🧠 Build a Resilient AI Research Agent with Gemini + LangChain (No APIs? No Problem.)
In this guide, you'll build a research super-agent that:
✅ Performs deep research on any topic
✅ Handles web/API failures gracefully
✅ Enhances its output with verifiable claims
✅ Generates a full professional report
✅ Uses Google Gemini, DuckDuckGo, Wikipedia, and optionally SerpAPI
🤔 Why This Matters
Most AI research agents work like this:
- Take a topic
- Query an LLM
- Dump the result
What’s wrong with that?
- They fail when web search APIs are down
- They hallucinate facts
- They don’t verify anything
- They can’t format output beyond a paragraph dump
This agent fixes all of that by combining multiple tools, smart rate-limiting, and multi-phase verification.
Let’s build it, step by step.
⚙️ Step 1: Install Required Tools
We need LangChain to connect LLMs and tools, Gemini for powerful reasoning, and dotenv for environment configs.
pip install langchain-google-genai python-dotenv requests
🧾 Step 2: Set Environment Variables
Create a .env file in your project directory with your API keys:
GOOGLE_API_KEY=your_google_api_key_here
SERPAPI_KEY=your_serpapi_key_here # optional
❗ You can still run the system without SerpAPI. DuckDuckGo and Wikipedia are enough for many topics.
🔍 Step 3: Create a Multi-Source Search System
We use DuckDuckGoSearchRun via LangChain for fast, no-auth search. If that fails, we fallback to Wikipedia. If SerpAPI is present, we use that too.
import time, os, requests
from langchain_community.tools import DuckDuckGoSearchRun
from dotenv import load_dotenv
load_dotenv()
class ImprovedSearchSystem:
def __init__(self):
self.search = DuckDuckGoSearchRun()
self.search_delay = 3.0
self.last_search_time = 0
def safe_web_search(self, query):
now = time.time()
if now - self.last_search_time < self.search_delay:
time.sleep(self.search_delay - (now - self.last_search_time))
self.last_search_time = time.time()
for name, method in [
("DuckDuckGo", self._try_duckduckgo),
("Wikipedia", self._try_wikipedia),
("SerpAPI", self._try_serpapi if os.getenv("SERPAPI_KEY") else None)
]:
if method is None: continue
try:
result = method(query)
if result and len(result.strip()) > 50:
return f"[{name}] {result}"
except Exception:
continue
return "Search unavailable or no results."
def _try_duckduckgo(self, query):
return self.search.invoke(query)
def _try_wikipedia(self, query):
q = "_".join(query.split()[:3])
r = requests.get(f"https://en.wikipedia.org/api/rest_v1/page/summary/{q}")
if r.status_code == 200:
return r.json().get("extract", "")
return ""
def _try_serpapi(self, query):
params = {"q": query, "api_key": os.getenv("SERPAPI_KEY")}
r = requests.get("https://serpapi.com/search", params=params)
results = r.json().get("organic_results", [])
return "\n".join(f"{r['title']}: {r.get('snippet', '')}" for r in results[:2])
✅ What this does:
- Avoids hitting the same search source every time
- Skips broken sources
- Doesn’t block if internet is slow
🧠 Step 4: Add Google Gemini via LangChain
We'll use Gemini to do the heavy lifting: writing, verifying, summarizing, polishing.
from langchain_google_genai import ChatGoogleGenerativeAI
class RobustResearchOrchestrator:
def __init__(self):
self.llm = ChatGoogleGenerativeAI(
model="gemini-1.5-flash-latest",
temperature=0.7,
google_api_key=os.getenv("GOOGLE_API_KEY")
)
self.search = ImprovedSearchSystem()
self.llm_delay = 2.0
self.last_llm_time = 0
def _rate_limit_llm(self):
now = time.time()
if now - self.last_llm_time < self.llm_delay:
time.sleep(self.llm_delay - (now - self.last_llm_time))
self.last_llm_time = time.time()
📚 Step 5: Research with Web Context
def research_with_llm_knowledge(self, topic, web_context=""):
self._rate_limit_llm()
prompt = f"""
Research the following topic thoroughly:
Topic: {topic}
{f"Web context: {web_context}" if web_context else ""}
Include:
1. Definitions
2. 2024-2025 breakthroughs
3. Key players
4. Challenges
5. Future outlook
6. Technical details
"""
return self.llm.invoke(prompt).content
✅ Why this matters:
You’re not just asking the LLM for trivia—you’re giving it timely web data as context.
🔎 Step 6: Enhance with Web-Verified Claims
def enhance_with_verification(self, content, topic):
self._rate_limit_llm()
extract_prompt = f"""
Extract 2 verifiable claims from this research on {topic}:
{content[:2000]}
"""
claims = self.llm.invoke(extract_prompt).content.splitlines()
verified = []
for claim in claims[:2]:
search_result = self.search.safe_web_search(claim)
if search_result and len(search_result) > 100:
verified.append(f"{claim}\nVerified: {search_result}")
if not verified:
return content
enhance_prompt = f"""
Update this research with the following verified information:
Original:
{content}
Verified Data:
{chr(10).join(verified)}
Mark updated lines with [VERIFIED].
"""
return self.llm.invoke(enhance_prompt).content
✅ Why this matters:
You extract and fact-check your own LLM output. Self-auditing makes your system more trustworthy.
📄 Step 7: Format as a Full Report
def create_report(self, enhanced, topic):
self._rate_limit_llm()
report_prompt = f"""
Convert the following into a full technical report on '{topic}' with:
- Executive Summary
- Introduction
- Breakthroughs
- Key Players
- Technical Details
- Challenges
- Future Outlook
- Conclusion
Content:
{enhanced}
"""
return self.llm.invoke(report_prompt).content
💅 Step 8: Polish for Professional Quality
def polish_report(self, report):
self._rate_limit_llm()
return self.llm.invoke(f"Polish this technical report:\n\n{report}").content
🧪 Step 9: Run the Entire Pipeline
def run_pipeline(self, topic):
context = self.search.safe_web_search(f"{topic} 2024 update")
content = self.research_with_llm_knowledge(topic, context)
enhanced = self.enhance_with_verification(content, topic)
report = self.create_report(enhanced, topic)
final = self.polish_report(report)
return final
🧯 Step 10: Command Line Runner
def run_robust_research(topic):
orchestrator = RobustResearchOrchestrator()
return orchestrator.run_pipeline(topic)
if __name__ == "__main__":
topic = input("Enter a research topic: ").strip()
print(run_robust_research(topic))
🧠 Example: Output for "AI Hardware in 2025"
Executive Summary:
AI hardware has rapidly evolved...
Breakthroughs:
- NVIDIA’s Blackwell GPU architecture...
- Intel’s neuromorphic Loihi chip...
Key Players:
- NVIDIA, Intel, Graphcore, Cerebras...
Challenges:
- Energy efficiency, heat dissipation...
[VERIFIED] Blackwell GPUs announced March 2025 with 4x efficiency...
🔚 Final Thoughts
This agent is:
✅ Self-healing (fallbacks and retries)
✅ Fact-aware (claim verification)
✅ Pro-quality (full report structure)
✅ Flexible (extendable to PDF, Slack, APIs)
You now have a bulletproof research pipeline that works even if APIs don’t.
Top comments (0)