Interact with bee-api through OpenAI Python SDK
⚠️ COMPATIBILITY DISCLAIMER⚠️ The bee-api is designed to overlap with OpenAI API to a degree that is sufficitent for most use-cases, however some types and operations are not 100% compatible, see the section on OpenAI Compatibility below. The examples provided in this repository are regularly tested to be working, however we will never guarantee full compatibility with OpenAI. Please create an issue if you stumbled upon non-compatible behavior that is blocking you.
We are using purely OpenAI SDK, so the only required package is openai
.
pip install openai
Make sure to have the following environment variables configured, see example.env:
BEE_API=http://localhost:4000
BEE_API_KEY=sk-testkey
# This is required to prevent some pydantic serialization errors
DEFER_PYDANTIC_BUILD=false
import os
from openai import OpenAI
# Instantiate OpenAI client with Bee credentials from env
client = OpenAI(base_url=f'{os.getenv("BEE_API")}/v1', api_key=os.getenv("BEE_API_KEY"))
# Create assistant
assistant = client.beta.assistants.create(
model="meta-llama/llama-3-1-70b-instruct", tools=[{"type": "code_interpreter"}]
)
# Create a thread with user message
question = "What is the unicode character U+1F41D? Create a haiku about it."
thread = client.beta.threads.create(messages=[{"role": "user", "content": question}])
# Create a run and wait for completion
run = client.beta.threads.runs.create_and_poll(thread_id=thread.id, assistant_id=assistant.id)
assert run.status == "completed"
# List messages and get an answer
messages = client.beta.threads.messages.list(thread_id=thread.id)
print("Answer:", messages.data[0].content[0].text.value)
If you want to run the examples in this repository, install all dependencies using poetry.
# Install dependencies
poetry install
Create a correct .env
file.
# Create .env file
cp example.env .env
# Insert your API key
open .env
Run examples through poetry
poetry run python -m examples.basic_usage
Here are the important differences from the official OpenAI SDK.
from openai import OpenAI
client: OpenAI = ...
# OpenAI
# Assistants
client.beta.assistants
client.beta.threads
client.beta.threads.messages
client.beta.threads.runs
client.beta.threads.runs.steps
client.beta.threads.runs.stream # early stage event compatibility
client.beta.vector_stores
client.beta.vector_stores.files
## Not supported:
# client.beta.vector_stores.file_batches
# Files
client.files
Extensions can be called using the low-level OpenAI client methods (get, post, put, delete). See the definition of a custom tool in examples/custom_tool.py as an example.
from openai import OpenAI, BaseModel
client: OpenAI = ...
# Tools
client.get('/tools', cast_to=BaseModel) # list tools
client.post('/tools', cast_to=BaseModel) # create tool
client.post('/tools/:tool_id', cast_to=BaseModel) # update tool
client.delete('/tools/:tool_id', cast_to=BaseModel) # delete tool
# Observe
client.get('/threads/:thread_id/runs/:run_id/trace') # Get trace ID for a run
Observe API module is designed to provide full trace of everything that happened during a run.
You can obtain the trace using a special /observe
endpoint, here is a brief example, for full code
see examples/download_trace.py:
import os
from openai import OpenAI, BaseModel
# Normal bee client
bee_client = OpenAI(base_url=f'{os.getenv("BEE_API")}/v1', api_key=os.getenv("BEE_API_KEY"))
thread = ...
run = ...
trace_info = bee_client.get(f"/threads/{thread.id}/runs/{run.id}/trace", cast_to=BaseModel)
# (!) Note different base_url
observe_client = OpenAI(base_url=f'{os.getenv("BEE_API")}/observe', api_key=os.getenv("BEE_API_KEY"))
# Get trace
params = {"include_tree": True, "include_mlflow": True}
trace = observe_client.get(f"/trace/{trace_info.id}", options={"params": params}, cast_to=BaseModel)
Use the official OpenAI documentation with caution (see caveats below), here are links to the relevant topics:
- streaming events are not fully compatible yet, the "With streaming" portions of OpenAI documentation will not work as expected (for example in file search)
- some features are not implemented:
- vector store file batches -
client.beta.vector_stores.file_batches
- adding message attachment to
file_search
without previously embedding the file in a thread vector store, see examples/vector_store.py - ... and more
- vector store file batches -
- some type unions are extended so the data returned does not match the original openai models:
- you may see pydantic warningns during serialization, you can avoid this by using
warnings="none"
when dumping a model, for exampleassistant.model_dump(warnings="none")
- you must set the
DEFER_PYDANTIC_BUILD=false
environment variable before all imports, if you see an error similar to 'MockValSer' object cannot be converted to 'SchemaSerializer', you are probably missing this configuration
- you may see pydantic warningns during serialization, you can avoid this by using
This is an open-source project and we ❤️ contributions.
If you'd like to contribute to Bee, please take a look at our contribution guidelines.