Refuel SDK Guide¶

Refuel SDK is a Python library to programmatically call Refuel's APIs. This guide describes how to install and use the Refuel SDK.

The SDK is currently an beta release and under rapid development. Please expect some sharp edges. Feedback and bug reports are always welcome! Shoot us an email at support@refuel.ai or ping us on our shared Slack.

Getting Started¶

Installation¶

Install the SDK using a package installer such as pip:

pip install refuel

Installing and using the SDK requires Python 3.6+.

Initializing a client session¶

Set the environment variable REFUEL_API_KEY. The SDK will read it during initialization, and use this value when sending requests:

import refuel


# Assuming you've set `REFUEL_API_KEY` in your env,
# init() will pick it up automatically
refuel_client = refuel.init()

Alternatively, you can supply the API key as a parameter during initializtion as shared below. In the cloud application, there is a top-level dropdown to select the project you’re working on currently. And this selection powers all the pages downstream (datasets, labeling tasks etc). The SDK allows you to do this by setting the project during initialization as well:

import refuel

options = {
    "api_key": "<YOUR API KEY>",
    "project": "<PROJECT NAME>",
}

refuel_client = refuel.init(**options)

Here’s the complete list of initialization options currently supported:

Option	Is Required	Default Value	Comments
`api_key`	Yes	None	Used to authenticate all requests to the API server
`project`	Yes	None	The name of the project you plan to use for the current session
`timeout`	No	60	Timeout in seconds
`max_retries`	No	3	Max number of retries for failed requests
`max_workers`	No	Num CPUs (os.cpu_count())	Max number of concurrent requests to the API server

Projects¶

These functions let you create a new project in your team’s Refuel account, or get information about projects already defined.

Get Projects¶

The get_projects API will return a list of all projects that belong to your team:

projects = refuel_client.get_projects()

You can also retrieve information about a specific project, either by name or by UUID

project = refuel_client.get_project(project='<PROJECT NAME>')

Create Project¶

refuel_client.create_project(
  project='<NEW PROJECT NAME>',
  description='<NEW PROJECT DESCRIPTION>',
)

Datasets¶

These functions let you upload/download a full dataset, or fetch rows and LLM labels within a dataset.

Get all current datasets¶

The get_datasets function will return a list of all datasets in the project:

import refuel

options = {
    "api_key": "<YOUR API KEY>",
    "project": "<PROJECT NAME>",
}

refuel_client = refuel.init(**options)

datasets = refuel_client.get_datasets()

Upload Dataset¶

This function lets you upload a local CSV file as a new dataset to Refuel.

import refuel

options = {
    "api_key": "<YOUR API KEY>",
    "project": "<PROJECT NAME>",
}

refuel_client = refuel.init(**options)

dataset = refuel_client.upload_dataset(
  file_path='<PATH TO CSV FILE>',
  dataset_name='<NEW DATASET NAME>',
  source='local|uri',
  wait_for_completion='True|False'
)

Some details about the function parameters:

Option	Is Required	Default Value	Comments
`file_path`	Yes	-	Path to the data you wish to upload
`dataset_name`	Yes	-	Unique name of the dataset being uploaded
`source`	Yes	local	Place where the file resides
`wait_for_completion`	No	False	Whether to poll for the completion of the dataset ingestion

Note: When uploading dataset from uri source, the file_path should be publicly accessible (eg. S3 presigned url) for Refuel to process it.

Download Dataset¶

This function create a snapshot of your dataset and generate presigned URL for secure download.

response = refuel_client.download_dataset(
  dataset='<DATASET NAME>',
  email_address='<YOUR EMAIL ADDRESS>',
)

This is an asynchronous workflow, and when the dataset is available for download, Refuel will send an email to the email address. Make sure the email address belongs to a valid user from your team. Depending on the size of the dataset, this export step can take a few minutes. Once the download link is created and emailed, it will be valid for 24 hours.

