APPC Archive Assistant

The APPC Archive Assistant helps Gettysburg College faculty search and analyze decades of Academic Policy and Program Committee minutes. It is a Retrieval-Augmented Generation pipeline that drastically reduces the administrative burden of researching historical academic policy. (The live demo is available to Gettysburg College faculty and students.)

Built with two teammates.

Highlights

• Self-querying RAG: LlamaIndex, Pydantic, and pgvector give faculty natural-language access to 500+ PDFs, with adaptive course/date filtering applied before vector search.
• Batch ingestion: a horizontally scalable Django/Celery pipeline with Docling OCR and LLM course extraction processed 500+ PDFs into pgvector in ~4 hours, replacing weeks of manual archival work.
• Fault tolerance: a 4-model OpenRouter fallback chain and service-level isolation (RAG, query parsing, cost tracking) for 99.9% uptime and <2s latency.
• Cheap to run: a multi-model prompting strategy cut inference costs 90%; the whole project cost under $1 in API spend.
• Multi-turn chat: an LLM query rewriter injects conversation history to resolve pronouns and temporal references.
• Verifiable answers: a React chat UI with hover-linked source citations back to the original PDF chunks.