SambaParse

SambaParse is a Python library that simplifies the process of extracting and processing unstructured data using the Unstructured.io API. It provides a convenient wrapper around the Unstructured.io CLI tool, allowing you to ingest data from various sources, perform partitioning, chunking, embedding, and load the processed data into a vector database. It's designed to be used within AI Starter kits and SN Apps, unifying our data ingestion and document intelligence platform. This allows us to keep our code base centralized for data ingestion kits.

Prerequisites

Before using SambaParse, make sure you have the following:

Docker installed on your machine (or access to another API server)
An Unstructured.io API key

Before using SambaParse, make sure you have the following:

Create a .env file in the ai-starter-kit root directory (not in the parsing folder root):
```
UNSTRUCTURED_API_KEY=your_api_key_here
```

Setup

Pre Reqs

Using pyenv to manage virtualenv's is recommended Mac install instructions. See pyenv-virtualenv repo for more detailed instructions.

brew install pyenv-virtualenv

Create a python venv using python version 3.10.12

pyenv install 3.10.12
pyenv  virtualenv 3.10.12 sambaparse
pyenv activate sambaparse

Clone the ai-starter-kit repo and cd:

git clone https://github.com/sambanova/ai-starter-kit

cd into utils/parsing and pip install the requirements
```
pip install -r requirements.txt
```
cd into the unstructured-api foder and Install the unstructured-api make-file:
```
cd  unstructured-api
```
Run
```
make install
```
Run The Web Server:
```
make run-web-app
```
This script will start the Unstructured API server using the specified API key and expose it on port 8005.
- Alternatively, if you have another Unstructured API server running on a different instance, make sure to update the partition_endpoint and unstructured_port values in the YAML configuration file accordingly.

Usage

Import the SambaParse class from the ai-starter-kit library:
```
from utils.parsing.sambaparse import SambaParse
```

Create a YAML configuration file (e.g., config.yaml) to specify the desired settings for the ingestion process. Here's the configuration for use cases 1 and 2 ie local files and folders:

processor:
verbose: True
output_dir: './output'
num_processes: 2

sources:
  local:
    recursive: True
  confluence:
    api_token: 'your_confluence_api_token'
    user_email: '[email protected]'
    url: 'https://your-confluence-url.atlassian.net'
  github:
    url: 'owner/repo'
    branch: 'main'
  google_drive:
    service_account_key: 'path/to/service_account_key.json'
    recursive: True
    drive_id: 'your_drive_id'

partitioning:
  pdf_infer_table_structure: True
  skip_infer_table_types: []
  strategy: 'auto'
  hi_res_model_name: 'yolox'
  ocr_languages: ['eng']
  encoding: 'utf-8'
  fields_include: ['element_id', 'text', 'type', 'metadata', 'embeddings']
  flatten_metadata: False
  metadata_exclude: []
  metadata_include: []
  partition_endpoint: 'http://localhost'
  unstructured_port: 8005
  partition_by_api: True

chunking:
  enabled: True
  strategy: 'basic'
  chunk_max_characters: 1500
  chunk_overlap: 300

embedding:
  enabled: False
  provider: 'langchain-huggingface'
  model_name: 'intfloat/e5-large-v2'

destination_connectors:
  enabled: False
  type: 'chroma'
  batch_size: 80
  chroma:
    host: 'localhost'
    port: 8004
    collection_name: 'snconf'
    tenant: 'default_tenant'
    database: 'default_database'
  qdrant:
    location: 'http://localhost:6333'
    collection_name: 'test'

additional_processing:
  enabled: True
  extend_metadata: True
  replace_table_text: True
  table_text_key: 'text_as_html'
  return_langchain_docs: True
  convert_metadata_keys_to_string: True

Make sure to place the config.yaml file in the desired folder.

