The Intelligent Web Information Retrieval Agent is an autonomous, accuracy-first AI system built for AMD AI League Match 4. Given a starting website URL and a natural-language query, the agent autonomously navigates complex multi-page websites, locates factual answers, and returns verbatim supporting proof with exact source attribution.
Unlike conventional scrapers or conversational agents prone to hallucination, this system prioritizes factual fidelity through an innovative Dual-Stage Guardrail Architecture:
Deterministic Code Check: A Python-level verbatim algorithm verifies that the extracted evidence quote appears character-for-character within the fetched web page, automatically rejecting fabricated quotes.
Entailment Verification Pass: A dedicated reasoning audit ensures the quote directly answers the query before accepting.
Zero-Hallucination Fallback: If information is absent or unverifiable, the agent reliably outputs 'found: false' rather than guessing, achieving a 0.0% hallucination rate across impossible test queries.
Key Technical Innovations:
Dynamic Search Form Discovery: Automatically discovers HTML forms, query fields, and dropdown selectors with full 302 and meta-refresh redirect traversal—handling complex sites like Wikipedia without site-specific heuristics.
BM25Okapi Efficiency: Semantically ranks anchor texts and page chunks to minimize context window consumption and latency.
Enterprise Security: Enforces RFC1918 SSRF protection, registered-domain confinement, and robots.txt compliance.
AMD ROCm Acceleration: Engineered for seamless deployment on AMD Instinct GPUs (MI210, MI250, MI300X) and Radeon hardware via vLLM with zero code changes.
The application features a responsive dashboard with a live visual navigation graph and an in-browser batch benchmark runner, achieving 96.6% accuracy on evaluation benchmarks.
.png&w=256&q=75)
.png&w=640&q=75)