Bravebird Assignment – Multi-Agent PDF Download & Query System

Overview

This project demonstrates a two-agent system that cooperates to:

1. Download a PDF from Google Drive using a browser-based flow (no API usage)
2. Index the downloaded PDF and make it queryable using natural language

The system emphasizes:

• Clean agent handoff
• Explicit context management
• Strong guardrails
• Clear logging and reproducibility

---

Architecture

Agent A – Download Agent

• Role: Downloads a Google Drive PDF using real browser interactions
• Executor: Playwright (Chromium)
• Reasoner: gpt-5-mini
• Key Features:
  • Uses Drive UI controls (no Google Drive API)
  • Retries on failure
  • Enforces a strict 10-minute execution limit
  • Logs every step
  • Stores all artifacts in a sandbox directory
  • Writes a structured handoff artifact with metadata

Agent B – Query Agent

• Role: Indexes the downloaded PDF and answers questions about its content
• Key Features:
  • Reads the file path and metadata from Agent A’s handoff
  • Builds embeddings and a FAISS vector index
  • Answers questions via a CLI interface
  • Uses retrieval-only guardrails to avoid hallucinations

---

Sample Output:

Planner(gpt-5-mini): {'action': 'CLICK_DOWNLOAD_BUTTON', ...} Saved to: sandbox/downloads/Bravebird Assignment.pdf Handoff written to: sandbox/handoff.json Agent B: Document is ready. Ask me anything about it.