Adding items (rows) to an uploaded dataset¶

This function lets you append new rows to an existing dataset. Keep in mind that the dataset schema is decided during initial upload and is not updated here.

import refuel

options = {
    "api_key": "<YOUR API KEY>",
    "project": "<PROJECT NAME>",
}

refuel_client = refuel.init(**options)

new_items_to_add = [
  {"column1": "value1", "column2": "value2", ...},
  {"column1": "value3", "column2": "value4", ...},
  ...
]

refuel_client.add_items(dataset='<DATASET NAME>', items=new_items_to_add)

Querying items (rows) in a dataset¶

In addition to downloading the entire dataset, you can also fetch a list of items (rows) from the dataset as follows:

import refuel

options = {
    "api_key": "<YOUR API KEY>",
    "project": "<PROJECT NAME>",
}

refuel_client = refuel.init(**options)

items = refuel_client.get_items(
  dataset='<DATASET NAME>',
  max_items=20,
  offset=0
)

This function will return a pandas dataframe. Some details about the function parameters:

Option	Is Required	Default Value	Comments
`dataset`	Yes	-	Name of the dataset you want to query and retrieve items (rows) from
`max_items`	No	20	Max number of rows you want to fetch
`offset`	No	0	If this is set to a positive number, say N, then the first N rows will be skipped and the API will return “max_items” number of rows after skipping the first N rows.

Querying items, along with labels from a labeling task¶

get_items() also allows you to provide an optional parameter - a labeling task. When provided, the function will also include the task results (LLM labels, confidence and manually confirmed labels, if any) for the returned items.

items = refuel_client.get_items(
  dataset='<DATASET NAME>',
  task='<LABELING TASK NAME>',
  max_items=100
)

Applying sort ordering when querying items¶

By default, the API will use Refuel’s sort order (by decreasing order of diversity). You can use the order_by param to sort by any other columns in the dataset or by the label or confidence score from a labeling task.

Sort by dataset column

items = refuel_client.get_items(
  dataset='<DATASET NAME>',
  max_items=100,
  order_by=[{'field': '<COLUMN NAME TO SORT BY>', 'direction': '<ASC or DESC>'}],
)

Sort by label or confidence score from a labeling task. Note that this requires a task name and a subtask name to be specified. field can be either 'label' or 'confidence'.

items = refuel_client.get_items(
  dataset='<DATASET NAME>',
  task='<LABELING TASK NAME>',
  max_items=100,
  order_by=[{'field': 'confidence', 'direction': '<ASC or DESC>', 'subtask': '<SUBTASK NAME>'}],
)

You may have multiple dicts in the order_by list if you would like to sort by multiple columns (used in the case of ties). Some details about the keys for each dict in the order_by list:

Key	Is Required	Default Value	Description	Comments
`field`	Yes		The name of the column in the dataset to sort by	In addition to the columns in the dataset, the field can also be 'label' or 'confidence', if the task and subtask names are specified.
`direction`	No	`ASC`	The direction that you would like to sort the specified column by	Should be `ASC` or `DESC`
`subtask`	No	null	The name of the subtask for which you would like to sort by label or confidence	This should only be provided if the field is 'label' or 'confidence' and requires a task name to be specified in the function params.

Applying filters when querying items¶

In addition to sorting options, you can also define filters to only fetch items in the dataset that match a specific criteria. The SDK supports three types of filters:

Metadata Filters:

Filter based on the value of a specific column, for e.g. a filter such as “column = value” is defined as a Python dictionary with three keys:

field: This is the column on which you want to apply the filter
operator: This is the comparison operator
value: This is the value to compare to

Here’s an example:

items_filter = {
  'field': 'transaction_category',
  'operator': '=',
  'value': 'Food'
}

items = refuel_client.get_dataset_items(
  dataset='<DATASET NAME>',
  max_items=100,
  filters = [items_filter]
)

LLM output value/confidence filter:

Filter items based on the LLM output value or confidence score from a specific task configured in Refuel.

Here's a concrete example: Let's say you configured a classification task called Sentiment Analysis in your Refuel account, which has two subtasks (output fields):

(i) predicted_sentiment - the predicted sentiment

(ii) explanation - a one sentence explanation of why the LLM output the predicted sentiment as Positive or Negative for the item.

