zquery — AI-Native Search Engine

Every token matters. Built for AI agents, not humans.

---

What is zquery?

zquery is a search engine designed specifically for AI agents. Unlike traditional search APIs that return walls of text optimized for human reading, zquery delivers progressive, token-efficient answers that let AI decide how deep to go.

# Install
curl -sSL https://zquery.dev/install | bash

# One-shot answer (~42 tokens)
$ zq "claude opus 4 context window"
1,000,000 tokens (since 2026-03)

# With sources
$ zq "claude opus 4 context window" -d 2
Claude Opus 4 context window:
  • Scale: 1M tokens
  • Released: 2026-03 official announcement  
  • Pricing: prompt caching -50%

sources: docs.anthropic.com, anthropic.com/news

# AI agent mode (JSON + token budget)
$ zq "..." --format json --budget 500

---

The Problem

Every AI search API today is built on top of human search engines — optimized for humans, not machines.

Service	What AI gets
Google Search API	Links + snippets, AI must fetch each one
Tavily	Fixed-format summaries, no control over depth
Perplexity API	English-first, poor Chinese quality
Serper	Essentially a Google proxy

Common problem: none of them understand what AI actually needs.

---

The Solution: Progressive Disclosure

AI agents default to Layer 1. They decide if they need more.

// Layer 1: Direct answer (~42 tokens, default)
GET /v1/query?q=claude+opus+4+context&depth=1
{
  "answer": "1,000,000 tokens (since 2026-03)",
  "confidence": 0.95,
  "tokens_used": 42,
  "more_available": true,
  "continuation": "eyJxIjoi..."
}

// Layer 2: Key points + data (~400 tokens)
GET /v1/query?continuation=...&depth=2
{
  "answer": "Claude Opus 4 context window:",
  "points": [
    { "text": "Scale: 1M tokens", "cite": 1 },
    { "text": "Released: 2026-03 official", "cite": 1 },
    { "text": "Pricing: prompt caching -50%", "cite": 2 }
  ],
  "sources_preview": [
    { "id": 1, "title": "Claude Opus 4 Release Notes", "domain": "docs.anthropic.com" },
    { "id": 2, "title": "What's New in Opus 4", "domain": "anthropic.com/news" }
  ],
  "tokens_used": 387
}

// Layer 3: Full synthesis with citations (~1500 tokens)
// Layer 4: Raw content chunks (~5000 tokens, on-demand)

---

Architecture

Query → [Intent Classification] → [Multi-Source Retrieval] → [Content Extraction] → [Layered Cache] → [Response]
          │                           │
          │                     ├─ Bing Search API (real-time index)
          │                     ├─ Whitelist fast-path (Wikipedia/MDN/Official Docs)
          │                     └─ Chinese-first sources (zhihu/WeChat/CSDN)
          │
          └─ Query type: factual / tutorial / news / code / opinion
             → different retrieval strategies

Tech stack:

• Backend: Go (single binary)
• Content extraction: go-readability
• LLM synthesis: Claude Haiku ($0.25/M tokens)
• Cache: Redis (Layer 1-2 cached 24h)
• Deploy: Docker, single VPS to start

Cost per query: ~$0.004 → price at $0.01 → 60% margin

---

Why Not Just Use Tavily?

	Tavily	zquery
Cost	Free tier → then pay	Always cheap (Bing API cost only)
Chinese	Mediocre	Native (zhihu/WeChat/MDN-CN priority)
Response format	Fixed	Progressive (Layer 1-4)
Token cost	Opaque	Returns tokens_used every call
Self-hostable	❌	✅
Source control	Black box	Customizable whitelist
CLI	❌	✅

---

Integrations

# OpenAI Function Calling
{
  "name": "zquery",
  "description": "Search the web for current information",
  "parameters": {
    "query": {"type": "string"},
    "depth": {"type": "integer", "enum": [1, 2, 3, 4], "default": 1},
    "budget": {"type": "integer", "description": "max tokens to use"}
  }
}

Works with:

• OpenAI Function Calling
• Anthropic Tool Use
• MCP (Claude Desktop / Cursor / Windsurf)
• LangChain / LlamaIndex plugins
• ZyHive (first showcase)

---

Pricing

Plan	Free	Pro $19/mo	Team $99/mo	Enterprise
Calls	100/day	10K/mo	100K/mo	Unlimited
Depth	1-2	All	All	All
Chinese optimization	✅	✅	✅	✅
Custom source whitelist	❌	❌	✅	✅
Self-hosted	❌	❌	❌	✅
SLA	best-effort	99%	99.5%	99.9%

---

Roadmap

Phase 1 — MVP (2 weeks)

• Go backend + Bing Search API + Haiku synthesis
• Chinese + English support
• CLI tool (zq)
• Open source: github.com/Zyling-ai/zquery
• ZyHive integration (first real user)

Phase 2 — Product (1 month)

• zquery.dev domain
• API key management + billing (Stripe)
• Layered cache (Redis)
• MCP protocol adapter
• Documentation site (VitePress)

Phase 3 — Ecosystem (3 months)

• Cursor / Claude Desktop integration guide
• LangChain / LlamaIndex plugin
• SDK: Python / Go / TypeScript
• Usage dashboard

Phase 4 — Differentiation (6 months)

• Vertical editions (finance, legal, medical, code)
• Enterprise private deployment
• Long-term memory (ZyHive memory tree integration)

---

Market

• Tavily (2023): $100M+ valuation (Series A)
• Exa.ai: $700M valuation
• Chinese market: completely empty
• Cursor, Windsurf, Manus, AutoGPT all looking for reliable search backends

---

Built by zyling.ai
Tagline: "Every token matters."