Office staff want AI help with everyday work, but pasting confidential company data into public chatbots is a privacy and compliance risk. Office Assistant is an internal AI agent that answers questions from a company's own documents, SharePoint and email, while keeping the data under the company's control.
What it does
Answers questions from uploaded PDF, DOCX, XLSX and CSV files, with citations to the source
Uses hybrid search (keyword plus vector search with pgvector) to find the right passages
Runs a tool-calling agent that decides whether to search, call a finance tool or answer directly
Drafts email replies that a person approves before anything is sent
What makes it different
Role-based access control down to the section level. One document can have general, manager and finance sections. Each user sees only what their role allows, and the filtering happens before the model sees anything, so a prompt can't get around it.
Automatic sensitivity classification. When a document is uploaded, an LLM suggests a sensitivity level for each unmarked section and an admin approves it.
Conflict detection. When two documents disagree, for example on how many days you can work from home, users see both claims and their sources until an admin decides which one is correct.
Full audit trail. Every query, retrieval and tool call is logged.
Tech stack
Vue 3, FastAPI, PostgreSQL with pgvector, and a hand-written agent loop with no LangChain. It works with Anthropic, OpenRouter or local Ollama models, and it uses self-hosted Langfuse for observability. The whole stack can run on the company's own servers.
Evaluation
The evaluation suite checks answer accuracy, access-control leaks and conflict handling. It uses test documents that mix sensitivity levels and contradict each other.
