The Problem
Agents that scrape raw HTML or ingest unstructured search snippets dump thousands of tokens into the context window per query. A single web page can consume 8,000-15,000 tokens after HTML-to-text conversion. With GPT-4o at $2.50 per million input tokens and Claude at $3, a 50-query session burns $0.50-$1.00 on search context alone. Most of those tokens are navigation bars, footers, and boilerplate the model ignores.
The Scavio Solution
Replace raw page fetches with Scavio's structured JSON endpoint. Each result returns title, snippet, link, and optional AI Overview text -- typically 50-100 tokens per result versus 8,000+ for a raw page. Feed the top five results into the context for under 500 tokens total. The agent gets the same factual grounding at 1/20th the token cost.
Before
Before switching, an agent fetched full pages via a scraper, converted HTML to text, and stuffed it into the prompt. A five-source grounding step consumed 40,000 tokens per query. At 200 queries per day, the team spent $40/day on input tokens alone.
After
After switching to structured search, the same five-source grounding step uses 500 tokens. Daily token cost dropped from $40 to $2 for the search context portion. Agent response latency improved by 3 seconds because the model processes less input.
Who It Is For
AI engineers building search-augmented agents who want to reduce LLM inference costs and latency by minimizing context window usage.
Key Benefits
- Cut search context from 40,000 tokens to 500 tokens per query
- Reduce LLM input cost by 95% for search-grounded agents
- Faster inference because the model processes less context
- Structured fields map directly to tool-call schemas
- No HTML parsing, no boilerplate stripping
Python Example
import requests, os
H = {'x-api-key': os.environ['SCAVIO_API_KEY']}
def lean_context(query: str, n: int = 5) -> str:
"""Return structured search context under 500 tokens."""
r = requests.post('https://api.scavio.dev/api/v1/search', headers=H,
json={'platform': 'google', 'query': query}, timeout=10).json()
lines = []
for o in r.get('organic', [])[:n]:
lines.append(f"[{o.get('title')}]({o.get('link')})\n{o.get('snippet')}")
return '\n\n'.join(lines)
# Before: fetched 5 full pages = ~40,000 tokens
# After: structured results = ~400 tokens
ctx = lean_context('best search api for ai agents 2026')
print(f'Context length: ~{len(ctx)//4} tokens')
print(ctx)JavaScript Example
const H = { 'x-api-key': process.env.SCAVIO_API_KEY, 'Content-Type': 'application/json' };
async function leanContext(query, n = 5) {
const r = await fetch('https://api.scavio.dev/api/v1/search', {
method: 'POST', headers: H,
body: JSON.stringify({ platform: 'google', query })
}).then(r => r.json());
return (r.organic || []).slice(0, n).map(o =>
`[${o.title}](${o.link})\n${o.snippet}`
).join('\n\n');
}
const ctx = await leanContext('best search api for ai agents 2026');
console.log(`Context length: ~${Math.ceil(ctx.length / 4)} tokens`);
console.log(ctx);Platforms Used
Web search with knowledge graph, PAA, and AI overviews
Community, posts & threaded comments from any subreddit
YouTube
Video search with transcripts and metadata