The Problem#
Most AI coding tools know the current prompt and maybe the open file. They don’t know the rest of the project’s memory: the PDFs, architecture notes, old decisions, spreadsheets and screenshots that explain why the code looks the way it does.
Solo Vault gives the agent that context. Ask “what did we decide about auth?” and it can search the project’s own files.
How It Works#
Files go through an ingestion pipeline before they become searchable:
validate → parse → chunk → embed → store → retrieve
An upload to S3 triggers a queue message, which starts a Step Functions workflow. The results land in PostgreSQL with pgvector, and indexing progress streams back to the IDE over WebSocket. Authentication runs through Cognito, and data is encrypted with customer-managed KMS keys.
The backend is deployed as 12 CloudFormation stacks, driven by a TypeScript orchestration layer.
My Part#
This was a team of four for NYU’s Cloud Computing course. I worked on the Vault backend, not on the IDE itself.
- Built the Step Functions indexing pipeline, the Lambda handlers and the CloudFormation deploy automation
- Wrote the indexer’s unit and integration test suite, plus bulk-indexing and pipeline benchmarking scripts
- Co-authored the systems paper Solo Vault: Cloud-Native Retrieval Infrastructure for the Solo IDE with Sachin Adlakha, Krish Kuchroo and Michael Wang
Demo#
Read More#
- 📝 Blog post: Building the Cloud Backend for Solo Vault, by teammate Sachin Adlakha
- 📄 Paper: Solo Vault: Cloud-Native Retrieval Infrastructure for the Solo IDE (PDF)