Assay Architecture

8. Operational Considerations

Error Handling & Fault Tolerance

Retry Logic: Pub/Sub automatically retries failed messages with exponential backoff. Workers are designed to be idempotent, ensuring safe re-execution.

Error States: Failed processing updates Firestore status to failed with error details. Users can see error messages in the UI and retry if needed.

Monitoring: All workers log structured events with document IDs, user IDs, and processing metadata. Errors are logged with full context for debugging.

Observability

Logging: Structured logging using Google Cloud Logging with:

• Document ID and user ID in every log entry
• Processing stage and progress percentage
• Error details with stack traces
• Performance metrics (processing time, token counts)

Monitoring: Cloud Functions provide built-in metrics:

• Invocation count
• Execution time
• Error rate
• Memory usage

Alerting: Can be configured for error rates, processing failures, or system health issues.

Cost Optimization

Serverless Scaling: Functions scale to zero during quiet periods, eliminating idle costs.

Summary-Only Storage: PDFs deleted after processing, only AI-generated summaries retained (no extracted text kept), reducing storage costs by 10-20x.

Efficient Queries: Firestore indexes optimized for common query patterns, reducing read costs.

Caching: Frontend uses static generation and CDN caching to minimize database reads.

Security

Authentication: Firebase Authentication for user management, API keys for programmatic access.

Authorization: Document-level access control (users can only access their own documents or public documents).

Data Privacy: Original PDFs deleted after processing, only summaries retained (no extracted text from PDFs is kept).

API Security: API keys use hashed secrets, header-only authentication, time-limited expiration.

User Tiers

Anonymous: Can browse public documents only, no uploads.

Member (Registered): Can upload documents, access private collection, use MCP integration.

Premium (Upgraded): Same as Member, plus access to Premium quality processing (Gemini 2.5 Pro).

Admin: Full system access for moderation and management.

Document Visibility

Public: Document summaries are visible to all users, enabling community discovery.

Private: Document summaries are visible only to the uploader.

Visibility Selection: Chosen at upload time, cannot be changed after upload (privacy and data integrity).

9. Conclusion

Assay's architecture demonstrates how event-driven, serverless design can create a scalable, maintainable, and user-friendly document intelligence system. By decoupling processing stages, enabling parallel execution, and providing real-time updates, we've built a system that scales automatically while delivering intelligent insights to users.

Future Enhancements:

• Additional processing stages (e.g., citation extraction, figure analysis)
• Enhanced similarity algorithms (e.g., semantic embeddings)
• Advanced search capabilities (e.g., full-text search, semantic search)
• Multi-language support
• Batch processing for large document sets

The architecture continues to evolve as we add new capabilities, but the core principles—event-driven design, serverless execution, and real-time updates—remain the foundation for everything we build.