ogonkem/naija-civic-guard_AI_Platform

A reusable civic-legal AI agent platform — built on top of the existing Naija Civic Guard RAG system — demonstrating agent orchestration, MCP-based tool integration, production-grade model serving, governance, and full observability # Naija Civic Guard **Grounded question-answering over the Constitution of the Federal Republic of Nigeria (1999), built as a running platform rather than a notebook.** Ask a question in plain English — *"What does Section 45 say about restrictions on fundamental rights?"* — and get an answer drawn **only** from the constitutional text, with the specific sections it came from, streamed token-by-token. Underneath that answer is a full operational stack: an agentic retrieval pipeline, a Model Context Protocol (MCP) tool server, an API gateway with authentication / rate limiting / audit logging, per-request metrics in Postgres and Prometheus, asynchronous quality evaluation on its own worker, Grafana dashboards, and dbt models over the logs — all standing up from a single `docker compose up`. > ⚖️ **It is an educational tool.** Retrieval precision is not at a production > bar for legal use (see Limitations). Treat every answer as a > pointer to sections worth reading, not as legal advice. --- ## Contents - Why this exists - The system at a glance - How a request flows - The build, layer by layer - 1. Ingestion & the vector store - 2. Hybrid retrieval - 3. The retrieval agent (LangGraph) - 4. The MCP tool server - 5. Generation & the provider switch - 6. The API gateway - 7. Request correlation - 8. Observability — two layers - 9. Asynchronous evaluation - 10. Analytics (dbt) - 11. Containerisation - Tech stack - Getting started - Using it - Observing it - Testing - Performance (measured) - Project layout - Limitations - Roadmap --- ## Why this exists A generic LLM will answer questions about Nigerian law from a blurry mixture of training data, other jurisdictions, and confident guessing. For a legal document you want the opposite: answers **anchored to the source**, with citations, that say *"I don't know"* when the text doesn't cover it. Retrieval-Augmented Generation gets you there — retrieve the relevant passages, put them in the prompt, instruct the model …