DGX Lab Agent

The DGX Lab Agent is a codebase-aware assistant embedded in the dashboard navbar. It uses RAG over the full repository, agent persona definitions, and skill documentation to answer questions about DGX Lab's architecture, tools, configuration, and team structure.

Stack

Component	Implementation
LLM	Claude 3.5 Haiku via AWS Bedrock (anthropic.claude-3-5-haiku-20241022-v1:0)
Embeddings	nvidia/llama-embed-nemotron-8b on CUDA (FAISS GPU)
Reranker	nvidia/llama-nemotron-rerank-1b-v2 on CUDA
Vector store	FAISS (GPU), persisted to ~/.dgx-lab/agent/faiss-index/
Framework	LangChain + LangSmith tracing
Frontend	Sheet panel in the navbar (480px slide-out)

How it works

1. Document collection -- On first invocation (or manual reindex), the agent walks the codebase starting from CODEBASE_ROOT, respecting .cursorignore patterns. It skips node_modules, .venv, __pycache__, .git, binary files, and lock files. Agent personas and skills get a metadata boost for retrieval priority.

2. Splitting -- Documents are split using language-aware splitters: Python, TypeScript, and Markdown each get their own RecursiveCharacterTextSplitter with appropriate chunk sizes. Everything else gets a 1500-char default splitter.

3. Embedding and indexing -- Chunks are embedded with nvidia/llama-embed-nemotron-8b and stored in a FAISS GPU index. The index is saved to disk and reloaded on subsequent starts.

4. Retrieval and reranking -- Queries retrieve the top 20 candidates from FAISS, then nvidia/llama-nemotron-rerank-1b-v2 (cross-encoder) reranks them down to the top 6.

5. Generation -- The reranked context, team directory, skills summary, and conversation history are injected into the prompt. Claude 3.5 Haiku generates the response with temperature 0.3 and a 2048 token limit.

6. Tracing -- Every invocation is traced. If LANGSMITH_API_KEY is set, traces go to LangSmith cloud. Regardless, a local JSONL export is written to ~/.dgx-lab/langsmith-traces/traces.jsonl.

Using the agent

Click the agent icon in the navbar to open the chat sheet. Type a question and press Enter or click Send.

The agent understands:

• Codebase structure, file locations, and configuration values
• How each tool works (Monitor, Control, AutoModel, Designer, Curator, Datasets, Traces, LangSmith)
• Agent persona roles and responsibilities (all 14 team members)
• Skills from .cursor/skills/
• DGX Spark hardware specs and constraints

Each response includes source citations -- the files the agent retrieved to ground its answer. These appear as chips below the message.

Use the New chat button to start a fresh conversation.

Environment variables

Variable	Default	Purpose
LANGSMITH_API_KEY	(none)	Enables LangSmith cloud tracing
DGX_LAB_CODEBASE_ROOT	Current working directory	Root directory for document collection
DGX_LAB_AGENT_INDEX_DIR	~/.dgx-lab/agent	FAISS index and conversation storage
DGX_LAB_LANGSMITH_TRACES_DIR	~/.dgx-lab/langsmith-traces	Local trace export directory
AWS_REGION / AWS_DEFAULT_REGION	us-east-1	Bedrock region for Claude

AWS credentials (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, or an instance profile) must be configured for Bedrock access.

API endpoints

All endpoints are under /api/agent.

Method	Path	Description
POST	/api/agent/chat	Send a message. Body: { "message": "...", "conversation_id": "..." }
GET	/api/agent/conversations	List all conversations (id, title, turn count, last activity)
GET	/api/agent/conversations/{id}	Get full conversation history
DELETE	/api/agent/conversations/{id}	Delete a conversation
POST	/api/agent/reindex	Force-rebuild the FAISS index from the current codebase

Chat response shape

{
  "conversation_id": "uuid",
  "answer": "The Monitor tool reads GPU data via...",
  "sources": ["backend/app/routers/monitor.py", "docs/setup.md"],
  "trace_id": "uuid",
  "duration_ms": 3200
}

Reindexing

The index is built lazily on first query and cached. To force a rebuild after code changes:

curl -X POST http://localhost:8000/api/agent/reindex

Or restart the backend -- the index will be rebuilt on the next query if the saved index file is missing.

Evaluation

The agent ships with a seed evaluation dataset (dgx-lab-agent-evals) and a keyword-overlap correctness evaluator. Both require LANGSMITH_API_KEY.

The dataset covers five categories: tool knowledge, hardware specs, codebase structure, team awareness, and skill documentation. The evaluator scores predictions by checking how many key terms from the reference answer appear in the agent's response.

Evaluation functions are in backend/app/agent/evals.py. They are not run automatically -- invoke them from a script or notebook when validating agent quality.

Source files

Layer	Path
Agent chain	backend/app/agent/chain.py
RAG pipeline	backend/app/agent/rag.py
Personas loader	backend/app/agent/personas.py
Skills loader	backend/app/agent/skills.py
Tracing config	backend/app/agent/tracing.py
Evaluation	backend/app/agent/evals.py
Chat router	backend/app/routers/agent_chat.py
Config paths	backend/app/config.py
Frontend sheet	frontend/apps/web/components/agent-sheet.tsx