Onboard AI - Employee Onboarding Assistant

ACME Onboard — AI-Powered Enterprise Onboarding Platform Problem Statement: Onboarding knowledge in most companies is scattered across HR policies, IT setup guides, security rules, and engineering handbooks. New employees often struggle to identify the right documents, determine which department owns the information, and verify whether it is current. Meanwhile, HR and IT teams spend considerable time answering repetitive questions. Our Solution: ACME Onboard centralizes enterprise onboarding knowledge through two integrated components: an HR document management portal and an AI-powered employee knowledge assistant. Built around ACME Corp, a fictional enterprise software company with 13 departments, the platform uses LLM-assisted document processing, semantic search, intelligent department routing, and Retrieval-Augmented Generation (RAG) to deliver relevant, citation-backed answers. HR Document Management: The HR portal allows administrators to select a department and upload onboarding documents in PDF, DOCX, Markdown, or TXT format. The Next.js frontend supports uploads in batches of 10, displays real-time progress, and provides individual file results. The FastAPI backend extracts document structure, including headings, lists, tables, and page information. DeepSeek V4.1 Flash generates grounded metadata and identifies semantic sections, which are validated against the original document to prevent missing, duplicated, or invented content. Documents are then divided into hierarchical, contextualized chunks and embedded using Qwen3 Embedding 8B. The resulting embeddings, document text, and metadata are stored in a persistent ChromaDB collection under the department selected by HR. The ingestion pipeline also supports idempotent uploads, skipping unchanged documents and replacing previous versions when files are modified. AI-Powered Employee Assistant: The employee assistant uses a department-aware RAG pipeline to answer questions using company onboarding documents. When an employee submits a question, the backend first resolves any contextual follow-up and generates a query embedding. A trained XGBoost classifier combines its predictions with evidence from Qwen3 Reranker 8B, using a 20/80 weighted fusion strategy to identify the relevant departments. DeepSeek V4.1 Flash then generates alternative search queries and a hypothetical document using HyDE. Both operations execute concurrently to improve retrieval coverage. The hybrid retrieval pipeline combines dense vector search in ChromaDB with BM25 keyword search and departmental metadata filtering. Results are merged using weighted Reciprocal Rank Fusion (RRF), filtered for relevance and evidence quality, and diversified using Maximum Marginal Relevance (MMR) to reduce redundancy. Finally, GLM 5.3 Flash generates a grounded Markdown answer using up to 20 retrieved source chunks. Responses are streamed through Server-Sent Events (SSE), with citations validated against the retrieved documents. When sufficient evidence is unavailable, the assistant explicitly communicates that limitation rather than inventing an answer. Conversations are persisted as JSON files and support history retrieval, renaming, and deletion.