This documentation covers how to get started with the API that backs OpenGPTs.
Sign in to like and favorite skills
This documentation covers how to get started with the API that backs OpenGPTs. This allows you to easily integrate it with a different frontend of your choice.
For full API documentation, see localhost:8100/docs after deployment.
If you want to see the API docs before deployment, check out the hosted docs here.
In the examples below, cookies are used as a mock auth method. For production, we recommend using JWT auth. Refer to the auth guide for production for more information. When using JWT auth, you will need to include the JWT in the
AuthorizationFirst, let's use the API to create an assistant. This should look something like:
import requests requests.post('http://127.0.0.1:8100/assistants', json={ "name": "bar", "config": {"configurable": {}}, "public": True }, cookies= {"opengpts_user_id": "foo"}).content
This is creating an assistant with name
"bar""foo"This should return something like:
b'{"assistant_id":"9c7d7e6e-654b-4eaa-b160-f19f922fc63b","name":"string","config":{"configurable":{}},"updated_at":"2023-11-20T16:24:30.520340","public":true,"user_id":"foo"}'
The config parameters allows you to set the LLM used, the instructions of the assistant and also the tools used.
{ "name": "bar", "config": { "configurable": { "type": "agent", "type==agent/agent_type": "GPT 3.5 Turbo", "type==agent/system_message": "You are a helpful assistant", "type==agent/tools": ["Wikipedia"] }, "public": True }
This creates an assistant with the name
"bar""You are a helpful assistant"Available tools names can be found in the AvailableTools class in backend/packages/gizmo-agent/gizmo_agent/tools.py Available llms can be found in GizmoAgentType in backend/packages/gizmo-agent/gizmo_agent/agent_types/init.py
Note: If a RAGBot assistant is created (
typechat_retrievalWe can now create a thread. Notably different from OpenAI's assistant API, we require starting the thread with an assistant ID.
import requests requests.post('http://127.0.0.1:8100/threads', cookies= {"opengpts_user_id": "foo"}, json={ "name": "hi", "assistant_id": "9c7d7e6e-654b-4eaa-b160-f19f922fc63b" }).content
This is creating a thread, named
"hi"This should return something like:
b'{"thread_id":"231dc7f3-33ee-4040-98fe-27f6e2aa8b2b","assistant_id":"9c7d7e6e-654b-4eaa-b160-f19f922fc63b","name":"hi","updated_at":"2023-11-20T16:26:39.083817","user_id":"foo"}'
We can check the thread, and see that it is currently empty:
import requests requests.get( 'http://127.0.0.1:8100/threads/231dc7f3-33ee-4040-98fe-27f6e2aa8b2b/state', cookies= {"opengpts_user_id": "foo"} ).content
b'{"values":[]}'
For RAGBot:
b'{"values":{"messages":[]}}'
Let's add a message to the thread!
import requests requests.post( 'http://127.0.0.1:8100/threads/231dc7f3-33ee-4040-98fe-27f6e2aa8b2b/state', cookies= {"opengpts_user_id": "foo"}, json={ "values": [{ "content": "hi! my name is bob", "type": "human", }] } ).content
For RAGBot:
{ "values": { "messages": [{ "content": "hi! my name is bob", "type": "human", }] } }
If we now run the command to see the thread, we can see that there is now a message on that thread
import requests requests.get( 'http://127.0.0.1:8100/threads/231dc7f3-33ee-4040-98fe-27f6e2aa8b2b/state', cookies= {"opengpts_user_id": "foo"} ).content
b'{"values":[{"content":"hi! my name is bob","additional_kwargs":{},"type":"human","example":false}],"next":[]}'
For RAGBot:
b'{"values":{"messages":[...]},"next":[]}'
We can now run the assistant on that thread.
import requests requests.post('http://127.0.0.1:8100/runs', cookies= {"opengpts_user_id": "foo"}, json={ "assistant_id": "9c7d7e6e-654b-4eaa-b160-f19f922fc63b", "thread_id": "231dc7f3-33ee-4040-98fe-27f6e2aa8b2b", "input": { "messages": [] } }).content
This runs the thread with the same id that we just created, with the assistant that we created, with no additional input messages (see below for how to add input messages).
