Back to Projects

MARA - Multi-Agent AI Research Assistant

LangGraph-based multi-agent system that classifies intent, orchestrates parallel retrieval, validates citations, and generates structured research reports with human-in-the-loop review.

GitHub Live Demo
Latency reduction>4min → <60s
Citation validationURL-verified sources
Retrieval sourcesWeb + Academic parallel

Overview

MARA is a LangGraph-based multi-agent research system that classifies query intent, orchestrates parallel academic and web retrieval, validates citations, and generates structured markdown research reports with human-in-the-loop review before final output.

System Demonstration

Architecture

Orchestration Design

1

Intent Classification

A zero-shot classifier routes queries at the entry point - separating casual inputs from research queries before any tool is invoked, reducing unnecessary latency and cost.

2

Modular Agentic Pipeline

A four-node LangGraph workflow (Planner → Researcher → Writer → Human Review) with clearly separated responsibilities per node.

3

Parallel Retrieval

Web search (Tavily) and academic search (Arxiv) execute concurrently, cutting retrieval time relative to sequential execution.

4

Citation Grounding

Source URLs are strictly validated and parsed at the writer stage, hallucinated or malformed references are filtered before the report is assembled.

5

Human-in-the-Loop Review

An explicit approval node allows a human reviewer to request revisions before the final output, improving reliability for research-oriented outputs.

Key Learnings

  • Task-specific agent separation improves orchestration reliability: Splitting planning, retrieval, synthesis, and review into dedicated LangGraph nodes produced more deterministic tool usage and more stable report generation than monolithic agent prompts.
  • Context ranking materially improves synthesis quality: Filtering retrieval results before generation reduced "lost-in-the-middle" degradation and reduced execution latency from >4 minutes to under 60 seconds.
  • Structured validation must be integrated into orchestration design: Pydantic-enforced outputs and URL validation significantly reduced malformed citations and improved report consistency.
  • Intent routing reduces unnecessary orchestration cost: Lightweight query classification prevented non-research requests from triggering the full agentic pipeline.
  • Model selection had larger impact than prompt tuning: Migrating synthesis workloads to DeepSeek-V3 produced larger latency improvements than iterative prompt optimization.

Tech Stack

Frontend

ReactTypeScriptTailwind CSS

Orchestration

LangGraphLangChainHuman-in-the-Loop

LLMs

GroqOpenRouter

Retrieval

Tavily Web SearchArxiv API

Validation

PydanticCitation Grounding

Want to Work Together?

Need a research automation pipeline, intelligent document synthesis, or multi-source retrieval system? Let's explore how agentic workflows can accelerate your research or product.

✉ Get in Touch
Previous ProjectHierarchical Multi-Agent Decision System for Wargaming
Next Project Knowledge Graph RAG for Aerospace Technical Documentation