Hermes 3 web search returns irrelevant results because the model reformulates your query before searching, often stripping critical keywords or adding hallucinated context. The fix is routing search through an API you control, where the exact query hits the search engine without LLM reformulation.
Root Cause: Query Reformulation
When Hermes 3 uses its built-in web search tool, it rewrites your query to what it thinks is a better search. "Best search API for agents 2026" might become "search API comparison" -- losing the year filter and the agent context. The reformulated query returns broader, less relevant results.
This is not a bug; it is how tool-use models work. The LLM decides what to search for. But when your pipeline depends on specific query terms, reformulation breaks result quality.
Fix 1: Bypass LLM Query Reformulation
Call the search API directly with the exact user query. Do not let the LLM choose what to search. Feed the raw results back into Hermes as context.
import requests, os
H = {"x-api-key": os.environ["SCAVIO_API_KEY"]}
def search_then_reason(user_query):
"""Search with exact query, then reason with Hermes."""
# Step 1: Exact query search -- no LLM reformulation
r = requests.post("https://api.scavio.dev/api/v1/search",
headers=H,
json={"platform": "google", "query": user_query},
timeout=10
).json()
context = "\\n".join([
f"- {item['title']}: {item.get('snippet', '')}"
for item in r.get("organic", [])[:5]
])
# Step 2: Feed exact results to Hermes for reasoning
prompt = f"""Search results for "{user_query}":
{context}
Based ONLY on these search results, answer the user's question.
Do not add information not present in the results."""
# Send to Hermes 3 via Ollama
hermes_r = requests.post("http://localhost:11434/api/generate",
json={
"model": "hermes3",
"prompt": prompt,
"stream": False,
},
timeout=30
).json()
return hermes_r.get("response", "")
answer = search_then_reason("best search api for agents 2026")
print(answer)Fix 2: Multi-Source Grounding
Hermes's built-in search uses a single source. Cross-reference with multiple platforms to improve result quality. If Google and Reddit agree on an answer, it is more likely correct than a single-source result.
def multi_source_search(query):
"""Search Google and Reddit for cross-referencing."""
sources = {}
for platform in ["google", "reddit"]:
r = requests.post("https://api.scavio.dev/api/v1/search",
headers=H,
json={"platform": platform, "query": query},
timeout=10
).json()
sources[platform] = [
{"title": item["title"], "snippet": item.get("snippet", "")}
for item in r.get("organic", [])[:3]
]
return sources
def grounded_answer(query):
"""Answer with multi-source grounding."""
sources = multi_source_search(query)
context = "Google results:\\n"
for r in sources.get("google", []):
context += f"- {r['title']}: {r['snippet']}\\n"
context += "\\nReddit discussions:\\n"
for r in sources.get("reddit", []):
context += f"- {r['title']}: {r['snippet']}\\n"
# Hermes reasons over multi-source data
return context # pass to Hermes
result = grounded_answer("hermes 3 web search quality issues")
print(result)Fix 3: Query Template Enforcement
If you must let Hermes reformulate queries, constrain the reformulation with a template. "Search for: [original query] [current year]" prevents the model from stripping year filters.
Why Built-In Search Breaks
Hermes's web search tool is typically backed by DuckDuckGo or a similar free search engine. These have smaller indexes than Google and return weaker results on niche queries. The combination of query reformulation plus a weaker search engine compounds the quality problem.
Quality Metrics to Track
Compare answers from Hermes's built-in search vs API-grounded answers on the same queries. Track: answer relevance (does it address the actual question), factual accuracy (can claims be verified), and freshness (does it reference current data). In testing, API-grounded answers score 30-40% higher on relevance for technical queries.