Here are a few filters we can define for this task:

"predicted sentiment is Positive":

{
  'field': 'llm_label',
  'subtask': 'predicted_sentiment'
  'operator': '=',
  'value': 'Positive'
}

"predicted sentiment confidence >= 80%":

{
  'field': 'confidence',
  'subtask': 'predicted_sentiment'
  'operator': '>=',
  'value': '0.8'
}

"predicted sentiment does not agree with the ground truth label (available in a column called ground_truth_sentiment)":

{
  'field': 'llm_label',
  'subtask': 'predicted_sentiment'
  'operator': '<>',
  'value': 'ground_truth_sentiment'
}

Semantic search filter:

Filter based on semantic similarity to a “query item” in a dataset. Semantic search queries are mapped to approximate nearest neighbor searches in an embedding space. These filters are defined as a Python dictionary with two keys:

operator: always set to "SIMILAR"
value: This is query item, identified by the uuid in the refuel_uuid column.

Here’s an example of how to define a semantic search filter. The results are returned in decreasing order of similarity score:

items_filter = {
  'operator': 'SIMILAR',
  'value': '020b434f-e1d7-4b0e-bfc1-beded36c806a'
}

items = refuel_client.get_dataset_items(
  dataset='<DATASET NAME>',
  task='<TASK NAME>'
  max_items=10,
  filters = [items_filter]
)

Here’s the complete list of filter operators that are currently supported:

Operator	Description
`>`	Greater than
`>=`	Greater than or equal to
`=`	Equals
`<>`	Not equal to
`<`	Less than
`<=`	Less than or equal to
`IS NULL`	True if field is undefined
`IS NOT NULL`	True if field is defined
`LIKE`	String matching: True if value is in field
`ILIKE`	String matching (case insensitive): True if value is in field
`NOT LIKE`	String does not match: True if value is not in field
`NOT ILIKE`	String does not match (case insensitive): True if value is not in field
`SIMILAR`	Semantic similarity filter operator

Labeling Tasks¶

These functions let you retrieve information about labeling tasks defined within a project, and start and cancel a task run.

Define a new Task¶

You can create a new task programmatically within a given project using the create_task function:

import refuel

options = {
    "api_key": "<YOUR API KEY>",
    "project": "<PROJECT NAME>",
}

refuel_client = refuel.init(**options)

refuel_client.create_task(
    task='<TASK NAME>',
    dataset='<DATASET NAME>',
    context = 'You are an expert at analyzing sentiment of an online review about a business...',
    fields = [
        {
          'name': '...',
          'type': '...',
          'guidelines': '...',
          'labels': [...],
          'input_columns': [...],
          'fallback_value': '...'
        },
        ...
    ],
    model = 'GPT-4 Turbo'
)

Some details about the various parameters you see in the function signature above:

Parameter	Is Required	Default Value	Comments
`task`	Yes	None	Name of the new task you're creating
`dataset`	Yes	None	Dataset (in Refuel) for which you are defining this task
`context`	Yes	None	Context is a high level description of the problem domain and the dataset that the LLM will be working with. It typically starts with something like 'You are and expert at ...'
`fields`	Yes	None	This is a list of dictionaries. Each entry in this list defines an output field generated in the task. See below for details about the schema of each field
`model`	No	team default	LLM that will be used for this task. If not specified, we will use the default LLM set for your team, e.g. GPT-4 Turbo

Next, let's take a look at the schema of each entry in the fields list above:

Parameter	Is Required	Default Value	Comments
`name`	Yes	None	Name of the output field, e.g. `llm_predicted_sentiment`
`type`	Yes	None	Type of output field. This is one of: [`classification`, `multilabel_classification`, `attribute_extraction`, `webpage_transform`, `web_search`]
`guidelines`	Yes	None	Output guidelines for the LLM for this field. Note that if the field type is a `web_search` type, the guidelines will be simply the query template
`labels`	Yes (for classification field types)	None	list of valid labels, this field is only required for classification type tasks
`input_columns`	Yes	None	Columns from the dataset to use as input when passing a "row" in the dataset to the LLM.
`ground_truth_column`	No	None	A column in the dataset that contains ground truth value for this field, if one exists. Note this is an optional parameter.
`fallback_value`	No	None	A fallback/default value that the LLM should generate for this field if a row cannot be processed successfully

Finally, here is the list of LLMs currently supported (use the model name as the parameter value):

