rohit-p27/healthcare-ai-india

Purpose and Entry Points

AI Sarthi is an India-focused healthcare-governance intelligence platform: operational data, analytics, RAG, and agent workflows support budget, epidemic, supply, insurance, geographic, and district analysis. Runtime entry points are the FastAPI backend at core_platform/backend/main.py (with application setup in core_platform/backend/app/main.py), Next.js UI under core_platform/frontend/app, and CLI-oriented ETL/RAG packages via core_platform/etl/__main__.py and core_platform/rag/__main__.py.

User Interface and API Boundary

The Next.js/TypeScript frontend presents dashboards, geographic drilldowns, domain screens, and an AI copilot/query experience. Shared API access is centralized in core_platform/frontend/lib/api.ts; app routes include core_platform/frontend/app/ask/page.tsx, core_platform/frontend/app/districts/page.tsx, and core_platform/frontend/app/knowledge-graph/page.tsx. It communicates with versioned FastAPI routes rather than accessing data stores directly.

Backend Domain Services

FastAPI route handlers define the external contract and delegate domain logic to service modules. The central route composition is core_platform/backend/api/v1/router.py; representative boundaries are chat/orchestration (core_platform/backend/api/v1/chat.py, core_platform/backend/api/v1/orchestrator.py), RAG (core_platform/backend/api/v1/rag.py), and analytical domains such as budget, epidemic, insurance, supply, district, and geo. Services own application-level queries and calculations, e.g. core_platform/backend/services/budget_service.py, core_platform/backend/services/epidemic_service.py, and core_platform/backend/services/rag_service.py.

Persistent Data and Governance Model

The backend separates ORM persistence models from request/response schemas: core_platform/backend/models covers health, governance, coordination, geo, RAG, and analytic domains, while core_platform/backend/schemas is the API serialization boundary. Database initialization/session handling lives in core_platform/backend/db/session.py; Alembic migrations establish core schemas, district-profile fields, and pgvector support in core_platform/backend/alembic/versions/0001_create_schemas_and_core_tables.py through 0003_rag_pgvector.py. SQLite database files are present for development; the intended production relational/vector target is PostgreSQL with pgvector.

Data Ingestion and Curated Marts

ETL converts source health/governance data into normalized, validated data products. The CLI orchestrates pipelines from core_platform/etl/cli.py; shared ingestion and normalization are in core_platform/etl/common/ingestor.py and core_platform/etl/common/normalize.py. Geographic identity is a cross-domain join boundary via core_platform/etl/transform/geo_spine.py; district profiles derive from it in core_platform/etl/transform/district_profile.py. Curated Parquet marts in datasets/curated/mart_geo_spine_v1/data.parquet and datasets/curated/mart_district_profile_v1/data.parquet feed operational loading and analytics.

AI, RAG, and Agent Control Flow

Natural-language requests enter API chat/orchestration, then the agentic gateway and orchestrator select specialized budget, epidemic, insurance, policy, research, or supply agents. Coordination and memory abstractions are in core_platform/agents/core/orchestrator.py, core_platform/agents/core/memory.py, and core_platform/agents/workflows/agentic_gateway.py. RAG ingestion loads documents, chunks, embeds, and indexes them; retrieval performs hybrid search, reranking, context construction, and citations via core_platform/rag/indexing/embedder.py, core_platform/rag/retrieval/pipeline.py, and core_platform/rag/retrieval/hybrid_search.py. The repository also contains Neo4j-oriented graph integration at core_platform/knowledge_graph/connector.py and ontology definitions at core_platform/knowledge_graph/ontology.py.

Analytics and Model Operations

Domain APIs/services expose predictive and anomaly-oriented functions described by the platform: epidemic detection, budget forecasting, and insurance analysis. MLOps primitives for features, registry, monitoring, and serving are isolated in core_platform/mlops/feature_store.py, core_platform/mlops/registry.py, and core_platform/mlops/serving.py. Healthcare SLM training assets are separate from serving infrastructure under healthcare_slm/training, including healthcare_slm/training/llm/finetune_qlora.py.

Deployment and External Dependencies

The main material runtimes are Python/FastAPI, Node.js/Next.js, PostgreSQL/pgvector, and optionally Neo4j for GraphRAG. Docker deployment artifacts are deployment/Dockerfile.backend, deployment/Dockerfile.frontend, and deployment/docker-compose.yml. External knowledge inputs are local policy, budget, and health PDFs/CSVs under data/raw; embedding/vector artifacts also exist in vector_store/chroma.sqlite3, indicating a local Chroma-style development vector store alongside the PostgreSQL vector migration path.