Building a coding chatbot with DeepSeek R1

Working with DeepSeek in Python

James Chapman

Curriculum Manager, DataCamp

Current approach

All of the input and output is stored to be sent with the next message.

Problems with reasoning

Reasoning outputs consist of thoughts and a final response.

Problems with reasoning

Best practices dictates that reasoning inputs should be short.

Problems with reasoning

Storing thoughts in the conversation history increases cost, runtimes, and will likely degrade performance.

The reasoning approach

Trimming the thoughts from the first model output.

The reasoning approach

Trimming the thoughts from the second model output.

Trimming thoughts

Extract thinking tokens

import re

match = re.search(r'<think>(.*?)<\/think>', response_content, re.DOTALL)
think_content = match.group(1).strip()
print(think_content)

Remove thinking tokens

final_response = re.sub(r'<think>[\s\S]*?<\/think>\s*', '', response_content, re.DOTALL)
print(final_response.strip())

Coding a reasoning chatbot

"system" messages shouldn't be used with reasoning models

messages = []
user_msgs = [code_to_debug, follow_up]

for q in user_qs:
    user_dict = {"role": "user", "content": q}
    messages.append(user_dict)

    response = client.chat.completions.create(model="deepseek-ai/DeepSeek-R1", messages=messages)


    final_response = re.sub(r'<think>[\s\S]*?<\/think>\s*', '', response.choices[0].message.content, re.DOTALL)
    assistant_dict = {"role": "assistant", "content": final_response.strip()}

    messages.append(assistant_dict)

Let's practice!

Working with DeepSeek in Python