Research Sentry — Autonomous Real-Time Research Agent

Overview

Research Sentry is a next-generation academic research tool that uses the TinyFish Web Agent to deploy autonomous web agents. These agents "scout" live academic portals—such as ArXiv, PubMed, and IEEE Xplore—to navigate complex search interfaces and extract "Deep Insights" that traditional static APIs often miss.

🖥️ System Preview

1. Goal (Intelligence Lifecycle)

The system acts as an expert agentic researcher. Its goal is to provide real-time, source-backed evidence for specific research hypotheses or broad discovery topics.

🔍 Discovery Process

Phase	Action
INTENT ANALYSIS	Use GPT-4o to map the user's natural language goal (voice or text) into a directed agentic mission.
AGENT DISPATCH	Deploy parallel TinyFish Web Agents to navigate the live DOM of targeted academic portals simultaneously.
DEEP EXTRACTION	Identify specific methodologies, results, and citation counts from unstructured web content.
SYNTHESIS	Normalize raw portal data into a structured Research Profile with "TinyFish Verified" provenance.

📄 Output Schema

{
  "research_context": {
    "topic": "LLM Hallucinations",
    "sources": ["arxiv", "pubmed", "semantic_scholar"],
    "intent": "Discovery"
  },
  "discovery_summary": {
    "yield": 20,
    "primary_themes": ["Mitigation", "Detection Pipelines", "Grounding"],
    "confidence": 0.94
  },
  "papers_discovered": [
    {
      "title": "Mitigating Hallucinations in Large Language Models",
      "source": "arxiv",
      "evidence": "Wait-time analysis on reasoning chains.",
      "url": "https://arxiv.org/abs/..."
    }
  ],
  "recommended_action": "Actionable Insight: Focus on 'Self-Correction' papers for immediate technical implementation."
}

---

2. Code Snippets

cURL

curl -X POST "http://localhost:3000/api/search/text" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "Recent advances in Transformer efficiency",
    "sources": ["arxiv", "semantic_scholar"]
  }'

TypeScript (Frontend Component)

const response = await fetch("/api/search/text", {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({
    query: "Quantum error correction in silicon",
    sources: ["ieee", "arxiv"]
  }),
});

const data = await response.json();

if (data.papers) {
  setResults(data);
  toast({ title: "Discovery Complete", description: "20 papers synthesized." });
}

Python (SDK Usage)

import requests

def initiate_discovery(topic, sources):
    url = "http://localhost:3000/api/search/text"
    payload = {"query": topic, "sources": sources}
    
    response = requests.post(url, json=payload)
    data = response.json()
    
    print(f"Total Papers Found for {topic}: {data['totalFound']}")
    print(f"Top Result: {data['papers'][0]['title']}")

initiate_discovery("LLM Hallucinations", ["arxiv", "pubmed"])

---

3. Operational Signals

As the agents browse, the TinyFish Live Terminal displays real-time status updates:

TinyFish Scout: Analyzing ArXiv Search Portal - Results stabilizing...
TinyFish Scout: Reading Semantic Scholar - Extracting direct PDF links...
TinyFish Scout: Navigating IEEE Xplore - Capturing DOI references...
Intelligence Engine: Aggregating 20 signals for "LLM Hallucinations"...
System Activity: Discovery yield complete. Generating Research Summary...

---

🛠 Key Features

Feature	Description
Parallel Discovery	Launches multiple agents simultaneously to analyze different academic portals in seconds.
Deep Content Scouring	Agents extract quantitative data (citations, dates) rather than just surface-level titles.
TinyFish Live Terminal	A high-fidelity UI component showing the agent's real-time reasoning and navigation steps.
Multimodal Intent	Seamlessly switch between detailed text prompts or direct voice-activated discovery.

---

🏗 Architecture Overview

Component Relationships

Component	Role	Logic
Home (page.tsx)	Core Dashboard managing input state and global agent orchestration.	app/page.tsx
SearchInterface	Multimodal input layer for text prompts and voice recording.	components/SearchInterface.tsx
TinyFishAgentTerminal	Visual terminal animation showing live scouting telemetry.	components/TinyFishAgentTerminal.tsx
Search Engine	Handles agentic scouting and hybrid API fallback execution.	lib/search.ts
TinyFish Core	Underlying autonomous browser navigation engine.	lib/mino.ts

System Architecture

graph TD
    User((User)) -->|Voice/Text| UI[Search Interface]
    UI -->|Intent| Parser[Intent Parser GPT-4]
    Parser -->|Plan| Engine[Search Engine]
    Engine -->|Dispatch| Agent1[TinyFish Agent: ArXiv]
    Engine -->|Dispatch| Agent2[TinyFish Agent: PubMed]
    Engine -->|Dispatch| Agent3[TinyFish Agent: Scholar]
    Agent1 -->|Scraping| Web[Live Web DOM]
    Agent2 -->|Scraping| Web
    Agent3 -->|Scraping| Web
    Web -->|Result| Aggregator[Synthesis & Deduplication]
    Aggregator -->|JSON Payload| UI
    UI -->|Visuals| Terminal[Live Log Terminal]

Loading

---

🔄 Data Flow

1. Context Entry → User provides a research goal (e.g., "Find Transformer papers from 2024").
2. Intent Analysis → GPT-4o parses the goal into specific keywords and agentic instructions.
3. Agent Dispatch → Parallel TinyFish Web Agents navigate to the target URLs to extract metrics.
4. Autonomous Reasoning → Agents read the DOM, handle pagination, and identify relevant signals.
5. Synthesis → The Search Engine deduplicates results and maps them to a unified schema.
6. Insight Output → A comprehensive discovery grid is returned with verified links and AI summaries.

---

💡 System Vision

Build the ultimate autonomous research monitor:

• Browser-Level Intelligence: Deploy TinyFish agents to browse official portals for live verification.
• Radical Transparency: Show a transparent "Scouting Map" of the agent's steps to build trust.
• Deep Metrics: Extract source-backed data (not just LLM hallucinations) to detect breakthroughs early.
• Actionable Synthesis: Transform unstructured web data into structured, ready-to-use research profiles.