How to setup Digital Ocean GPU Droplet using RTX 6000

				
					import os
from fastapi import FastAPI
from pydantic import BaseModel

from duckduckgo_search import DDGS

from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent
from langgraph_supervisor import create_supervisor

# ---------
# Tools
# ---------
def web_search(query: str) -> str:
    """Lightweight web search tool using DuckDuckGo snippets."""
    with DDGS() as ddgs:
        results = list(ddgs.text(query, max_results=5))
    # Return condensed snippets for the agent to cite/ground itself.
    lines = []
    for r in results:
        title = r.get("title", "")
        href = r.get("href", "")
        body = r.get("body", "")
        lines.append(f"- {title}\n  {href}\n  {body}")
    return "\n".join(lines)

def safe_calc(expression: str) -> str:
    """Very small calculator tool (avoid arbitrary code execution)."""
    allowed = set("0123456789+-*/(). %")
    if any(ch not in allowed for ch in expression):
        return "Unsupported characters in expression."
    try:
        return str(eval(expression, {"__builtins__": {}}, {}))
    except Exception as e:
        return f"Error: {e}"

# ---------
# Model: points to local vLLM OpenAI-compatible server
# ---------
VLLM_BASE_URL = os.getenv("VLLM_BASE_URL", "http://127.0.0.1:8000/v1")
MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-7B-Instruct")

llm = ChatOpenAI(
    model=MODEL_NAME,
    api_key=os.getenv("OPENAI_API_KEY", "not-used"),
    base_url=VLLM_BASE_URL,
    temperature=0,
)

# ---------
# Worker agents
# ---------
research_agent = create_react_agent(
    model=llm,
    tools=[web_search],
    name="research_expert",
    prompt="You are a research expert. Use web_search to gather facts and relevant links."
)

math_agent = create_react_agent(
    model=llm,
    tools=[safe_calc],
    name="math_expert",
    prompt="You are a math expert. Use safe_calc for numeric calculations."
)

# ---------
# Supervisor
# ---------
workflow = create_supervisor(
    [research_agent, math_agent],
    model=llm,
    prompt=(
        "You are a supervisor managing a research expert and a math expert. "
        "Delegate research questions to research_expert. "
        "Delegate calculations to math_expert. "
        "Synthesize a final answer for the user."
    ),
)

graph_app = workflow.compile()

# ---------
# FastAPI wrapper
# ---------
app = FastAPI(title="Agentic GPU API", version="0.1.0")

class ChatIn(BaseModel):
    message: str

@app.post("/chat")
def chat(payload: ChatIn):
    result = graph_app.invoke(
        {"messages": [{"role": "user", "content": payload.message}]}
    )
    # Return the last model message
    messages = result.get("messages", [])
    final = messages[-1].content if messages else ""
    return {"answer": final}