Build a Production Chatbot with DeepSeek API

AI customer support is one of the most mature LLM use cases. This guide builds a production-grade bot with DeepSeek API — streaming replies, multi-turn memory, tool calling, and knowledge-base retrieval.

Step 1: API Key and First Call

Register at platform.deepseek.com for an API key. DeepSeek API is OpenAI-compatible — reuse the OpenAI SDK:

from openai import OpenAI

client = OpenAI(api_key="YOUR_KEY", base_url="https://api.deepseek.com/v1")

resp = client.chat.completions.create(
    model="deepseek-chat",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

Step 2: Enable Streaming

Great support UX streams tokens as they're generated:

stream = client.chat.completions.create(
    model="deepseek-chat",
    messages=messages,
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="", flush=True)

Step 3: Multi-Turn Context

Append history to messages. V4's long context lets you keep more conversation without aggressive truncation:

messages = [
    {"role": "system", "content": "You are a patient, professional support agent. Answer product questions only."},
]
messages.append({"role": "user", "content": user_input})
# Append assistant reply after each turn

Step 4: Function Calling for Business Systems

Let the model query orders and create tickets:

tools = [{
    "type": "function",
    "function": {
        "name": "get_order_status",
        "description": "Look up shipping status by order ID",
        "parameters": {
            "type": "object",
            "properties": {"order_id": {"type": "string"}},
            "required": ["order_id"],
        },
    },
}]

When the model returns tool_calls, execute the function and return results — enabling real "track my order" and returns workflows.

Step 5: Knowledge Base (RAG)

Chunk product docs, embed, retrieve relevant passages, and inject into the system prompt. The model answers from your knowledge with fewer hallucinations.

Production Tips

• Cost: Run non-real-time jobs (ticket summaries) during API off-peak hours for lower rates.
• Reliability: Time out and retry tool calls; rate-limit during peak API hours.
• Evaluation: Build a regression set from historical tickets.

---

📅 Updated July 2, 2026. Full sample code in API docs.