Build a PDF Search Application with Vector Search and LLMs

Solution Overview

Insurance firms rely heavily on data processing. To make investment decisions or handle claims, they leverage vast amounts of mostly unstructured data. Underwriters and claims adjusters need to comb through numerous pages of guidelines, contracts, and reports, typically in PDF format. Manually finding and reviewing every piece of information is time-consuming and can easily lead to expensive mistakes, such as incorrect risk estimations.

Retrieval-augmented generation (RAG) applications are a game changer for insurance companies, enabling them to harness the power of unstructured data while promoting accessibility and flexibility. They are especially helpful for PDFs, which are common yet difficult to search. RAG makes PDF search more efficient and accurate. Now, users can type a question in natural language and the application will sift through the company data, provide an answer, summarize the content of the documents, and indicate the source of the information, including the page and paragraph where it was found.

In this GitHub repo, you will find detailed, step-by-step instructions on how to build a PDF search application that combines MongoDB, Superduper, and LLMs. Our use case for this solution focuses on a claims adjuster or an underwriter handling a specific case. Analyzing a PDF of guidelines that is associated with a specific customer helps determine the loss amount in the event of an accident, or the new premium in the case of a policy renewal. The application answers questions and displays the relevant sections of the document.

Combining Atlas Vector Search and LLMs to build RAG apps can directly impact the bottom line of an insurance company. To try our semantic search tool, visit the Atlas Vector Search Quick Start.

Reference Architecture

Combining MongoDB and Superduper allows you to build an information retrieval system with ease. The process involves the following steps:

1. The user adds the PDFs that need to be searched.

2. A script scans the PDFs, creates the chunks, and vectorizes them (see Figure 1). To ensure that you don't lose transitional data between chunks, the script uses the sliding window methodology to produce overlapping chunks.

3. Vectors and chunk metadata are stored in MongoDB, and a Vector Search index is created (see Figure 2).

4. The PDFs are now ready to be queried. The user selects a customer, asks a question, and the system returns an answer, displaying the page and paragraph where the information was found and highlighting the specific section with a red frame (see Figure 2).

Figure 1. PDF chunking, embedding creation, and storage, orchestrated with Superduper.io

Each customer has a guidelines PDF associated with their account based on country of residency. When the user selects a customer and asks a question, the system runs a vector search query only on that particular document, seamlessly filtering out the non-relevant ones. This is made possible by the pre-filtering field included in the index and the search query (see code snippets below).

Atlas Vector Search also takes advantage of MongoDB's new Search Nodes dedicated architecture, enabling better optimization for the right level of resourcing for specific workload needs. Search Nodes provide dedicated infrastructure for Atlas Search and Vector Search workloads, allowing you to optimize compute resources and fully scale search needs independent of the database. Search Nodes provide workload isolation, higher availability, and the ability to better optimize resource usage.

Figure 2. PDF querying flow, orchestrated with Superduper.io

{
   "fields": [
      {
         "numDimensions": 1536,
         "path": "_outputs.elements.text-embedding.0",
         "similarity": "cosine",
         "type": "vector"
      },
      {
         "path": "_outputs.elements.chunk.0.source_elements.metadata.filename",
         "type": "filter"
      }
   ]
}

vector_query = [
   {
      "$vectorSearch": {
         'index': INDEX_NAME,
         'path': "_outputs.elements.text-embedding.0",
         'queryVector': query_embedding,
         'numCandidates': vector_search_top_k,
         'limit': vector_search_top_k,
         'filter': {'_outputs.elements.chunk.0.source_elements.metadata.filename': filename}
      }
   },
   {
      "$project": {
         "_outputs.elements.text-embedding.0": 0,
         "score": { "$meta": "vectorSearchScore" }
      }
   }
 ]

Superduper.io

Superduper.io is an open-source Python framework for integrating AI models and workflows directly with and across major databases for more flexible and scalable custom enterprise AI solutions. It enables developers to build, deploy, and manage AI on their existing infrastructure and data while using their preferred tools, eliminating data migration and duplication.

With Superduper.io, developers can:

• Bring AI to their databases, which eliminates data pipelines and minimizes engineering efforts, time to production, and computation resources.

• Implement AI workflows with any AI models and APIs on any type of data.

• Safeguard data by switching from APIs to hosting and fine-tuning their own models on their own infrastructure.

• Switch from embedding models to other API providers. They can also host their models on HuggingFace or other platforms.

Superduper.io provides an array of sample use cases and notebooks that developers can use to get started, including vector search with MongoDB, embedding generation, multimodal search, RAG, transfer learning, and many more. This solution's demo is adapted from a previously developed app from Superduper.io.

Build the Solution

Build the solution following the instructions in this Github repo. The solution is composed of two steps:

1. The initialization script breaks down the PDFs into chunks and then turns them into vector embeddings.

2. The querying step allows the user to interrogate the documents.

Key Learnings

• Embeddings can use different models and deployment options: You can deploy a model locally if your data needs to remain on the servers. Otherwise, you can call an API and get your vector embeddings back, as explained in this tutorial. You can use Voyage AI or open-source models. When building your system, consider privacy and security requirements.

• Superduper integrates AI models and workflow: Superduper is the framework that provides a simple and standard interface to interact with Vector Search and LLMs.

Authors

• Luca Napoli, Industry Solutions, MongoDB

• Clarence Ondieki, Solutions Architect, MongoDB

• Pedro Bereilh, Industry Solutions, MongoDB