Indexify Extractors

Overview

Extractors are modules that give Indexify data processing capabilities such as metadata or embedding extraction from document, videos and audio. This repository hosts a collection of extractors for Indexify.

For the main Indexify project, visit: Indexify Main Repository.

Available Extractors

We have built some extractors based on demand from our users. You can write a new or a custom extractor for your use-case too, instructions for writing new extractors are below.

• Embedding Extractors
• Video Extractors
• Invoice Extractors
• Audio Extractors
• PDF Extractors

Usage

Install

pip install indexify-extractor-sdk

List Available extractors

indexify-extractor list

Download an Extractor

Find the name of the extractor you want.

indexify-extractor download tensorlake/minilm-l6

Load and Run in Notebook or Python Applications

from indexify_extractor_sdk import load_extractor, Content
extractor, config_cls = load_extractor("indexify_extractors.minilm-l6.minilm_l6:MiniLML6Extractor")
content = Content.from_text("hello world")
out = extractor.extract(content)

Extractors can be parameterized when they are called. The input parameters are Pydantic Models. Inspect the config class programatically or in the docs of the corresponding extractor -

ex, config = load_extractor("indexify_extractors.chunking.chunk_extractor:ChunkExtractor")
config.schema()
#{'properties': {'overlap': {'default': 0, 'title': 'Overlap', 'type': 'integer'}, 'chunk_size': {'default': 100, 'title': 'Chunk Size', 'type': 'integer'}, 'text_splitter': {'default': 'recursive', 'enum': ['char', 'recursive', 'markdown', 'html'], 'title': 'Text Splitter', 'type': 'string'}, 'headers_to_split_on': {'default': [], 'items': {'type': 'string'}, 'title': 'Headers To Split On', 'type': 'array'}}, 'title': 'ChunkExtractionInputParams', 'type': 'object'}

Extract Locally on shell -

indexify-extractor run-local indexify_extractors.minilm-l6.minilm_l6:MiniLML6Extractor --text "hello world" // or --file

Run Extractors as a Service for Continous Extraction and Indexing with Indexify Server

To run the extractor with Indexify's control plane such that it can continuously extract from content -

indexify-extractor join-server --coordinator-addr localhost:8950 --ingestion-addr localhost:8900

The coordinator-addr and ingestion-addr above are the default addresses exposed by the Indexify server to get extraction instructions and to upload extracted data, they can be configured in the server configuration.

Build a new Extractor

If want to build a new extractor to give Indexify new data processing capabilities you can write a new extractor by cloning this repository - https://github.com/tensorlakeai/indexify-extractor-template

Clone the template

curl https://codeload.github.com/tensorlakeai/indexify-extractor-template/tar.gz/main | tar -xz  indexify-extractor-template-main

Implement the extractor interface

class InputParams(BaseModel):
    a: int = 0
    b: str = ""


class MyExtractor(Extractor):
    name = "yourorgname/myextractor"
    description = "Description of the extractor goes here."
    system_dependencies = []
    input_mime_types = ["text/plain"]

    def __init__(self):
        super().__init__()

    def extract(self, content: Content, params: InputParams) -> List[Union[Feature, Content]]:
        return [
            Content.from_text(
                text="Hello World", features=[Feature.embedding(values=[1, 2, 3])]
            ),
            Content.from_text(
                text="Pipe Baz", features=[Feature.embedding(values=[1, 2, 3])]
            ),
            Content.from_text(
                text="Hello World",
                features=[Feature.metadata({"key": "value"})],
            ),
        ]

    def sample_input(self) -> Tuple[Content, Type[BaseModel]]:
        Content.from_text(text="Hello World")

All the Python dependencies of the extractor goes into requirements.txt file adjacent to the extractor file.

Once you have developed the extractor you can test the extractor locally by running the indexify-extractor run-local command as described above.

Test and Deploy the extractor

You can test your extractor without running the Indexify server!

ex, config = load_extractor("custom_extractor:MyExtractor")
config.schema()
ex.extract(Content(...), config(...))# or ignore config if you don't have config

Run the extractor on shell

indexify-extractor run-local custom_extractor:MyExtractor --text "hello world" // or --file /path to file

Install your Extractor

You can install your extractor locally

indexify-extractor install-local custom_extractor:MyExtractor

When you are ready to deploy the extractor in production, package the extractor and deploy as many instances you want on your cluster for parallelism, and point it to the indexify server.

indexify-extractor join-server --coordinator-addr localhost:8950 --ingestion-addr localhost:8900

Package the Extractor

Once you build a new extractor, and have tested it and it's time to deploy this in production, you can build a container with the extractor -

indexify-extractor package indexify_extractors.<folder_name>.custom_extractor:MyExtractor

If you want to package an extractor in a container that support Nvidia CUDA GPU, you can pass the --gpu flag to the package command.

Running Your packaged extractor

To run your packaged extractor image you can run the following command

docker run ExtractorImageName indexify-extractor join-server --coordinator-addr=host.docker.internal:8950 --ingestion-addr=host.docker.internal:8900

If you have a GPU enabled extractor, you might need to set up your machine to support running the container with the GPU. This might involve installing the Nvidia Container Toolkit and setting up the Nvidia runtime for Docker. You can find more information on how to do this in the Nvidia Container Toolkit Documentation.

Finally, to run your GPU enabled extractor, you can add the --gpus all flag to the docker run command.