Provider	Name
OpenAI	GPT-4 Turbo
OpenAI	GPT-4o
OpenAI	GPT-4o mini
OpenAI	GPT-4
OpenAI	GPT-3.5 Turbo
Anthropic	Claude 3.5 (Sonnet)
Anthropic	Claude 3 (Opus)
Anthropic	Claude 3 (Haiku)
Google	Gemini 1.5 (Pro)
Mistral	Mistral Small
Mistral	Mistral Large
Refuel	Refuel LLM-2
Refuel	Refuel LLM-2-small

Get Tasks¶

You can retrieve a list of all tasks within a given project as follows

tasks = refuel_client.get_tasks()

Start a Labeling Task Run¶

You can begin running a labeling task on a dataset with the following:

response = refuel_client.start_task_run(
  task='<TASK NAME>',
  dataset='<DATASET NAME>',
  num_items=100
)

This will kick off a bulk labeling run for the specified task and dataset, and label 100 items in the dataset. If num_items parameter is not specified, it will label the entire dataset.

Cancel an ongoing labeling task run¶

You can also cancel an ongoing labelling task with the same function as follows.

response = refuel_client.cancel_task_run(
  task='<TASK NAME>',
  dataset='<DATASET NAME>'
)

Get Task run status/progress¶

To check on the status of an ongoing labeling task run, you can use the following function

task_run = refuel_client.get_task_run(
  task='<TASK NAME>',
  dataset='<DATASET NAME>'
)

The task_run object has the following schema:

{
 'id': '...',
 'task_id': '...',
 'task_name': '...',
 'dataset_id': '...',
 'dataset_name': '...',
 'model_name': '...',   # This is the LLM used for the run
 'status': 'active',    # Status of the task run
 'metrics': [           # Metrics related to the task run
      {'name': 'num_labeled', 'value': ... },
      {'name': 'num_remaining', 'value': ...},
      {'name': 'time_elapsed_seconds', 'value': ...},
      {'name': 'time_remaining_seconds', 'value': ...},
 ]
}

status enum shows the current task run status. It can be one of the following values:

not_started: this is the starting state before the batch run has been kicked off
active: a batch task run is ongoing
paused: the batch run was paused before the full dataset was labeled.
failed: the batch run failed due to a platform error. This should ideally never happen
completed: the batch run was completed successfully

metrics is a list containing all metrics for the current task run. Currently the platform supports the following metrics:

num_labeled: number of rows from the dataset that have been labeled
num_remaining number of rows from the dataset that are remaining
time_elapsed_seconds: time (in seconds) since the task run started. This is only populated when the task run is active (since this metric is not valid when there is no active run).
time_remaining_seconds estimated time (in seconds) remaining for the task run to complete. This is only populated when the task run is active (since this metric is not valid when there is no active run).

Applications¶

Refuel allows you to deploy a labeling task as an application. Applications allow you to label data synchronously on demand, primarily for online workloads.

Deploy labeling application¶

To deploy an existing task as a labeling application, you can use the following function

import refuel

options = {
    "api_key": "<YOUR API KEY>",
    "project": "<PROJECT NAME>",
}

refuel_client = refuel.init(**options)

response = refuel_client.deploy_task(task='<TASK NAME>')

Get all labeling application¶

To get all labeling applications that are currently deployed, use the following function

applications = refuel_client.get_applications()

Label using a deployed application¶

You can use the deployed application for online predictions as follows:

inputs = [
  {'col_1': 'value_1', 'col_2': 'value_2' ...},
  {'col_1': 'value_1', 'col_2': 'value_2' ...},
]

response = refuel_client.label(application='<APPLICATION NAME>', inputs=inputs, explain=False)

Each element in inputs is a dictionary, with keys as names of the input columns defined in the task. For example, let's consider an application for sentiment classification called my_sentiment_classifier, with two input fields - source and text. You can use it as follows:

inputs = [
  {'source': 'yelp.com', 'text': 'I really liked the pizza at this restaurant.'}
]

response = refuel_client.label(application='my_sentiment_classifier', inputs=inputs)

response has the following schema:

refuel_output[i] contains the output for inputs[i]
refuel_fields is a list whose length is equal to the number of fields defined in the application. For example, let's say my_sentiment_classifier has just one field, sentiment. In this case the output will be:

