Case study · RAG and agents over enterprise documents
AI Business Operations Copilot
An enterprise RAG platform with AI agents that lets teams question their own documents in plain language, and automates document analysis and reporting.
- less time spent searching documents
- 80%+
- file types ingested: PDF, DOCX, Excel and CSV
- 4
- model providers: OpenAI, Claude and Gemini
- 3
The problem
Enterprise knowledge sits in PDFs, Word files and spreadsheets. Finding one answer means opening many files, and turning them into reports is slow, manual work.
How it works
- 1
Ingest documents
ingest_docs()
PDF, DOCX, Excel and CSV files enter through dedicated ingestion pipelines.
- 2
Index for meaning
index()
Content is embedded and stored in ChromaDB for semantic search.
- 3
Retrieve
retrieve()
A plain-language question pulls back the most relevant passages.
- 4
Agentic workflows
agents()
LangGraph agents automate document analysis and reporting.
- 5
Serve securely
serve()
FastAPI APIs expose the AI services to a Next.js workspace.
Engineering decisions
Ground answers in the company's own documents
Retrieval-Augmented Generation means answers come from the organisation's documents, not from the model's general knowledge.
Ingest the formats people actually use
Separate ingestion paths for PDF, DOCX, Excel and CSV, so spreadsheets are as searchable as reports.
Agents for the multi-step work
LangGraph agentic workflows take on analysis and reporting that would otherwise take several manual steps.
Secure API boundary
All AI features are served through secure FastAPI APIs, with PostgreSQL and Redis behind them and Docker for deployment.
What I built
- Designed an enterprise-grade RAG architecture, with AI agents and agentic workflows that automate business operations.
- Semantic search over company documents with ChromaDB.
- Ingestion pipelines for PDF, DOCX, Excel and CSV files.
- Secure FastAPI APIs serving the AI features.
Outcome
Document search time dropped by more than 80%. Document analysis and reporting workflows are automated, and teams can query their knowledge repositories in natural language.