Production-style Retrieval-Augmented Generation service with sentence chunking, local hashed embeddings, hybrid retrieval, reranking, grounded answer generation, and lightweight retrieval evaluation.

Loading

GET /events protected when API_KEY is set

POST /query with { "query": "remote work policy" }

See DEMO.md for terminal demo steps, curl commands, and sample request/response files.

Set API_KEY to require X-API-Key on query/event endpoints. Set APP_DB_PATH to control the SQLite event database location.

With the server running, use a second terminal for the smoke check:

Docker:

Kubernetes manifests live in k8s/deployment.yaml and include probes, resource limits, a Service, and a PVC for the SQLite event store. The default manifest uses one replica because SQLite is the default event store.

Dockerfile, Docker Compose, and Kubernetes configuration are validated by static inspection/YAML parsing in this workspace. Runtime container and cluster validation remains a CI or cloud-environment step.

Purpose: retrieval service for source-backed HR policy answers.

Quickstart: run tests/eval, start uvicorn api.server:app --reload --port 8000, then run python scripts/smoke_test.py.

Demo path: use DEMO.md for curl examples and sample request/response files.

Deployment status: local tests and smoke tests pass; Docker/Compose/Kubernetes config is present; Docker image builds are validated in CI; cloud deployment is pending.

Remaining gaps: production corpus, managed retrieval infrastructure, managed auth/secrets, observability, cloud deployment, and production data governance.

Portfolio index: https://github.com/Adityansh-Chand/ai-engineering-portfolio

Overlapping sentence chunking.

Local deterministic embedding provider.

Hybrid semantic and lexical retrieval.

Query-aware reranking.

Source-bearing structured responses.

Evaluation runner with bundled HR policy queries.

SQLite event audit trail for query metadata.

GitHub Actions CI for tests, eval, and container build.

Production data contract in datasets/production_schema.json.

MIT