Build AI agents in India (LLM + tools + loops)

Build AI agents in India (₹-native API)

AI agents combine an LLM, function calling (tools), and reasoning loops to complete multi-step tasks. Build a customer-support agent or sales bot—all billed in ₹ rupees on a GST invoice.

Here's the 3-step setup + real ₹ costs.

Why unoblox for agents

• Function calling: Claude and Qwen support tool calls natively
• ₹ billing: No international payments, GST invoice, input-tax-credit claimable
• One key, one endpoint: Manage all agent calls in rupees
• OpenAI-compatible: Works with LangChain agent loops
• Streaming: Real-time agent thinking for interactive apps

Recommended models

• Qwen3 Max: ₹120.95 input / ₹604.77 output per 1M tokens (fast reasoning, high-volume)
• Claude Sonnet: ₹201.6 input / ₹1,008 output per 1M tokens (superior reasoning, complex)

Step 1: Define your tools

An agent needs external tools to act (search APIs, DBs, email). Define as JSON schema.

tools = [
    {
        "type": "function",
        "function": {
            "name": "search_docs",
            "description": "Search your knowledge base",
            "parameters": {
                "type": "object",
                "properties": {"query": {"type": "string"}},
                "required": ["query"]
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "send_email",
            "description": "Send an email to a customer",
            "parameters": {
                "type": "object",
                "properties": {
                    "to": {"type": "string"},
                    "subject": {"type": "string"},
                    "body": {"type": "string"}
                },
                "required": ["to", "subject", "body"]
            }
        }
    }
]

Step 2: Call the LLM with function calling

The LLM receives tools and decides which to call.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.unoblox.ai/v1",
    api_key="ub-gw-..."
)

messages = [{"role": "user", "content": "Customer asked about refunds. Check our policy and reply."}]

response = client.chat.completions.create(
    model="qwen/qwen3-max",
    messages=messages,
    tools=tools,
    tool_choice="auto"
)

Step 3: Loop until task completes

Execute the tool, feed the result back, and repeat.

while response.choices[0].finish_reason == "tool_calls":
    for tool_call in response.choices[0].message.tool_calls:
        if tool_call.function.name == "search_docs":
            result = search_docs(tool_call.function.arguments["query"])
        else:
            result = send_email(**tool_call.function.arguments)
        
        messages.append({"role": "assistant", "content": response.choices[0].message.content})
        messages.append({"role": "user", "content": f"Result: {result}"})
    
    response = client.chat.completions.create(
        model="qwen/qwen3-max",
        messages=messages,
        tools=tools
    )

Real ₹ cost: customer-support agent

Scenario:

• 100 support tickets/day
• 3 tool calls per ticket
• Avg: 200 input tokens, 150 output tokens per turn

Daily cost (Qwen3 Max):

• 300 turns × (200 input + 150 output) = 105k tokens
• Input: ₹7.26 | Output: ₹27.21
• Daily: ₹34.47 = ₹1,034/month

Daily cost (Claude Sonnet):

• Input: ₹12.10 | Output: ₹45.36
• Daily: ₹57.46 = ₹1,724/month

Qwen saves ₹690/month.

Common patterns

• Research agent: Search docs → synthesize → output
• Data analyst: Query DB → transform → visualize
• Sales agent: Lookup customer → check inventory → email quote

Frequently asked questions

Q: Tool calling vs RAG—what's the difference? Tool calling: LLM decides to call external functions. RAG: You retrieve docs before calling LLM. Agents are more flexible.

Q: Can an agent use 5+ tools? Yes. Define as many tools as needed. LLM picks the right ones per task.

Q: Does agent loop cost scale with tool calls? Yes. Each LLM turn costs tokens. Monitor to avoid runaway costs.

Q: How do I prevent infinite loops? Set max_iterations limit (e.g., 10). If agent doesn't finish, return "max iterations reached" error.

Q: Does LangChain work with unoblox? Yes. LangChain's AgentExecutor works as the LLM backend. Swap base_url and api_key.

Q: What if a tool call fails? Return error message as tool result. LLM learns and retries or pivots to another tool.

Get started in rupees → https://unoblox.ai/sign-in