Mastra
Look for workflow mode adapter and the five typed agents in this page.
Five agents, one source-grounded workflow.
SchemeSeva is a TypeScript-native civic AI app built on Mastra-style orchestration, Qdrant retrieval and memory, Gemini embeddings, Featherless AI reasoning, OpenAI/OpenRouter fallback, Enkrypt validation, Langfuse tracing, and Upstash rate limiting. The core report path is simple: Qdrant retrieves, Featherless reasons, and Enkrypt validates before display.
Judge verification
How to verify the mandatory stack
Qdrant retrieval
Run the Farmer demo and confirm Retrieval: qdrant-vector.
Featherless AI
Confirm Reasoning: featherless when Featherless is configured.
Qdrant memory
Confirm Memory: qdrant and Memory write: success after a report.
Enkrypt AI
Confirm Safety: enkrypt on reports and Vigilance alerts.
Collections
schemeseva_schemes, citizen_sessions, and pending_alerts.
Debug proof
Open /debug/integrations for live provider status without secrets.
Pipeline
How a request moves through the system
Profile
Structure citizen details
Discovery
Retrieve candidate schemes
Eligibility
Check rules
Report
Explain likely matches
Safety
Validate output
Vigilance
Watch for new matches
Stack
The production proof points
Frontend
TanStack Start (React 19) + Tailwind v4
Landing page, agent runner, report view, scheme catalog, and alert banner.
Orchestration
Mastra-style TypeScript workflow
Chains five agents through typed server functions that are testable in isolation.
Reasoning
Featherless AI + OpenAI/OpenRouter fallback
Featherless powers open-source model reasoning for reports and Vigilance alert reasons; OpenAI, OpenRouter, and local grounded logic remain fallbacks.
Embeddings
Google Gemini embeddings
Semantic vectors for Qdrant retrieval when GEMINI_API_KEY is configured.
Retrieval
Qdrant + local catalog fallback
Verified schemes indexed with eligibility rules, keywords, and metadata.
Memory
Qdrant citizen_sessions + local fallback
Citizen profile and found schemes persisted by browser-generated session key.
Safety
Enkrypt AI + fallback validator
Citizen-facing reports and Vigilance alerts checked before display.
Operations
Langfuse + Upstash + Vercel
Tracing, rate limiting, and deployed hackathon runtime.
| Layer | Technology | Purpose |
|---|---|---|
| Frontend | TanStack Start (React 19) + Tailwind v4 | Landing page, agent runner, report view, scheme catalog, and alert banner. |
| Orchestration | Mastra-style TypeScript workflow | Chains five agents through typed server functions that are testable in isolation. |
| Reasoning | Featherless AI + OpenAI/OpenRouter fallback | Featherless powers open-source model reasoning for reports and Vigilance alert reasons; OpenAI, OpenRouter, and local grounded logic remain fallbacks. |
| Embeddings | Google Gemini embeddings | Semantic vectors for Qdrant retrieval when GEMINI_API_KEY is configured. |
| Retrieval | Qdrant + local catalog fallback | Verified schemes indexed with eligibility rules, keywords, and metadata. |
| Memory | Qdrant citizen_sessions + local fallback | Citizen profile and found schemes persisted by browser-generated session key. |
| Safety | Enkrypt AI + fallback validator | Citizen-facing reports and Vigilance alerts checked before display. |
| Operations | Langfuse + Upstash + Vercel | Tracing, rate limiting, and deployed hackathon runtime. |
Agents
What each agent contributes
1. Profile Agent
structureExtracts state, age, gender, category, income, occupation, land, disability, and other profile facts from guided or plain-language input.
2. Discovery Agent
retrieveSearches the 28-scheme Central + Telangana catalog through Qdrant retrieval or local fallback.
3. Eligibility Agent
evaluateChecks deterministic rules such as age, gender, category, income cap, occupation, state, Aadhaar status, and bank account status.
4. Report Agent
explainCreates a simple source-grounded report that uses likely eligible language and includes documents, steps, sourceUrl, and lastVerified.
5. Vigilance Agent
watchScans saved profiles against unseen schemes and creates proactive alerts when a new likely match appears.
Guardrails
Honest fallback behavior
Retrieval
Qdrant is primary; local catalog fallback keeps the demo runnable.
Memory
Qdrant memory is primary; local session fallback is visible in badges.
Safety
Enkrypt validates outputs; fallback status stays honest when needed.
Operations
Langfuse and Upstash signals are inspectable on the debug page.