ChatBotGC — AI Assistant for Civil Engineering Standards

Engineers lose time looking for a value or an article in hundreds of pages of PDF standards, and a general-purpose LLM can make answers up. The goal was an assistant that relies only on the company’s documents and whose every answer can be checked.

An end-to-end RAG pipeline: PDF extraction (PyMuPDF), regex cleaning and article-based chunking (about 300 tokens with a 50-token overlap) that keeps the printed page and article numbers; a ChromaDB vector store with multilingual-e5-base embeddings; hybrid retrieval that reranks 30 semantic candidates (65 % semantic, 35 % lexical, bonus for exact numeric matches) and keeps the best 8; generation through a fallback chain of 6 free LLM providers; and a component-based Streamlit interface.

Designed and built the whole system on my own: corpus preparation, vector store, hybrid retrieval, multi-provider generation with fallback, index synchronisation, interface, a 50-question evaluation benchmark with an LLM as judge, Playwright tests, and the technical documentation (developer, user and API guides, LaTeX report).

On a 50-question benchmark graded by an LLM judge, the application never crashed, the average retrieval confidence was 0.792 and no answer was judged incorrect: when the system fails, it declines to answer instead of making something up. Direct article lookups score 4.62/5. Mistral Small answered all 50 questions (3.54/5); Groq Compound scored highest (3.90/5) but answered only 29 because of rate limits — a quality-versus-reliability trade-off that justifies the fallback chain. Next steps: OCR for scanned PDFs, multi-document questions, streaming answers and public deployment.