Python Client Library#

Full documentation of the reka python client library.

Generation#

reka.chat(human=None, media_url=None, media_filename=None, conversation_history=None, retrieval_dataset=None, model_name='reka-flash', request_output_len=None, temperature=None, random_seed=None, runtime_top_k=None, runtime_top_p=None, frequency_penalty=None, presence_penalty=None, length_penalty=None, stop_words=None, extra_request_args=None)#

Chat endpoint.

Example usage:

import reka

reka.API_KEY = "APIKEY"

conversation_history = [
    {"type": "human", "text": "Hi, my name is John."},
    {"type": "model", "text": "Hi, I'm Reka's assistant."},
]
response = reka.chat(
    human="What was my name?",
    conversation_history=conversation_history,
)
print(response) # {"type": "model", "text": "Your name is John.\n\n"}
Parameters:

  • human (str | None) – latest message from human, this is optional if using conversation_history instead.

  • media_url (str | None) – an optional URL for the media (image, video, or audio) to chat about. This may only be set for the first turn (when conversation_history is empty). You can also send base64 media in the format data:image/{image_format};base64,{base64_image}

  • media_filename (str | None) – alternative to the media_url parameter, the location of a local file.``

  • conversation_history (List[HumanTurn | ModelTurn] | None) – list of dicts, where each dict has a key “type” indicating the speaker, either “human” or “model”, and a key “text” containing the message from the speaker. If not set, will default to an empty history. The first turn may also have “media_url” set.

  • retrieval_dataset (str | None) – Previously uploaded dataset to do retrieval on.

  • model_name (str) – Name of model. You can check available models with reka.get_models(). Defaults to flash.

  • request_output_len (int | None) – Completion length in tokens.

  • temperature (float | None) – Softmax temperature, higher is more diverse.

  • random_seed (int | None) – Seed to obtain different results.

  • runtime_top_k (int | None) – Keep only k top tokens when sampling.

  • runtime_top_p (float | None) – Keep only top p quantile when sampling.

  • frequency_penalty (float | None) – Penalize repetitions. 0 means no penalty.

  • presence_penalty (float | None) – Penalize repetitions. 0 means no penalty.

  • length_penalty (float | None) – Penalize short answers. 1 means no penalty.

  • stop_words (List[str] | None) – Optional list of words on which to stop generation.

  • extra_request_args (Dict[str, Any] | None) – Optional extra arguments to include in the request, useful for experimental API features.

Raises:

InvalidConversationError – if the conversation history is not valid.

Returns:

Return type:

ModelTurn

A dict containing {"type": "model", "text": <response from the model>}. If retrieval_dataset is set, then this will also contain "retrieved_chunks". :rtype: ModelTurn

Datasets#

Details of valid dataset formats are given in Dataset format.

reka.add_dataset(filepath, name, description=None)#

Upload a dataset to run jobs on it later.

NOTE: If the name is deemed inappropriate by the server, e.g. “../../etc/shadow”, it will be changed to a secure name, which is returned in the response.

Parameters:

  • filepath (str) – str, local path to a text file or a zipped collection of text files.

  • name (str) – str, what should the dataset be called

  • description (str | None) – Optional[str], optional metadata description

Returns:

Dictionary object representing what happened with the uploaded file.

Return type:

Dict[str, Any]

reka.list_datasets()#

List all datasets available to the user of the API_KEY.

Returns:

List of dataset names.

Return type:

List[str]

reka.delete_dataset(name)#

Delete a dataset with a given name for the user of this API_KEY.

Parameters:

name (str) – name of the dataset to delete.

Return type:

Dict[str, Any]

Returns: Dictionary object with keys name (str, the dataset name), ok (bool•), and ``info`` (*str).

Retrieval#

See the guide on retrieval for more information.

reka.prepare_retrieval(dataset_name)#

Prepare dataset_name for retrieval later, creating a prepare-retrieval job.

Parameters:

dataset_name (str) – name of previously uploaded dataset to prepare for retrieval

Raises:

  • RetrievalError if there is something wrong with retrieval preparation, e.g. –

  • no such dataset uploaded, or retrieval already prepared –

Returns:

ID of the prepare-retrieval job, to be used for tracking.

Return type:

str

reka.retrieval_job_status(job_id)#

Given a prepare retrieval job id, check the status.

Parameters:

job_id (str) – id of the prepare-retrieval job.

Returns:

The current state of the job.

Return type:

PrepareRetrievalStatusResponse

class reka.PrepareRetrievalStatusResponse(job_status, detail, history)#

Status of a prepare-retrieval job.

Parameters:

  • job_status (JobStatus) – The current status.

  • detail (str) – Further details, if any.

  • history (List[Dict[str, Any]]) –

is_complete()#

Whether the job has completed successfully.

Return type:

bool

is_done()#

Whether the job has completed, either succesfully or with an error.

Return type:

bool

is_error()#

Whether the job has ended with an error.

Return type:

bool

is_running()#

Whether the job is still running.

Return type:

bool