Local LLM research stacks (Ollama, LMStudio, LocalAI) provide privacy and zero per-token cost but lack access to current web data. Adding a search API as a tool gives your local model live grounding without sending your prompts to cloud LLMs. One HTTP call returns structured results your local model can cite.
Prerequisites
- Ollama or LMStudio running locally
- Python 3.8+
- A Scavio API key (free tier: 50 credits on signup)
Walkthrough
Step 1: Create the search tool function
Build a simple function that your local LLM can call for web search.
Python
import requests, os
# Scavio has one endpoint per platform - there is no dispatcher endpoint and no
# `platform` request param, so the selector lives in your code.
SCAVIO = "https://api.scavio.dev"
SCAVIO_ENDPOINTS = {
"google": "/api/v2/google",
"reddit": "/api/v1/reddit/search",
"youtube": "/api/v1/youtube/search",
"amazon": "/api/v1/amazon/search",
"walmart": "/api/v1/walmart/search",
}
SCAVIO_QUERY_KEY = {"youtube": "search"}
SCAVIO_RESULTS_KEY = {"google": "organic_results", "reddit": "results",
"youtube": "results", "amazon": "products", "walmart": "products"}
def scavio_url(platform):
return SCAVIO + SCAVIO_ENDPOINTS[platform or "google"]
def scavio_body(platform, query):
return {SCAVIO_QUERY_KEY.get(platform or "google", "query"): query}
def scavio_payload(payload, platform="google"):
"""Google v2 passes Google's response through as-is; every other endpoint
wraps its payload in `data`. Item fields differ per platform (see
https://scavio.dev/docs), so only the result list is normalised here."""
platform = platform or "google"
out = payload if platform == "google" else payload["data"]
return {**out, "results": out.get(SCAVIO_RESULTS_KEY[platform], [])}
SCAVIO_KEY = os.environ.get('SCAVIO_API_KEY', 'your_key_here')
def web_search(query: str, platform: str = 'google') -> str:
"""Search the web and return structured results for the local LLM."""
resp = requests.post(scavio_url(platform), headers={'Authorization': f'Bearer {SCAVIO_KEY}', 'Content-Type': 'application/json'}, json=scavio_body(platform, query), timeout=10)
results = scavio_payload(resp.json(), platform).get('results', [])[:5]
# Format for local LLM context (plain text, token-efficient)
lines = []
for r in results:
lines.append(f"- {r.get('title','')}: {r.get('snippet','')} ({r.get('link','')})")
return '\n'.join(lines) if lines else 'No results found.'Step 2: Integrate with Ollama
Use Ollama's tool calling to invoke web search when needed.
Python
import ollama
def research_with_search(question: str) -> str:
# First, ask the model if it needs search
response = ollama.chat(model='llama3.2', messages=[
{'role': 'system', 'content': 'You are a research assistant. If you need current information, say SEARCH: <query>. Otherwise answer directly.'},
{'role': 'user', 'content': question}
])
answer = response['message']['content']
# If model requests search, fetch and re-prompt
if 'SEARCH:' in answer:
query = answer.split('SEARCH:')[1].strip()
search_results = web_search(query)
response = ollama.chat(model='llama3.2', messages=[
{'role': 'system', 'content': 'Answer using the search results below.'},
{'role': 'user', 'content': f'Question: {question}\n\nSearch results:\n{search_results}'}
])
return response['message']['content']
return answer
print(research_with_search('What is the current version of Python?'))Step 3: Add multi-platform research
Extend to search Reddit for opinions and YouTube for tutorials.
