Parsing & Transformation in LlamaCloud

Once data is loaded from a Data Source, it is pre-processed before being sent to the Data Sink. There are many pre-processing parameters that can be tweaked to optimize the downstream retrieval performance of your index. While LlamaCloud sets you up with reasonable defaults, you can dig deeper and customize them as you see fit for your specific use case.

Parser Settings

A key step of any RAG pipeline is converting your input file into a format that can be used to generate a vector embedding. There are many parameters that can be used to tweak this conversion process to optimize for your use case. LlamaCloud sets you up from the start with reasonable defaults for your parsing configurations, but also allows you to dig deeper and customize them as you see fit for your specific application.

Transformation Settings

The transform configuration is used to define the transformation of the data before it is ingested into the Index. it is a JSON object which you can choose between two modes auto and advanced and as the name suggests, the auto mode is handled by LlamaCloud which uses a set of default configurations and the advanced mode is handled by the user with the ability to define their own transformation.

Auto Mode

You can set the mode by passing the transform_config as below on index creation or update.

transform_config = {
    "mode": "auto"
}

Also when using the auto mode, you can configure the chunk size being used for the transformation by passing the chunk_size and chunk_overlap parameter as below.

transform_config = {
    "mode": "auto",
    "chunk_size": 1000,
    "chunk_overlap": 100
}

Advanced Mode

The advanced mode provides a variation of configuration options for the user to define their own transformation. The advanced mode is defined by the mode parameter as advanced and the segmentation_config and chunking_config parameters are used to define the segmentation and chunking configuration respectively.

transform_config = {
    "mode": "advanced",
    "segmentation_config": {
        "mode": "page",
        "page_separator": "\n---\n"
    },
    "chunking_config": {
        "mode": "sentence",
        "separator": " ",
        "paragraph_separator": "\n"
    }
}

Segmentation Configuration

The segmentation configuration uses the document structure and/or semantics to divide the documents into smaller parts following natural segmentation boundaries. The segmentation_config parameter include three modes none, page and element.

None Segmentation Configuration

The none segmentation configuration is used to define no segmentation.

transform_config = {
    "mode": "advanced",
    "segmentation_config": {
        "mode": "none"
    }
}

Page Segmentation Configuration

The page segmentation configuration is used to define the segmentation by page and the page_separator parameter is used to define the separator, which will split your document into pages.

transform_config = {
    "mode": "advanced",
    "segmentation_config": {
        "mode": "page",
        "page_separator": "\n---\n"
    }
}

Element Segmentation Configuration

The element segmentation configuration is used to define the segmentation by element which identifies the elements from the document as title, paragraph, list, table, etc.

transform_config = {
    "mode": "advanced",
    "segmentation_config": {
        "mode": "element"
    }
}

Chunking Configuration

Chunking configuration is mainly used to deal with context window limitaitons of embeddings model and LLMs. Conceptually, it’s the step after segmenting, where segments are further broken down into smaller chunks as necessary to fit into the context window. It include a few modes none, character, token, sentence and semantic.

Also all chunk config modes allow the user to define the chunk_size and chunk_overlap parameters. In the examples below we are not always defining the chunk_size and chunk_overlap parameters but you can always define them.

None Chunking Configuration

The none chunking configuration is used to define no chunking.

transform_config = {
    "mode": "advanced",
    "chunking_config": {
        "mode": "none"
    }
}

Character Chunking Configuration

The character chunking configuration is used to define the chunking by character and the chunk_size parameter is used to define the size of the chunk.

transform_config = {
    "mode": "advanced",
    "chunking_config": {
        "mode": "character",
        "chunk_size": 1000
    }
}

Token Chunking Configuration

The token chunking configuration is used to define the chunking by token and uses OpenAI tokenizer behind the hood. Alsochunk_size and chunk_overlap parameters are used to define the size of the chunk and the overlap between the chunks.

transform_config = {
    "mode": "advanced",
    "chunking_config": {
        "mode": "token",
        "chunk_size": 1000,
        "chunk_overlap": 100
    }
}

Sentence Chunking Configuration

The sentence chunking configuration is used to define the chunking by sentence and the separator and paragraph_separator parameters are used to define the separator between the sentences and paragraphs.

transform_config = {
    "mode": "advanced",
    "chunking_config": {
        "mode": "sentence",
        "separator": " ",
        "paragraph_separator": "\n"
    }
}

Embedding Model

The embedding model allows you to construct a numerical representation of the text within your files. This is a crucial step in allowing you to search for specific information within your files. There are a wide variety of embedding models to choose from, and we support quite a few on LlamaCloud.

Sparse Model Configuration

The sparse model configuration enables hybrid search by combining dense embeddings with sparse embeddings for improved retrieval accuracy. This configuration is particularly useful for scenarios where you want to leverage both semantic similarity (dense) and keyword matching (sparse) capabilities.

Available Sparse Models

LlamaCloud supports three sparse model types:

auto (default): Automatically selects the appropriate sparse model (Default: Splade)
splade: Uses SPLADE model for learned sparse representations
bm25: Uses Qdrant’s FastEmbed BM25 model for traditional keyword-based sparse embeddings

Configuration

You can configure the sparse model when creating or updating a pipeline:

from llama_cloud import LlamaCloudClient

client = LlamaCloudClient(api_key="your_api_key")

# Create pipeline with sparse model configuration
pipeline = client.pipelines.create_pipeline(
    name="my-hybrid-pipeline",
    # ... other pipeline configuration ...
    sparse_model_config={
        "model_type": "splade"  # or "bm25", "auto"
    }
)

Usage in Retrieval

When using hybrid search with configured sparse models, you can control the balance between dense and sparse retrieval:

from llama_cloud_services import LlamaCloudIndex

# Connect to your pipeline
index = LlamaCloudIndex("my-hybrid-pipeline", project_name="Default")

# Configure retriever for hybrid search
retriever = index.as_retriever(
    dense_similarity_top_k=5,      # Number of results from dense search
    sparse_similarity_top_k=5,     # Number of results from sparse search
    alpha=0.5,                     # Balance between dense (0.0) and sparse (1.0)
    enable_reranking=True,         # Optional reranking for better results
    rerank_top_n=10               # Number of results to rerank
)

nodes = retriever.retrieve("your search query")

After Pre-Processing, your data is ready to be sent to the Data Sink ➡️