Skip to content

A lib of pre and post processing filters for LLM applications

Notifications You must be signed in to change notification settings

piratos/llmfilters

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

11 Commits
 
 
 
 
 
 

Repository files navigation

LLM Filters

LLM Filters is a pipeline library for processing and filtering text data that go in and out from LLM engines. It provides a framework for building custom filters, and includes managers for integrating with messaging systems such as RabbitMQ and Kafka.

When building an LLM based app, you want to control data going into the model, like the length, profanity removal, fact checking etc. and you want to control data out, to verify whether the LLM is hallucinating, fact checking etc.

the building blocks are Filterblocks, the message is passed through a set of filterblocks and each one of them has a couple of possible actions:

  • Modify the message (like removing profanity)
  • Passing the message to the next block (with or without modifying it)
  • Halting the pipeline and preventing the message to go to the LLM
  • Adding metadata to the message, which can be used by other filterblocks or can be used to prepare the final prompt going to the user or to the LLM.

Getting Started

  1. Install the lib pip install git+https://github.com/piratos/llmfilters
  2. Create a pipeline file or use one from the Examples folder (see below)
  3. Run the pipeline using the utility llmfilters-run --pipeline-file path/to/pipeline.yaml

Usage

This lib should be imported, augmented with needed blocks then used inside the LLM app/infrastructure rather then being executed in standalone way.

Project Structure

The project structure is as follows:

  • llmfilters/ is the main package directory.
  • blocks/ directory contains the custom filter blocks implemented.
  • managers/ directory contains the managers for integrating with RabbitMQ and Kafka messaging systems.
  • models directory contains the objects structure used by the filters, mainly the message format.
  • pipeline.yaml is the YAML configuration file that defines the pipeline structure.

Blocks

The following custom filter blocks are available:

FilterBlock

The FilterBlock is a base class for creating custom filters. It provides methods for applying changes, refusing input, and handling refused input.

To create a custom filter block, you can inherit from the FilterBlock class and implement the required methods.

Managers

The following managers are provided for integrating with messaging systems:

RabbitMQManager

The RabbitMQManager allows processing text data using the LLM pipeline with RabbitMQ as the messaging system. It connects to a RabbitMQ server, consumes input events, processes them through the pipeline, and produces output events.

KafkaManager

The KafkaManager allows processing text data using the LLM pipeline with Kafka as the messaging system. It connects to a Kafka broker, consumes input events, processes them through the pipeline, and produces output events.

Example with simple manager

Simple manager is useful to test the pipeline rather than using it in production

from llmfilters.managers import SimpleManager

def run_manager():

    config_file = "pipeline.yaml"
    manager = SimpleManager(manager_config, config_file)
    manager.run()

if __name__ == '__main__':
    run_manager()

Pipeline Configuration

The pipeline configuration is defined in the pipeline.yaml file. It specifies the sequence of filter blocks to be applied in the pipeline. Each block is defined with its type and parameters.

Example pipeline.yaml configuration:

manager:
  type: simple

blocks:
- type: llmfilters.blocks.FilterBlock1
  params:
    param1: value1
    param2: value2
- type: llmfilters.blocks.FilterBlock2
  params:
    param1: value3
    param2: value4

In this example, two FilterBlock instances are defined in the pipeline. Each block can have specific parameters based on the requirements. See examples folder for examples of piplines

The manager entry exposes the type and configuration of the manager handling the events

Happy filtering!

About

A lib of pre and post processing filters for LLM applications

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages