A searchable memory of government resolutions
Retrieval over state government circulars and resolutions, built in two editions from one codebase: one on a cloud model API, one fully on-premise with a switch between GPU and CPU.
The situation
Departments issue a steady stream of circulars and Government Resolutions. The answer to a simple administrative question can be spread across years of formal PDFs in Marathi and English.
Constraints we designed around
- Circulars are long, formal and bilingual.
- Some offices can use cloud APIs. Many cannot.
- Not every office has a GPU, and the system has to run anyway.
What we built
Two editions, one codebase
An API edition for fast rollout, and an on-premise edition for offices where documents must stay inside the network.
GPU or CPU by configuration
A single setting moves inference between GPU and CPU, so the same system runs on a server or an ordinary workstation.
Answers that point to the circular
Each answer links back to the resolution and passage it came from, so staff can quote the source directly.
Deployment
Cloud API edition, or on-premise with GPU or CPU inference.
Outcomes
- deployment editions from one codebase
- 2
- deployment editions from one codebase
- setting to switch GPU and CPU
- 1
- setting to switch GPU and CPU
- languages: Marathi and English
- 2
- languages: Marathi and English
Demonstrated to the state government.
Stack
- RAG pipeline
- Vector search
- Open-weight LLM
- Hosted LLM API
- Docker