline

Document
    │
    ▼
Chunking ──── markdown (by headers)
               paragraph (by double newlines)
               sliding (fixed window + overlap)
    │
    ▼
Embedding ─── Workers AI (@cf/baai/bge-base-en-v1.5)
               Free embeddings, ~100ms per call
    │
    ▼
Storage ───── Vectorize (vectors) + D1 (metadata)
               Free tier: 5M vectors + 5GB SQLite
    │
    ▼
Query ─────── Embed query → Vectorize search → D1 metadata lookup
               ~20ms edge latency

Chunking Strategies

Strategy	Best For	Split Logic
markdown	Documentation	Split on ##, ### headers
paragraph	Articles	Split on double newlines
sliding	Unstructured	Fixed window (512 chars) + overlap (64 chars)

Cost Comparison

Service	TideRAG	Pinecone	Supabase
Monthly cost	$0	$70+	$25+
Free vectors	5M	0	~100K
Embeddings	Free	$0.02/1K	$0.02/1K
Latency	~20ms	~80ms	~50ms

Cloudflare Free Tier Limits

Service	Free Limit	Overage
Vectorize	5M vectors	$0.50/M
Workers AI	10K embeddings/day	~$0.001/1K
D1	5GB storage	$0.75/GB

Agent Integration

Wrap TideRAG as a Tinkuy tool:

const searchDocs = defineTool({
  name: 'search_docs',
  description: 'Search documentation',
  execute: async (args) => {
    const results = await rag.query(args.query, { topK: 3 });
    return results.map(r => r.text).join('\n---\n');
  },
});