How abinesha312/gemmagenai works
Project type and purpose
• This is a full-stack AI assistant: a real-time web UI built with Chainlit talks to a Python multi-agent backend, which in turn offloads heavy neural inference to a vLLM service running on NVIDIA H100 GPUs.
• Purpose: provide specialized academic support (email composition, research guidance, campus info) via modular, domain-specific agents.
File structure
• Top-level directories and files:
• .chainlit/ (UI config & locales)
• demo/ (video & screenshots)
• docker-compose.yml, podman-compose.yml, Containerfile.chainlit, Containerfile.vllm, podman_run.sh (deployment scripts)
• run.py, setup.py, requirements.txt
• src/ (application code)
• src/ breakdown:
• src/app.py – Chainlit server entrypoint
• src/agents/ – base_agent.py, registry.py, specialized_agents.py
• src/config/ – prompts.py, redis_config.py, settings.py
• src/models/ – classification.py (TF-IDF & cosine similarity), query_models.py
• src/utils/ – redis_manager.py, vector_db.py
• src/services/ – (any orchestrators or wrappers)
README for additional insights
• Frontend: Chainlit real-time web interface, handles routing & formatting
• Inference: vLLM hosting Gemma-3-27B & LLaMa 3.2 90B Vision models, optimized for parallel GPU inference
• Data stores: Redis for caching/session state; FAISS vector store for retrieval-augmented generation
• Architecture: modular multi-agent pipeline, intelligent query classification, containerized via Podman/Docker Compose on multi-GPU hosts
Detail
• Break out the Application Backend into sub-modules: Agent Registry, BaseAgent, Specialized Agents, config & utils
• Orchestration layer (docker-compose/podman-compose) as the deployment boundary
• Protocols and ports (WebSocket: 8000, gRPC: 50051, Redis: 6379)