Token Optimization

Every token costs money. QANATIX is designed to minimize token usage while maximizing information density.

The problem

A typical web-scraped result sends 2,000-5,000 tokens to your LLM — HTML fragments, navigation, ads, boilerplate. Most of it is noise. At scale, this burns budget fast.

QANATIX sends ~120 tokens per result in compact format. That's 15-40x more efficient than web scraping.

How it works

1. Structured data in, structured data out

Your data is ingested as structured fields (vertical_data), not free text. QANATIX returns exactly the fields that matter — no parsing needed.

2. Three output formats

3. Compact format details

The compact format returns a markdown table:

| # | Name | Score | Key Data |
|---|------|-------|----------|
| 1 | Stainless Steel Bolt M8x40 A2 | 0.87 | part_number: SS-M8-40-A2, material: Stainless Steel A2, price_eur: 0.12 |

Key optimizations:

• Only top-level vertical_data fields included (no nested objects)
• Field names abbreviated where possible
• No entity IDs, timestamps, or metadata unless needed
• Score included for relevance context

4. MCP uses compact automatically

The MCP server returns compact format by default. When Claude calls qanatix_search(), it gets the markdown table — maximum information, minimum tokens.

5. Limit your results

Default limit is 20, but most AI queries need 3-5 results. Set limit explicitly:

{"query": "M8 bolt", "limit": 5, "format": "compact"}

5 results in compact = ~600 tokens. Same query in full JSON = ~4,000 tokens.

Token budget comparison

At $3/M input tokens (Claude Sonnet), 1,000 queries/day with 10 results each:

• Web scraping: ~$90/day
• QANATIX compact: ~$3.60/day

Best practices

1. Use compact format for all AI agent integrations
2. Set explicit limits — don't fetch 20 results if you need 5
3. Use filters to narrow results before search, reducing irrelevant matches
4. Use the MCP server — it handles format and limit optimization automatically
5. Cache on your side if you're making repeated identical queries (QANATIX also caches for 30 seconds server-side)