Skip to main content

This python package help to interact with Generative AI - Large Language Models. It interacts with AIaaS LLM , AIaaS embedding , AIaaS Audio set of APIs to cater the request.

Project description

AIaaS Falcon Logo

AIaaS Falcon-Light

Installation | Quickstart |

Documentation Coverage

Description

AIaaS_Falcon_Light is Generative AI - Logical & logging framework support AIaaS Falcon library

:shield: Installation

Ensure you have the requests and google-api-core libraries installed:

pip install aiaas-falcon-light

if you want to install from source

git clone https://github.com/Praveengovianalytics/falcon_light && cd falcon_light
pip install -e .

Methods

Light Class

  • __init__ (config) Intialise the Falcon object with endpoint configs.
    Parameter:

    • config: A object consisting parameter:
      • api_key : API Key
      • api_name: Name for endpoint
      • api_endpoint: Type of endpoint ( can be azure, dev_quan, dev_full, prod)
      • url: url of endpoint (eg: http://localhost:8443/)
      • log_id: ID of log (Integer Number)
      • use_pii: Activate Personal Identifier Information Limit Protection (Boolean)
      • headers: header JSON for endpoint
      • log_key: Auth Key to use the Application
  • current_pii() Check current Personal Identifier Information Protection activation status

  • switch_pii() Switch current Personal Identifier Information Protection activation status

  • list_models() List out models available

  • initalise_pii() Download and intialise PII Protection.
    Note: This does not activate PII but initialise dependencies

  • health() Check health of current endpoint

  • create_embedding(file_path) Create embeddings by sending files to the API.
    Parameter:

    • file_path: Path to file
  • generate_text(query="", context="", use_file=0, model="", chat_history=[], max_new_tokens: int = 200, temperature: float = 0, top_k: int = -1, frequency_penalty: int = 0, repetition_penalty: int = 1, presence_penalty: float = 0, fetch_k=100000, select_k=4, api_version='2023-05-15', guardrail={'jailbreak': False, 'moderation': False}, custom_guardrail=None)
    Generate text using LLM endpoint. Note: Some parameter of the endpoint is endpoint-specific.
    Parameter:

    • query: a string of your prompt
    • use_file: Whether to take file to context in generation. Only applies to dev_full and dev_quan. Need to create_embedding before use.
    • model: a string on the model to use. You can use list_models to check for model available.
    • chat_history: an array of chat history between user and bot. Only applies to dev_full and dev_quan. (Beta)
    • max_new_token: maximum new token to generate. Must be integer.
    • temperature: Float that controls the randomness of the sampling. Lower values make the model more deterministic, while higher values make the model more random. Zero means greedy sampling.
    • top_k: Integer that controls the number of top tokens to consider.
    • frequency_penalty: Float that penalizes new tokens based on their frequency in the generated text so far.
    • repetition_penalty: Float that penalizes new tokens based on whether they appear in the prompt and the generated text so far.
    • presence_penalty: Float that penalizes new tokens based on whether they appear in the generated text so far
    • fetch_k: Use for document retrival. Include how many element in searching. Only applies when use_file is 1
    • select k: Use to select number of document for document retrieval. Only applies when use_file is 1
    • api_version: Only applies for azure endpoint
    • guardrail: Whether to use the default jailbreak guardrail and moderation guardrail
    • custom_guardrail: Path to custom guardrail .yaml file. The format can be found in sample.yaml
  • evaluate_parameter(config) Carry out grid search for parameter
    Parameter:

    • config: A dict. The dict must contain model and query. Parameter to grid search must be a list.
      • model: a string of model
      • query: a string of query
      • **other parameter (eg: "temperature":list(np.arange(0,2,0.5))
  • decrypt_hash(encrypted_data) Decret the configuration from experiment id. Parameter:

    • encrypted_data: a string of id

:fire: Quickstart

from aiaas_falcon import Falcon
model=Falcon(api_name="azure_1",protocol='https',host_name_port='example.com',api_key='API_KEY',api_endpoint='azure',log_key="KEY")
model.list_models()
model.generate_text_full(query="Hello, introduce yourself",model='gpt-35-turbo-0613-vanilla',api_version='2023-05-15')

Conclusion

AIaaS_Falcon_Light library simplifies interactions with the AIaaS Falcon, providing a straightforward way to perform various operations such as fact-checking and logging.

Authors

Google Colab

Badges

MIT License

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

aiaas_falcon_light-0.2.5.tar.gz (10.9 kB view hashes)

Uploaded Source

Built Distribution

aiaas_falcon_light-0.2.5-py3-none-any.whl (10.5 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page