If we now check the thread, we can see (after a bit) that there is a message from the AI.
import requests requests.get('http://127.0.0.1:8100/threads/231dc7f3-33ee-4040-98fe-27f6e2aa8b2b/state', cookies= {"opengpts_user_id": "foo"}).content
b'{"values":[{"content":"hi! my name is bob","additional_kwargs":{},"type":"human","example":false},{"content":"Hello, Bob! How can I assist you today?","additional_kwargs":{"agent":{"return_values":{"output":"Hello, Bob! How can I assist you today?"},"log":"Hello, Bob! How can I assist you today?","type":"AgentFinish"}},"type":"ai","example":false}],"next":[]}'
For RAGBot:
b'{"values":{"messages":[...]},"next":[]}'
We can also run the assistant on a thread and add new messages at the same time. Continuing the example above, we can run:
import requests requests.post('http://127.0.0.1:8100/runs', cookies= {"opengpts_user_id": "foo"}, json={ "assistant_id": "9c7d7e6e-654b-4eaa-b160-f19f922fc63b", "thread_id": "231dc7f3-33ee-4040-98fe-27f6e2aa8b2b", "input": { "messages": [{ "content": "whats my name? respond in spanish", "type": "human", }] } }).content
Then, if we call the threads endpoint after a bit we can see the human message - as well as an AI message - get added to the thread.
import requests requests.get('http://127.0.0.1:8100/threads/231dc7f3-33ee-4040-98fe-27f6e2aa8b2b/state', cookies= {"opengpts_user_id": "foo"}).content
b'{"values":[{"content":"hi! my name is bob","additional_kwargs":{},"type":"human","example":false},{"content":"Hello, Bob! How can I assist you today?","additional_kwargs":{"agent":{"return_values":{"output":"Hello, Bob! How can I assist you today?"},"log":"Hello, Bob! How can I assist you today?","type":"AgentFinish"}},"type":"ai","example":false},{"content":"whats my name? respond in spanish","additional_kwargs":{},"type":"human","example":false},{"content":"Tu nombre es Bob.","additional_kwargs":{"agent":{"return_values":{"output":"Tu nombre es Bob."},"log":"Tu nombre es Bob.","type":"AgentFinish"}},"type":"ai","example":false}],"next":[]}'
For RAGBot:
b'{"values":{"messages":[...]},"next":[]}'
One thing we can do is stream back responses. This works for both messages as well as tokens. Below is an example of streaming back tokens for a response.
import requests import json response = requests.post( 'http://127.0.0.1:8100/runs/stream', cookies= {"opengpts_user_id": "foo"}, json={ "assistant_id": "9c7d7e6e-654b-4eaa-b160-f19f922fc63b", "thread_id": "231dc7f3-33ee-4040-98fe-27f6e2aa8b2b", "input": { "messages": [{ "content": "have a good day!", "type": "human", }] } }) res = [] if response.status_code == 200: # Iterate over the response for line in response.iter_lines(): if line: # filter out keep-alive new lines string_line = line.decode("utf-8") # Only look at where data i returned if string_line.startswith('data'): json_string = string_line[len('data: '):] # Get the json response - contains a list of all messages json_value = json.loads(json_string) if "messages" in json_value: # Get the content from the last message # If you want to display multiple messages (eg if agent takes intermediate steps) you will need to change this logic print(json_value['messages'][-1]['content']) else: print(f"Failed to retrieve data: {response.status_code}")
This streams the following:
You You too You too! You too! If You too! If you You too! If you have You too! If you have any You too! If you have any other You too! If you have any other questions You too! If you have any other questions, You too! If you have any other questions, feel You too! If you have any other questions, feel free You too! If you have any other questions, feel free to You too! If you have any other questions, feel free to ask You too! If you have any other questions, feel free to ask. You too! If you have any other questions, feel free to ask. You too! If you have any other questions, feel free to ask.