Create an instance of the SambaParse class, passing the path to the YAML configuration file:
```
sambaparse = SambaParse('path/to/config.yaml')
```
Use the run_ingest method to process your data:

For a single file:

source_type = 'local' 
input_path = 'path/to/your/file.pdf' 
additional_metadata = {'key': 'value'} 
texts, metadata_list, langchain_docs = sambaparse.run_ingest(source_type, input_path=input_path, additional_metadata=additional_metadata)

For a folder:

source_type = 'local' 
input_path = 'path/to/your/file.pdf' 
additional_metadata = {'key': 'value'} 
texts, metadata_list, langchain_docs = sambaparse.run_ingest(source_type, input_path=input_path, additional_metadata=additional_metadata)

For Confluence:

source_type = 'confluence' 
additional_metadata = {'key': 'value'} 
texts, metadata_list, langchain_docs = sambaparse.run_ingest(source_type, additional_metadata=additional_metadata)

Note that for conflence you must enable embedding and destinatation connectors automatically ie Chroma and turn off additional processing (ie langchain), an example yaml to do that is below

  processor:
  verbose: True
  output_dir: './output'
  num_processes: 2

  sources:
    local:
      recursive: True
    confluence:
      api_token: 'your_confluence_api_token'
      user_email: '[email protected]'
      url: 'https://your-confluence-url.atlassian.net'
    github:
      url: 'owner/repo'
      branch: 'main'
    google_drive:
      service_account_key: 'path/to/service_account_key.json'
      recursive: True
      drive_id: 'your_drive_id'

  partitioning:
    pdf_infer_table_structure: True
    skip_infer_table_types: []
    strategy: 'auto'
    hi_res_model_name: 'yolox'
    ocr_languages: ['eng']
    encoding: 'utf-8'
    fields_include: ['element_id', 'text', 'type', 'metadata', 'embeddings']
    flatten_metadata: False
    metadata_exclude: []
    metadata_include: []
    partition_endpoint: 'http://localhost'
    unstructured_port: 8005
    partition_by_api: True

  chunking:
    enabled: True
    strategy: 'basic'
    chunk_max_characters: 1500
    chunk_overlap: 300

  embedding:
    enabled: True
    provider: 'langchain-huggingface'
    model_name: 'intfloat/e5-large-v2'

  destination_connectors:
    enabled: True
    type: 'chroma'
    batch_size: 80
    chroma:
      host: 'localhost'
      port: 8004
      collection_name: 'snconf'
      tenant: 'default_tenant'
      database: 'default_database'
    qdrant:
      location: 'http://localhost:6333'
      collection_name: 'test'

  additional_processing:
    enabled: False
    extend_metadata: True
    replace_table_text: True
    table_text_key: 'text_as_html'
    return_langchain_docs: True
    convert_metadata_keys_to_string: True

In addition for confluence you will need to have a Chroma Server running on port 8004, you can do this by running the docker command below

docker run -d --rm --name chromadb -v ./chroma:/chroma/chroma -e IS_PERSISTENT=TRUE -e ANONYMIZED_TELEMETRY=TRUE -p 8004:8000 chromadb/chroma:latest

The run_ingest method returns a tuple containing the extracted texts, metadata, and LangChain documents (if return_langchain_docs is set to True in the configuration).

Process the returned data as needed:
- texts: A list of extracted text elements from the documents.
- metadata_list: A list of metadata dictionaries for each text element.
- langchain_docs: A list of LangChain Document objects, which combine the text and metadata.

Configuration Options

The YAML configuration file allows you to customize various aspects of the ingestion process. Here are some of the key options:

processor: Settings related to the processing of documents, such as the output directory and the number of processes to use.
sources: Configuration for different data sources, including local files, Confluence, GitHub, and Google Drive.
partitioning: Options for partitioning the documents, including the strategy, OCR languages, and API settings.
chunking: Settings for chunking the documents, such as enabling chunking, specifying the chunking strategy, and setting the maximum chunk size and overlap.
embedding: Options for embedding the documents, including enabling embedding, specifying the embedding provider, and setting the model name.
additional_processing: Configuration for additional processing steps, such as extending metadata, replacing table text, and returning LangChain documents.

Make sure to review and modify the configuration file according to your specific requirements.