assistants-api

DISCORD

Video Demo

Replicate and improve the OpenAI Assistants API Note: currently support client openai==1.26.0 (excluding custom tools like web_retriever) or use our fork by pip install git+https://github.com/OpenGPTs-platform/openai-python.git

Video Demo (full OpenGPTs-platform)

Architecture

Quickstart

0. Clone the repo git clone https://github.com/OpenGPTs-platform/assistants-api.git and navigate into assistants-apidirectory
1. Create a copy of .env.example and name it .env. Fill in necessary values.
2. Start docker-compose docker-compose -f .\docker-compose.dev.yml up
3. Its running 🥳!
4. In a new directory and environment, install the openai client fork with pip install IPython git+https://github.com/OpenGPTs-platform/openai-python.git, and try it with the following demo (NOTE: update YOUR_FILE_PATH with your file path that you want to test retrieval with). Also it may now

from openai import OpenAI
client = OpenAI(
    base_url="http://localhost:8000",
    api_key="NO_KEY_NEEDED",
)

# Upload file to the server
file = client.files.create(
    file=open('YOUR_FILE_PATH', 'rb'), # Input your file path (currently accepts .txt and .pdf files)
    purpose='assistants'
)

# Create vector store with the file id
vs = client.beta.vector_stores.create(
    name='my_info',
    file_ids=[file.id]
)

# Create an assistant with the vector store passed in
asst = client.beta.assistants.create(
    name="Demo Assistant",
    instructions="Always start your responses for with the word 'APPLE'",
    model="gpt-4-turbo",
    tools=[{"type": "file_search"}, {"type": "web_retrieval"}],
    tool_resources={
        "file_search": {
            "vector_store_ids": [vs.id]
        }
    },
)

# Create a thread with or without messages
thr = client.beta.threads.create(
    messages=[
        {
            "role": "user",
            "content": "I am curious what is in the file I provided"
        }
    ],
)

# Execute the run (Adds run to RabbitMQ for to be dequeued and processed by run_executor_worker)
run = client.beta.threads.runs.create(
    thread_id=thr.id,
    assistant_id=asst.id
)

# Poll the response untill complete (streaming not yet supported)
from IPython.display import clear_output
import time
while run.status not in ['completed', 'failed']:
    time.sleep(1)
    clear_output(wait=True)
    run = client.beta.threads.runs.retrieve(thread_id=thr.id, run_id=run.id)
    print("RUN STATUS:\n",run.status)
    messages = client.beta.threads.messages.list(thread_id=thr.id, order='desc')
    print("THREAD MESSAGES:\n",messages.model_dump_json(indent=2))

More Comprehensive Demo

assistants_api

View full Figma spec

Handle the business logic (store and retrieve data, store files, enque runs) for the Assistants API according to the official OpenAI OpenAPI specification.

The OpenAI OpenAPI specification is the source of truth for this API.

run_executor_worker

View full Figma spec

Agent that executes runs according to CoALA architecture and ReAct prompting strategy.

Major Objectives

1. Function calling for run_executor_worker using Mistral-7B-Instruct-v0.3
2. Connect chat-ui
3. Optimize prompting in run_executor_worker
4. Open-source web_retrieval and add annotations to messages for citation purposes

Helper "Assistants API" Dev Assistant GPT

Helper assistant for developing the "Assistants API". Normally conversation will flow like so:

Human: Lets work on /assistants GET endpoint, begin with a test. Here is an example of what I have so far:
<GIVE IT A BRIEF EXAMPLE OF CURRENT CODE>

Assistant: <RESPONDS WITH TEST CODE>

Human: Ok lets move on to the endpoint. Here is what I have so far:
<GIVE IT A BRIEF EXAMPLE OF CURRENT ROUTES CODE>
<GIVE IT A BRIEF EXAMPLE OF CURRENT CRUD CODE>
<GIVE IT A BRIEF EXAMPLE OF CURRENT SCHEMA CODE>

Assistant: <RESPONDS WITH RELEVANT CODE>

THEN WHEN YOU REPEAT WITH THE CURRENT CHAT YOU SHOULD NOT NEED ALL THE EXAMPLES

Instruction

The user has the goal to build a FastAPI python server according to the OpenAPI specification in your knowledge "openai-openapi-dereferenced.json". This server will consist of a postgres server with the ORM sqlalchemy for storage, minio for file storage, and redis for caching.

Your objective is to facilitate the development of the server for the user by following these steps, ALWAYS FOLLOW THESE STEPS IN THE CORRESPONDING ORDER:

1. Using code_interpreter tool download "openai-openapi-dereferenced.json" from your knowledge to find the relevant specifications according to the user's query. Programmatically navigate the JSON. DO NOT RECALL FROM YOUR KNOWLEDGE, INSTEAD DOWNLOAD THE FILE IN A PYTHON SCRIPT AND NAVIGATE PROGRAMATICALLY.
2. Asking the user questions if more information is needed
3. Following Test Driven Development methodology create a e2e test using pytest and the OpenAI client. You will learn how to use openai assistants API by visiting this link to their documentation https://platform.openai.com/docs/assistants/overview with web browsing tool. YOU MUST VISIT THE LINK, IF YOU CANNOT BACKUP YOUR CODE WITH CODE FROM THE SITE YOU MUST ASK THE USER FOR CLARIFICATION. (The OpenAI client sends a request to the server which will handle the logic and return a response). YOU MUST WRITE THE TEST.
4. Asking the user questions if more information is needed.
5. Creating a plan for execution and providing the code.

Knowledge

Add openai-openapi-dereferenced.json