https://jerma88.github.io/tags/vllm/VLLMHugo Blox Builder (https://hugoblox.com)en-usTue, 25 Mar 2025 00:00:00 +0000https://jerma88.github.io/media/icon_hu_982c5d63a71b2961.pnghttps://jerma88.github.io/tags/vllm/https://jerma88.github.io/project/course-recommender-system/Tue, 25 Mar 2025 00:00:00 +0000https://jerma88.github.io/project/course-recommender-system/<h1 id="agentic-graphrag-recommender-system">Agentic GraphRAG Recommender System</h1> <h2 id="overview">Overview</h2> <p>Published at the <strong>AAAI 2025 Spring Symposium (Human-Compatible AI)</strong> and supported by the <strong>Robert Mayer Undergraduate Research Fellowship</strong>, this project develops an autonomous multi-agent advising architecture designed to solve complex combinatorial university degree planning challenges. By fusing <strong>Knowledge Graph representations (Neo4j)</strong> with <strong>dense semantic vector retrieval (Qdrant)</strong> and <strong>vLLM continuous batching</strong>, the system eliminates prerequisite hallucinations and slashes advising turnaround times.</p> <div class="mermaid">flowchart TD Student[Student Preferences & Transcript] --> Agent[Orchestration Agent\nLangChain / Tool Calling] Agent <-->|Graph Traversal\nPrerequisites & Constraints| Neo4j[(Neo4j Knowledge Graph)] Agent <-->|Dense Semantic Search\nCourse Descriptions| Qdrant[(Qdrant Vector DB)] Agent <-->|Fast High-Throughput Inference\nContinuous Batching| vLLM[vLLM Inference Engine] Agent --> HITL[Human-in-the-Loop (HITL)\nAdvisor Verification Portal] HITL --> Plan[Optimized Graduation Plan] </div> <h2 id="system-highlights--innovations">System Highlights &amp; Innovations</h2> <h3 id="1-hybrid-graphrag-architecture">1. Hybrid GraphRAG Architecture</h3> <ul> <li>Modeled complex university curriculum catalogs, prerequisite DAGs, core requirements, and major/minor electives into <strong>Neo4j</strong>.</li> <li>Leveraged <strong>Qdrant</strong> dense vector indexing to semantically align student interests, career trajectories, and course descriptions.</li> <li>Hybrid retrieval prevents standard LLM hallucinations regarding course prerequisites and degree credit requirements.</li> </ul> <h3 id="2-high-throughput-vllm-inference-engine">2. High-Throughput vLLM Inference Engine</h3> <ul> <li>Deployed self-hosted <strong>vLLM</strong> inference endpoints utilizing continuous batching and PagedAttention to serve multiple concurrent student sessions.</li> <li>Supported 2,000+ concurrent LLM advising interactions with sub-second time-to-first-token (TTFT).</li> </ul> <h3 id="3-human-in-the-loop-hitl-verification">3. Human-in-the-Loop (HITL) Verification</h3> <ul> <li>Incorporated human-in-the-loop validation checkpoints where academic advisors can review, amend, and approve generated degree plans with 1-click workflows.</li> <li>Reduced student advising turnaround times by <strong>50%</strong> across trial departments.</li> </ul> <h2 id="publication">Publication</h2> <ul> <li><strong>A Recommender System Architecture for University Curriculum Advising</strong><br> Zerui Ma, Michael Hahsler, Peter Moore.<br> <em>Proceedings of the AAAI Symposium Series</em>, 5(1), 235–241 (2025).<br> DOI: .</li> </ul>