RAG assistant for shipping and logistics documents

Deliverables

• Google Drive ingestion with automatic download
• Document parsing to structured JSON with validation
• Text splitting and embeddings
• Vector store setup on Supabase and Qdrant
• Chat agent with memory and a vector store search tool

The challenge

The answers staff needed were buried in PDFs such as shipping documents, and finding them meant opening files one by one. A first version of the retrieval system worked on standard documents but became unreliable when a PDF had an unusual layout. The model was not the problem: the text going into the vector store was not consistently structured, so retrieval returned the wrong passages.

Our approach

A Google Drive trigger watches for new files, downloads each one and passes it through a data loader and text splitter. Each chunk is embedded with OpenAI or Google Gemini embeddings and written to the vector store. On the question side, a chat trigger feeds an AI agent that has a chat model, short-term memory and a vector store tool, so it pulls the most relevant passages before it answers. To fix the unreliable documents, we added a dedicated parsing layer in front of the store: documents are converted to structured JSON, the extracted information is checked before it moves on, and only validated content is indexed. Document processing and retrieval are kept as separate stages, which makes each one easier to test.