Country-aware RAG

Company policies differ by country. Our first RAG searched the whole knowledge base, so it sent one country's overtime policy to people in another country. People were confused by information that didn't apply to them.

The internal documents live in Vertex AI Search, and each document carries its country as metadata. The user's country filters which documents are searched. It comes from the company's own record of each user, which the user can't change. The retrieved passages then ground the model's answer, which comes back through the orchestrator into the same user session. A question about overtime or leave therefore returns the policy that applies to that person.

The documents change too. The policy documents are updated every year, so I keep track of them, properly remove the existing data and feed in the new version, so an old policy never sits next to the current one.

I chose Vertex AI Search because the company is partnered with Google and prefers GCP and Google services. I had thought about Chroma or Qdrant, but I went with Vertex AI Search for that reason, and it does a great job.

Chroma and Qdrant. I went with Vertex AI Search because of the company's partnership with Google.

Designed and implemented a Retrieval-Augmented Generation (RAG) system that allows employees to query organizational knowledge and receive answers grounded in internal company information.

Implemented country-aware retrieval using metadata associated with internal documents, enabling the system to filter relevant knowledge based on the user's country before providing information to the AI model.

This was particularly important where organizational information or policies differed between countries, allowing responses to be grounded in the appropriate country-specific context rather than retrieving indiscriminately from the entire knowledge base.

• Country metadata must be right. The filter only works if every document is set up with the proper country and the setup is done well.
• Yearly updates are a process. When the policies change, the old data has to be removed carefully and the new data fed in.

Answers are grounded in the right country's policy, and people get information that applies to them.