Project 3: A Research Assistant Agent with Tools

Build a LangChain agent that plans research steps, uses tools, and produces a cited summary.

Quick Reference

Attribute Value
Difficulty Level 3: Advanced
Time Estimate 12-20 hours
Language Python or JavaScript
Prerequisites Tool calling, retrieval basics
Key Topics agents, tools, planning, citations

1. Learning Objectives

By completing this project, you will:

  1. Build a tool-using agent that plans steps.
  2. Add web or document retrieval tools.
  3. Enforce citation requirements for claims.
  4. Log and replay agent traces.
  5. Evaluate answer quality with a rubric.

2. Theoretical Foundation

2.1 Agentic Research

Research agents must separate planning from execution and ensure every claim is grounded in evidence.

3. Project Specification

3.1 What You Will Build

An agent that takes a research question, performs multi-step retrieval, and outputs a cited response.

3.2 Functional Requirements

  1. Planner to decompose questions.
  2. Tool set for search and retrieval.
  3. Citation enforcement in output.
  4. Trace log of steps and tool calls.
  5. Evaluation rubric for quality.

3.3 Non-Functional Requirements

  • Safe tool use with allowlists.
  • Deterministic mode for testing.
  • Evidence-first responses.

4. Solution Architecture

4.1 Components

Component Responsibility
Agent Plan and execute steps
Tools Search, read, summarize
Trace Logger Record decisions
Evaluator Score output quality

5. Implementation Guide

5.1 Project Structure

LEARN_LANGCHAIN_PROJECTS/P03-research-assistant/
├── src/
│   ├── agent.py
│   ├── tools.py
│   ├── prompts.py
│   ├── trace.py
│   └── eval.py

5.2 Implementation Phases

Phase 1: Agent + tools (4-6h)

  • Implement agent loop with tool calls.
  • Checkpoint: agent completes a 3-step task.

Phase 2: Citations + trace (4-6h)

  • Add citations in output.
  • Log tool calls and decisions.
  • Checkpoint: trace file is replayable.

Phase 3: Evaluation (4-8h)

  • Add rubric and test questions.
  • Checkpoint: quality scores recorded.

6. Testing Strategy

6.1 Test Categories

Category Purpose Examples
Unit tools tool schemas valid
Integration agent multi-step completion
Regression citations every claim cited

6.2 Critical Test Cases

  1. Missing citation triggers output failure.
  2. Tool failure is logged and recovered.
  3. Agent stops within step budget.

7. Common Pitfalls & Debugging

Pitfall Symptom Fix
Hallucinated claims no sources enforce citations
Infinite loops repeated tools add step limit
Overly broad plan no progress break into tasks

8. Extensions & Challenges

Beginner

  • Add a local document search tool.
  • Add structured report output.

Intermediate

  • Add reranking of sources.
  • Add summary + full evidence sections.

Advanced

  • Add multi-agent debate for research.
  • Add confidence scores per claim.

9. Real-World Connections

  • Research copilots require traceable evidence.
  • Compliance teams need audit trails for conclusions.

10. Resources

  • LangChain agent docs
  • RAG system design guides
  • “AI Engineering” (agent safety)

11. Self-Assessment Checklist

  • I can build a tool-using agent.
  • I can enforce citations in outputs.
  • I can log and replay agent traces.

12. Submission / Completion Criteria

Minimum Completion:

  • Agent executes multi-step research
  • Cited output

Full Completion:

  • Trace logging
  • Evaluation rubric

Excellence:

  • Reranking or multi-agent debate
  • Confidence scoring per claim

This guide was generated from project_based_ideas/AI_AGENTS_LLM_RAG/LEARN_LANGCHAIN_PROJECTS.md.