TinDog Production AI Agent Backend

A production-ready TypeScript/Node.js Express backend for a streaming LangChain tool-calling agent using NVIDIA's OpenAI-compatible API, LangGraph orchestration, LangSmith tracing, Tavily search, Wikipedia, calculator, validation, memory, logging, metrics, Docker, and tests.

Architecture

Frontend
  ↓ SSE / POST /chat
Express Server
  ↓
Request Validation (Zod)
  ↓
Conversation Manager (replaceable in-memory store)
  ↓
LangSmith Trace + metadata/tags
  ↓
LangGraph Workflow
  ↓
Tool Calling Agent (createToolCallingAgent + AgentExecutor)
  ↓
Tools: Tavily Search, Wikipedia, Calculator
  ↓
SSE Streaming Response
  ↓
Client

Folder Structure

• src/agent — LangChain agent factory.
• src/graph — LangGraph workflow wrapper.
• src/tools — Modular Tavily, Wikipedia, and calculator tools.
• src/prompts — System prompt.
• src/memory — Conversation memory abstraction and in-memory implementation.
• src/langsmith — LangSmith tracing setup.
• src/services — Metrics service.
• src/config — Central validated configuration.
• src/middleware — Rate limiting and error handling.
• src/routes / src/controllers — HTTP API.
• src/schemas — Zod request schemas.
• src/events — SSE writer.
• src/logger — Pino logging.
• src/errors — Custom errors.

Installation

npm install
cp .env.example .env

Fill in NVIDIA_API_KEY, TAVILY_API_KEY, and optionally LANGSMITH_API_KEY.

Environment Variables

See .env.example for all configuration. Model, temperature, max tokens, timeouts, retry settings, rate limits, tool sizes, LangSmith project, and CORS origin are configurable.

Running

npm run dev
npm run build
npm start

Development

npm test

Production

Use npm run build then npm start, or Docker Compose.

LangSmith Setup

Set:

LANGSMITH_TRACING=true
LANGSMITH_API_KEY=your_key
LANGSMITH_PROJECT=tindog-agent

Every agent invocation includes request ID, session ID, conversation ID, user ID, tags, tool calls, latency, errors, and token usage when returned by the provider.

NVIDIA Setup

Set:

NVIDIA_API_KEY=your_key
NVIDIA_BASE_URL=https://integrate.api.nvidia.com/v1
NVIDIA_MODEL=mistralai/mistral-medium-3.5-128b

The backend uses ChatOpenAI with NVIDIA's OpenAI-compatible endpoint.

Tavily Setup

Set:

TAVILY_API_KEY=your_key
TAVILY_MAX_RESULTS=3

API

POST /chat

Streams SSE.

Request:

{"query":"What is NVIDIA's latest AI news?","userId":"user-123","tags":["demo"]}

POST /memory/clear

{"conversationId":"00000000-0000-4000-8000-000000000000"}

GET /health

Returns service health.

GET /metrics

Returns total requests, successes, failures, average latency, tool counts, token usage, and LLM calls.

SSE Event Format

Each event is JSON and uses one of:

• reasoning
• tool_start
• tool_end
• response
• metadata
• done
• error

Example:

event: response
data: {"type":"response","requestId":"...","conversationId":"...","timestamp":"...","data":{"token":"Hello"}}

Deployment

docker compose up --build

Troubleshooting

• Invalid configuration: verify .env and required API keys.
• No streaming: ensure the proxy disables response buffering.
• Tool errors: verify Tavily key and outbound network access.
• LangSmith missing traces: verify LANGSMITH_API_KEY and project name.

Next.js Frontend

A polished animated Next.js UI lives in web/. It provides:

• A glassmorphism landing page for the production agent.
• Animated aurora gradients, floating tool chips, and streaming message transitions.
• A chat composer that posts to the Express /chat endpoint and parses the SSE stream from reasoning, tool_start, tool_end, response, metadata, done, and error events.
• Configurable backend URL via NEXT_PUBLIC_API_BASE_URL.

Running the UI

cd web
npm install
cp .env.local.example .env.local
npm run dev

The default UI runs at http://localhost:3000 and expects the backend at http://localhost:3001.

Full Stack Docker Compose

docker compose up --build

This starts:

• ai-agent-api on port 3001.
• ai-agent-ui on port 3000.