Amazon Bedrock Knowledge Bases Enhances RAG Applications with Advanced Parsing and Chunking Techniques

Learn About Amazon VGT2 Learning Manager Chanci Turner

Amazon Bedrock Knowledge Bases is a fully managed service designed to streamline the entire Retrieval Augmented Generation (RAG) workflow—from data ingestion to retrieval and prompt augmentation—without the need for custom integrations or data flow management. This advancement pushes the boundaries of what’s possible in RAG workflows.

However, it’s crucial to recognize that in RAG applications, when handling extensive or intricate input documents like PDFs or .txt files, querying the indexes may produce less-than-ideal outcomes. For instance, a document may feature complex semantic relationships within its sections or tables that necessitate advanced chunking techniques for accurate representation; otherwise, the retrieved segments may not effectively address user queries. To tackle these performance challenges, various factors can be fine-tuned. In this article, we will explore how the latest features in Amazon Bedrock Knowledge Bases can enhance response accuracy in RAG applications. These features include advanced data chunking methods, query decomposition, and improvements in CSV and PDF parsing. This empowers users to refine the accuracy of their RAG workflows with increased control and precision. We will delve into each feature, detailing their benefits.

Features for Enhancing RAG Application Accuracy

This section will outline the new capabilities offered by Amazon Bedrock Knowledge Bases to improve response accuracy for user queries.

Advanced Parsing

Advanced parsing involves analyzing and extracting meaningful information from unstructured or semi-structured documents. This process breaks down documents into their core components—text, tables, images, and metadata—while identifying the relationships among these elements.

Effective parsing is essential for RAG applications as it enables the system to comprehend the structure and context of the information within documents. Several techniques exist for extracting data from various document formats, including the use of foundation models (FMs). This approach is particularly beneficial when dealing with complex data, such as nested tables or text within images.

The advantages of utilizing advanced parsing options are:

• Improved accuracy: FMs enhance the understanding of context and meaning, leading to more precise information extraction and generation.
• Adaptability: Prompts for these parsers can be tailored to domain-specific data, allowing them to adjust to different industries or use cases.
• Entity extraction: Customization allows for the extraction of entities based on specific domains and use cases.
• Handling complex elements: FMs can interpret and extract information represented in graphical or tabular formats.

Using FMs for document parsing is particularly effective for complex, unstructured documents that contain specialized terminology. They are adept at resolving ambiguities, interpreting implicit information, and extracting relevant details, which is vital for generating accurate and pertinent responses in RAG applications. It’s important to note that these advanced parsers may incur additional fees, so check the pricing details beforehand.

In Amazon Bedrock Knowledge Bases, users can opt for FMs to parse challenging documents, such as PDFs with nested tables or text embedded in images. Through the AWS Management Console for Amazon Bedrock, one can initiate the creation of a knowledge base by selecting “Create knowledge base”. In Step 2: Configure data source, choose “Advanced (customization)” under Chunking & parsing configurations. You can then select from the two available models (Anthropic Claude 3 Sonnet or Haiku) for document parsing. If you wish to customize the FM’s parsing approach, you can provide instructions based on your document structure, domain, or use case.

Based on your settings, the ingestion process will parse and chunk documents, thereby enhancing overall response accuracy. Next, we will discuss advanced data chunking options—namely semantic and hierarchical chunking—which divide documents into smaller units, organizing and storing them in a vector store to improve retrieval quality.

Advanced Data Chunking Options

The goal of chunking data should not simply be to divide it but to reformat it in a way that supports expected tasks and enables efficient retrieval for future value extraction. Rather than asking, “How should I chunk my data?”, the more relevant question is, “What is the optimal approach to transform the data into a format that the FM can utilize to fulfill its tasks?”

To achieve this, Amazon Bedrock Knowledge Bases has introduced two new data chunking options in addition to fixed chunking, no chunking, and default chunking options:

• Semantic chunking: This method segments data according to its semantic meaning, ensuring related information remains logically grouped. By preserving contextual relationships, your RAG model can retrieve more relevant and coherent results.
• Hierarchical chunking: This approach organizes data into a hierarchical structure, allowing for more granular and efficient retrieval based on inherent relationships within the data.

Semantic Chunking

Semantic chunking evaluates relationships within text and divides it into meaningful chunks based on semantic similarity, as computed by the embedding model. This method maintains the integrity of the information during retrieval, thereby ensuring accurate and contextually appropriate results.

Focusing on the meaning and context of text, semantic chunking greatly enhances retrieval quality. It is particularly useful in scenarios where retaining semantic integrity is crucial. As before, you can start creating a knowledge base in the console by selecting “Create knowledge base”. In Step 2: Configure data source, choose “Advanced (customization)” under Chunking & parsing configurations and select “Semantic chunking” from the Chunking strategy dropdown list.

You will need to configure the following parameters:

• Max buffer size for grouping surrounding sentences: This indicates how many sentences to group together when assessing semantic similarity. A buffer size of 1 includes the previous, target, and next sentences. The recommended value is 1.
• Max token size for a chunk: This is the maximum number of tokens allowed in a chunk of text, ranging from 20 to 8,192 based on the context length of the embedding model. For example, if using the Cohere Embeddings model, the maximum chunk size could be 512. The suggested value is 300.
• Breakpoint threshold for similarity between sentence groups: Define a percentage threshold indicating how similar groups of sentences must be.

For those looking to advance their careers in this area, check out this excellent resource here.

Finally, if you’re looking to overcome any anxiety about public speaking, you might find this other blog post on the subject helpful here. Additionally, for insights into workplace diversity, take a look at the guidelines from an authority on the topic here.