Python
def deep_research(topic: str) -> dict:
google = web_search(topic, 'google')
reddit = web_search(topic, 'reddit')
youtube = web_search(topic, 'youtube')
context = f"""Research on: {topic}
Google results:
{google}
Reddit discussions:
{reddit}
YouTube videos:
{youtube}"""
response = ollama.chat(model='llama3.2', messages=[
{'role': 'system', 'content': 'Synthesize a research brief from the sources below. Cite sources.'},
{'role': 'user', 'content': context}
])
return {'topic': topic, 'brief': response['message']['content'], 'credits_used': 3}Python Example
Python
import requests, os, ollama
# Scavio has one endpoint per platform - there is no dispatcher endpoint and no
# `platform` request param, so the selector lives in your code.
SCAVIO = "https://api.scavio.dev"
SCAVIO_ENDPOINTS = {
"google": "/api/v2/google",
"reddit": "/api/v1/reddit/search",
"youtube": "/api/v1/youtube/search",
"amazon": "/api/v1/amazon/search",
"walmart": "/api/v1/walmart/search",
}
SCAVIO_QUERY_KEY = {"youtube": "search"}
SCAVIO_RESULTS_KEY = {"google": "organic_results", "reddit": "results",
"youtube": "results", "amazon": "products", "walmart": "products"}
def scavio_url(platform):
return SCAVIO + SCAVIO_ENDPOINTS[platform or "google"]
def scavio_body(platform, query):
return {SCAVIO_QUERY_KEY.get(platform or "google", "query"): query}
def scavio_payload(payload, platform="google"):
"""Google v2 passes Google's response through as-is; every other endpoint
wraps its payload in `data`. Item fields differ per platform (see
https://scavio.dev/docs), so only the result list is normalised here."""
platform = platform or "google"
out = payload if platform == "google" else payload["data"]
return {**out, "results": out.get(SCAVIO_RESULTS_KEY[platform], [])}
def search(q, platform='google'):
r = scavio_payload(requests.post(scavio_url(platform), headers={'Authorization': 'Bearer ' + os.environ['SCAVIO_API_KEY'], 'Content-Type': 'application/json'}, json=scavio_body(platform, q)).json(), platform)
return '\n'.join(f"- {x['title']}: {x.get('snippet','')}" for x in r.get('results',[])[:5])
def research(q):
ctx = search(q)
return ollama.chat(model='llama3.2', messages=[{'role':'user','content':f'{q}\n\nContext:\n{ctx}'}])['message']['content']JavaScript Example
JavaScript
// Scavio has one endpoint per platform - there is no dispatcher endpoint and no
// `platform` request param, so the selector lives in your code.
const SCAVIO = "https://api.scavio.dev";
const SCAVIO_ENDPOINTS = {
google: "/api/v2/google",
reddit: "/api/v1/reddit/search",
youtube: "/api/v1/youtube/search",
amazon: "/api/v1/amazon/search",
walmart: "/api/v1/walmart/search",
};
const SCAVIO_QUERY_KEY = { youtube: "search" };
const SCAVIO_RESULTS_KEY = { google: "organic_results", reddit: "results",
youtube: "results", amazon: "products", walmart: "products" };
const scavioUrl = (platform) => SCAVIO + SCAVIO_ENDPOINTS[platform || "google"];
const scavioBody = (platform, query) =>
({ [SCAVIO_QUERY_KEY[platform || "google"] || "query"]: query });
// Google v2 passes Google's response through as-is; every other endpoint wraps
// its payload in `data`. Item fields differ per platform (see
// https://scavio.dev/docs), so only the result list is normalised here.
function scavioPayload(json, platform = "google") {
const p = platform || "google";
const out = p === "google" ? json : json.data;
return { ...out, results: out[SCAVIO_RESULTS_KEY[p]] || [] };
}
async function search(query, platform = 'google') {
const r = await fetch(scavioUrl("google"), {
method: 'POST', headers: {'Authorization': `Bearer ${process.env.SCAVIO_API_KEY}`, 'Content-Type': 'application/json'},
body: JSON.stringify(scavioBody("google", query))
});
return (scavioPayload(await r.json(), "google")).organic?.slice(0,5).map(x => `- ${x.title}: ${x.snippet}`).join('\n');
}Expected Output
JSON
A local LLM research stack that can search the live web for current information, combining Ollama's privacy with Scavio's structured search data.