AsoTora – rag-project-showcase

About

Production-style RAG system for SRE/DevOps knowledge. It ingests runbooks and docs into Chroma, serves answers through a FastAPI + Streamlit Ops Copilot, and is designed for GKE deployment with Terraform and GitHub Actions CI/CD.

Architecture

flowchart LR
    A[Runbooks and Docs] --> B[Ingestor Service]
    B --> C[Embeddings]
    C --> D[(Chroma Vector DB)]
    E[Retriever API] --> D
    F[DevOps Agent API] --> E
    F --> D
    F --> G[LLM Provider]
    F --> H[(Cloud SQL Postgres Logs/Feedback)]
    I[Streamlit UI] --> F
    U[User] --> I

Loading

Flow: documents are embedded and indexed in Chroma, the UI sends questions to the DevOps agent, and the agent retrieves context, calls the LLM, and logs interactions.

Development

uv sync --all-groups
docker compose up --build

Optional local runs (without Docker):

make ingest-local
make run-retriever
make run-devops-agent
make run-ui

Metrics endpoints:

• DevOps Agent: http://localhost:8080/metrics
• Retriever: http://localhost:8081/metrics

Evaluation

Run the RAG evaluation pipeline:

make eval

Run three experiment variants and compare in MLflow:

make eval-experiments
make mlflow-ui

Results are saved under eval/results/YYYY-MM-DD.json.

Validation

Validation checklist and latest run results: VALIDATION.md

Infrastructure (GKE)

Low-cost single-cluster infrastructure is defined in infra/terraform.

See infra/README.md for:

• Terraform apply flow
• required GitHub Actions variables/secrets
• DNS + ingress setup
• observability stack rollout (Grafana + in-cluster Prometheus + GKE managed metrics/SLOs)