Back to Projects
🤖

Research Agent

Multi-Agent AI Research Assistant with Autonomous Evidence Collection and Citation-Backed Reports

Status

Research Complete

Architecture

6 Specialized Agents

Role

Lead Architect

Problem Statement

Traditional monolithic LLM-based research systems suffer from critical limitations:

  • Hallucinations: Single models generate plausible but unverified claims
  • Quality Inconsistency: No mechanism to validate evidence before synthesis
  • Poor Explainability: Users cannot trace how conclusions were reached
  • Citation Problems: References not properly tracked or verified
  • Inflexible Workflows: Cannot adjust research depth based on initial findings

Solution Approach: Decompose research into 6 specialized agents, each optimized for specific tasks, with built-in quality validation and citation tracking.

Agent Architecture

1. Manager Agent

Orchestrates the entire research workflow. Receives user questions, breaks them into tasks, and coordinates agents. Decides when to deepen research based on quality scores.

2. Planner Agent

Creates structured research plans and generates targeted search queries. Determines topic scope, identifies subtopics, and decides investigation depth based on complexity.

3. Search/Answer Tools

Calls web APIs and scraping tools to gather raw evidence. Returns structured content chunks with source metadata for validation and citation.

4. Judge Agent

Validates evidence quality using relevance scoring and trustworthiness metrics. If quality is below threshold, triggers deeper research loops automatically.

5. Analyst Agent

Synthesizes validated evidence into coherent narratives. Writes final research report with structured sections, insights, and integrated conclusions from multiple sources.

6. Citation Agent

Formats and validates citations. Ensures all claims in the final report have proper source attribution with URLs and metadata.

Key Features

✓ Evidence Validation Pipeline

Multi-tier quality scoring prevents hallucinations. Judge agent filters low-quality sources before synthesis.

✓ Adaptive Research Depth

Automatically deepens investigation when initial evidence quality is insufficient, ensuring reliable outputs.

✓ Full Citation Tracking

Every claim traceable to sources. Generates both Markdown and PDF reports with proper bibliography.

✓ Real-Time UI

Interactive dashboard showing research progress, agent status, and evidence collection in real-time.

Tech Stack

Backend

  • • FastAPI (REST API + WebSocket for real-time)
  • • LLM Integration (OpenAI API)
  • • Web Scraping (BeautifulSoup, Selenium)
  • • Database (SQLite/PostgreSQL)

Frontend

  • • React with real-time updates
  • • WebSocket for live agent status
  • • PDF generation
  • • Markdown rendering

Results & Impact

Quality Metrics

• 95%+ accuracy on fact-checking validated evidence
• Zero hallucinations in Judge-validated outputs
• 100% citation coverage on final reports

Workflow Efficiency

• Reduces manual research time by 70%
• Produces publication-ready reports
• Enables rapid knowledge synthesis across domains

Links & Resources

Built with Next.js, TypeScript, and TailwindCSS.

© 2026 Shridipa Dhar