{
  'application_id': '...',
  'application_name': 'my_sentiment_classifier',
  'refuel_output': [
    {'refuel_uuid': '...',
     'refuel_api_timestamp': '...',
     'refuel_fields':
        {
          'sentiment': {
            'label': ['positive'],
            'confidence': 0.9758
          }
        }
    }]
}

You can also set the optional explain parameter to True to get an explanation for why the provided label was returned. The explanation will be returned in the explanation field in the response, along with the label and confidence:

'sentiment': {
    'label': ['positive'],
    'confidence': 0.9758,
    'explanation': 'The model predicted a positive sentiment because the text contains positive words like "liked" and "good".'
  }

If you would only like to get an explanation for certain fields, you can optionally provide a list of field names in the explain_fields parameter for which you want an explanation returned. If explain_fields is provided, explanations will be returned regardless of whether explain is set to True or False. Here's an example of how to get explanations for the sentiment field in the my_sentiment_classifier application:

response = refuel_client.label(application='my_sentiment_classifier', inputs=inputs, explain=True, explain_fields=['sentiment'])

You can also set the optional telemetry parameter to True to get additional info such as the model, provider, and number of tokens used (prompt, output, and total) in the request. The telemetry data will be returned in the usage field in the response.

response = refuel_client.label(application='my_sentiment_classifier', inputs=inputs, telemetry=True)

Async Labeling¶

If you do not want to wait for the labeling to be completed, you can instead use the method alabel with the exact same parameters as with label. This will submit the inputs for labeling with Refuel and returns the refuel_uuid to get the labeled item back.

response = refuel_client.alabel(application='my_sentiment_classifier', inputs=inputs)

The output will be:

{
  'application_id': '...',
  'application_name': 'my_sentiment_classifier',
  'refuel_output': [
    {'refuel_uuid': '...',
     'refuel_api_timestamp': '...',
     'uri': '...'
    }]
}

You can eith use the refuel_uuid from the output to get the labeled item back using get_labeled_item method.

response = refuel_client.get_labeled_item(application='my_sentiment_classifier', refuel_uuid='dcbb0266-aeaa-4c0e-87b3-4341b5f3b7bc')

You can also directly call the returned uri to get the labeled item back.

Async Labeling is also useful when the input data is large or the LLM generates large amount of data which can lead to timeouts on Refuel's Application.

The SDK allows users to log feedback for online predictions. When logging predictions, it is important to identify the input request for which you're logging feedback using refuel_uuid from the response above:

label = {'sentiment': ['negative']}
refuel_client.feedback(application='my_sentiment_classifier', refuel_uuid='...', label=label)

Any row with logged feedback will appear with the verified check mark ("✓") in the cloud application.

Finetuning¶

Refuel allows you to finetune a model based on all the human reviewed data and optionally data labeled by an LLM. Finetuned models allow you to reduce labeling cost and latency while achieving similar, and in some cases better, performance than LLMs like GPT-4.

Starting a finetuning run¶

To start a model finetuning run, you can use the following function:

import refuel

options = {
    "api_key": "<YOUR API KEY>",
    "project": "<PROJECT NAME>",
}

refuel_client = refuel.init(**options)

hyperparameters = {"num_epochs": 1}
datasets = ["dataset_0_id", "dataset_1_id"]
response = refuel_client.finetune_model(task_id='<TASK ID>', model='<BASE MODEL>', hyperparameters=hyperparameters, datasets=datasets)

Supported Base Models: ['refuel-llm-v2-large', 'refuel-llm-v2-small'] Supported hyperparameters: ['lora_r', 'lora_alpha', 'lora_dropout', 'weight_decay', 'learning_rate', 'cosine_min_lr_ratio']

Get all finetuned models¶

You can retrieve a list of all models within a given project as follows

models = refuel_client.get_finetuned_models(task_id='<TASK ID>')

Cancel a finetuning run¶

You can also cancel an ongoing finetuning run as follows.

response = refuel_client.cancel_finetuning(model_id='<FINETUNING MODEL ID>')