I’m slowly but surely wading through examples of the OpenAI Responses API. Some people can learn just by reading, but I can only learn by coding.

Briefly, to have a conversation, you need to remember the responses of the previous queries. This is accomplished using a response ID fed to the previous_response_id parameter. My demo asks the GPT LLM to tell a joke, and then asks for an explanation, and then asks for another joke.

There are three requests. In order to avoid repeating the parameters that don’t change (the model to use, the temperature, and so on), I set up an info dictionary:

  # set up basic stuff in dictionary to be unpacked
  info = {
    "model": "gpt-4.1",
    "input": [],
    "temperature": 0.5,  # mildly creative (default top_p)
    "max_output_tokens": 100,
    "store": True,  # is the default
  }

Only the “input” parameter will change. The “store” parameter means remember the response. The first query is:

  query1 = "Tell me a joke."
  print("query1 = ")
  print(query1)
  info["input"] = [{ "role": "user", "content": query1 }]
  response1 = client.responses.create(**info)

The input is set and then the info dictionary is unpacked to the create() call using the special ** syntax. The second query is set up like so:

  query2 = "Briefly, explain to me why this is funny"
  print("query2 = ")
  print(query2)
  info["input"] = [{ "role": "user", "content": query2 }]
  info["previous_response_id"] = response1.id
  response2 = client.responses.create(**info)

The previous_response_id parameter is set to the ID of the previous response, and so the second query knows what the joke was.

In the same way, the third query is constructed so that it knows what the original joke and explanation are, so that it doesn’t repeat the first joke:

  query3 = "Tell me a different joke."
  print("query3 = ")
  print(query3)
  info["input"] = [{ "role": "user", "content": query3 }]
  info["previous_response_id"] = response2.id
  response3 = client.responses.create(**info)

Interesting stuff. One step further in my journey to understand AI using large language models.

Complete demo program:

# conversation_openai.py
# conversation using response.id to maintain state

from openai import OpenAI

print("\nBegin conversation demo ")
try:
  key = "sk-proj-_AX7bGTXUwg-qojh2T5Z2CVXrox" + \
  "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" + \
  "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" + \
  "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
 
  client = OpenAI(api_key=key)

  # set up basic stuff in dictionary to be unpacked
  info = {
    "model": "gpt-4.1",
    # "tools": [],
    # "instructions": are per request
    "input": [],
    "temperature": 0.5,  # mildly creative (default top_p)
    "max_output_tokens": 100,
    "store": True,  # is the default
  }
  
  query1 = "Tell me a joke."
  print("\nquery1 = ")
  print(query1)
  info["input"] = [{ "role": "user", "content": query1 }]
  response1 = client.responses.create(**info)

  print("\nThe reply is: ")
  print(response1.output_text)

  query2 = "Briefly, explain to me why this is funny"
  print("\nquery2 = ")
  print(query2)
  info["input"] = [{ "role": "user", "content": query2 }]
  info["previous_response_id"] = response1.id
  response2 = client.responses.create(**info)
  
  print("\nThe reply is: ")
  print(response2.output_text)

  query3 = "Tell me a different joke."
  print("\nquery3 = ")
  print(query3)
  info["input"] = [{ "role": "user", "content": query3 }]
  info["previous_response_id"] = response2.id
  response3 = client.responses.create(**info)

  print("\nThe reply is: ")
  print(response3.output_text)  

except Exception as e:
  print("Fatal error " + str(e))
 
print("\nEnd conversation demo ")

Humor is a funny thing. What is funny to one person is not funny to another.

Within seconds, the game descends into absolute chaos, with cards and hands and arms flying everywhere. I often end up laughing so hard I cry. Every single person I know, who has been introduced to “Pit”, loves the game.