agent builder RAG

[Rajavelu]

Hi Experts,

I would like to understand difference between RAG and Agent Builder Data store.
I understand RAG uses vector data for retrieval and data store contains indexed documents.

1)Now which one i have to choose if I am having 80 PDF documents and which approach is more accurate if I need some information from on of the document?

2)Sometimes the table data inside pdf documents retrieval using agent app and data store is not accurate. It couldn't retrieve data of columns inside some contents of pdf document.

Kindly help me to understand.

[ruthseki]

Hi @Rajavelu,

Welcome to Google Cloud Community!

If you have 80 PDFs and need precise information from specific parts of those documents, RAG is generally the better choice. RAG focuses on efficient retrieval of the exact piece of information that matches your query, even if it's a single sentence within a document. Agent Builder tends to give you the entire document or parts that contain related keywords, which may be too much content for a simple query.

For the 80 PDFs, your process would likely be:

• Document Preprocessing: Extract relevant data (text, tables, specific fields).
• Embedding Creation: Create embeddings of this extracted information (e.g., a sentence about a table entry).
• Indexing: Index those embeddings.
• RAG Retrieval: Use the embeddings to retrieve specific information.

With regard to your concern about the table data retrieval from PDFs within Agent Builder or general data stores, Agent Builder stores documents but might not be extracting the tabular data.

Here are some workarounds that you may try within a RAG pipeline:

• Advanced document preprocessing: Use libraries or custom code to extract table data from the PDF. 
• Structured data representation: Convert the table data into a more structured format (JSON, CSV), potentially with column names and indices. This is crucial for the embedding process, making a query about the quantity from the table effective.
• Embedding the table data: Embed columns of data in the table and/or entire rows separately to make retrieval of data from that table possible. If you're retrieving table entries, focus on embedding the specific piece of table data you want to query.
• Query refinement: If the LLM fails to identify the table data correctly, modify your prompts/queries accordingly.

Additionally, you may refer to these documentations for more information:

• RAG Engine overview
• Use advanced PDF parsing with RAG Engine
• Data stores

Was this helpful? If so, please accept this answer as “Solution”. If you need additional assistance, reply here within 2 business days and I’ll be happy to help.

[Rajavelu]

Thanks for the details!

1)Agent builder has an option to integrate to google chat or messenger easily. How RAG integration to google chat or messenger works?

2)Can I store the embedded vector data in bigquery table and use it as for information retrieval instead of dedicated vector DB?

3)Where can i deploy my solution / RAG code in GCP?

4)Why agent app doesn't support bigquery table as data store?
Kindly help me to understand.