How nvidia-ai-blueprints/aiq-research-assistant works
Project Type and Purpose
• This is a full-stack AI research assistant blueprint. It combines a demo frontend web app, a Python backend service (the aiq_aira package), a middleware proxy, and integrations with external AI/RAG microservices and web search. Its purpose is to enable on-premise deep research using internal documents and optional web search.
File Structure Analysis
• Top-level directories:
• aira: Python package containing core AI-Q Research Assistant logic and REST API entrypoints.
• data: Sample datasets and scripts for on-premise document ingestion.
• demo: Demo frontend container (source not included) exposing the web UI.
• deploy/compose: Docker Compose files and NGINX proxy templates.
• notebooks: Jupyter notebooks for quick experimentation.
• Key configs: aira/configs/config.yml, Dockerfile in aira and data.
• Tests in aira/test_aira exercise each backend function.
README Insights
• Components described:
• Demo Frontend (web app container)
• Backend Service (aiq-aira Python package, exposed via REST)
• Middleware Proxy (NGINX routing to backend and NVIDIA RAG)
• External Services: NVIDIA RAG Blueprint, NeMo Retriever Microservices, NIM Microservices, Tavily web search.
Main Components
a. User & Demo Frontend (Docker container hosting web UI)
b. NGINX Middleware Proxy (routes HTTP between frontend, backend, RAG services)
c. AI-Q Research Assistant Backend (Python aiq_aira service; includes modules for plan generation, RAG queries, summary)
d. NVIDIA RAG Services (hosted or local microservices for document retrieval)
e. NVIDIA NeMo Retriever & NIM Microservices (multi-modal ingest and embedding)
f. External Web Search (Tavily API)
g. On-Premise Data Store (mounted volume with internal documents)
h. Docker Compose Orchestrator (defines service graph)
Relationships and Interactions
• User → Frontend (HTTP)
• Frontend → NGINX Proxy → Backend API endpoints
• Backend → RAG Services (via REST)
• Backend → Web Search (Tavily HTTP)
• Backend → Data Store (local file I/O)
• NGINX → Routes to NVIDIA RAG vs AI-Q backend based on path
Architectural Patterns & Technologies
• Microservices (separate containers for each service)
• Layered separation: Presentation (frontend), API layer (backend), integration layer (proxy), external AI services.
• Config-as-code with YAML (config.yml), Dockerfiles, Docker Compose.
• Python with LangGraph toolkit (observability, config).
• NGINX as reverse-proxy and TLS terminator.
• RESTful APIs for service communication.