ENTERPRISE-GRADE RAG
The GroundX Platform
APIs and tools for building advanced Retrieval-Augmented Generation (RAG) applications that meet the demands of regulated enterprises: accurate, secure and scaled.
APIs and tools for building advanced Retrieval-Augmented Generation (RAG) applications that meet the demands of regulated enterprises: accurate, secure and scaled.
Built on Kubernetes and powered by fine-tuned open source models, the GroundX Platform consists of four interlocking systems: ingest, store, search and eval
GroundX Ingest directly tackles the largest cause of hallucination in RAG applications... garbage in / garbage out.
GroundX merges two state-of-the-art-models: a vision model and a multimodal model. We then fine tune the system on nearly 1M pages of enterprise documents across many verticals.
This lets us correctly parse complex visual pages by identifying the different text, tabular and image objects on each page.
The colored boxes in the image below are generated by our vision model. Notice how the model identifies different text blocks on the page and correctly recognizes the complex flow chart.
Watch GroundX turn this Walmart chart into LLM-ready data
Each object is then sent to the correct pipeline for transformation into LLM-ready data. In this phase, we use a multimodal modal and chain of thought to explain complex objects like tables and graphics, generate important metadata used to improve RAG search and LLM completion and create several versions of each chunk that can be used with a wide array of LLMs for completion.
We call this output a semantic object. You could think of it as a super chunk. Here is a snippet of JSON we generate to describe the flow chart to the completion model.
[
{
"figure number": 1,
"figure title": "Walmart National Brand Transit Testing Flow Chart"
},
{
"connected_to": "Submit a sample in the EXACT packaging for production to an ISTA-certified lab for testing.",
"direction": "down",
"element": "New or existing item"
},
{
"connected_to": "Is Item required to be SIOC",
"direction": "down",
"element": "Submit a sample in the EXACT packaging for production to an ISTA-certified lab for testing."
},
{
"connected_to": [
"Does the item ship LTL from the Fulfillment Center?",
"Is the item a multi-pack of glass/ceramic bottles/jars not sold through eCommerce?"
],
"direction": [
"yes",
"no"
],
"element": "Is Item required to be SIOC"
},
{
"connected_to": [
"Follow ISTA 6A Fed-Ex protocol",
"Follow ISTA 3B protocol"
],
"direction": [
"yes",
"no"
],
"element": "Does the item ship LTL from the Fulfillment Center?"
},
Notice the JSON is describing both the structure and contents of the chart. We also generate a field called narrative text, which describes an object in natural language. Like humans, language models perform better if you explain your content to them then provide the data. This reduces downstream hallucination. Here is a snippet of the narrative text for the chart.
Vital Context
In the final pass, GroundX compares each semantic object to surrounding objects and also a summary of the entire document. The system uses the extra context to improve its understanding of each object.
This is critical when information from one part of a page is important to another. For example, a financial table might be on the bottom of a page and rich text explaining its purpose might be on the top. This context pass lets us enrich our understanding of the table from the text above. This is one of nearly a dozen techniques inside GroundX to reduce downstream hallucinations.
The World's First Custom Fine Tune for Ingest Models
GroundX performs industry leading ingest automatically across a wide array of documents. However, we also make it possible to fine tune the ingest model based on your unique documents. This is a powerful capability that is totally unique to GroundX.
Try Your Docs for Free Here
GroundX Store is the ultimate solution for secure, scalable, and cost-effective data storage tailored for enterprise RAG applications.
Secure at the Core
Your data is encrypted at rest and in transit, ensuring it stays protected every step of the way. Hosted within a dedicated Virtual Private Cloud (VPC), GroundX provides enterprise-grade isolation and control. With your own encryption keys, you maintain full ownership over your data's security.
GroundX On-Prem customers have total control of their security environment. Deploy GroundX in your managed cloud, on premises data center and even air-gapped installations.
Built on Open Source
Built on trusted open-source technologies like OpenSearch, MySQL, MiniIO and Redis, GroundX Store combines reliability with flexibility. It’s a solution enterprises worldwide trust for their most critical applications, offering seamless integration and unmatched performance.
Natively Multimodal
GroundX Store doesn’t just hold your files; it empowers them. Seamlessly store source files, multimodal objects, and vectors in a unified system optimized for rapid retrieval. Whether you're working with text, images, or high-dimensional vectors, GroundX ensures your data is always ready for advanced RAG workflows.
Autoscales without Sacrifice
GroundX Store delivers lightning quick retrievals even as you scale your RAG applications to millions of documents.
Storage is 100% Free
And the best part? GroundX Store is 100% free. You’ll never pay for storage, no matter how much data you manage or how far you scale your applications. This cost-free approach lets you focus on innovation, not infrastructure costs. On-prem customers need to provision storage, of course, but there are no additional charges from EyeLevel.
Enterprise-Grade Accuracy
GroundX Search is engineered for precision. In a head-to-head study, GroundX was 50% more accurate than other RAG frameworks, especially on complex visual documents with tables, graphics and forms. Whether you’re working with intricate datasets or large-scale archives, GroundX ensures your search results are not only fast but highly accurate.
Unmatched Scalability
Most RAG systems falter as the volume grows, losing search accuracy in just 10K pages. GroundX Search maintains its precision across millions of pages, giving you consistent and reliable performance no matter how much your data grows. GroundX showed 6X less accuracy degradation at scale than Pinecone, in recent research.
A Unique Hybrid Approach to Search
GroundX blends the strengths of text, vector, and micro graph search into a unified framework. This hybrid approach offers fine-tuned control, allowing you to achieve optimal results tailored to your application’s unique needs.
Fine-Tuned Reranker for Optimal Results
At the core of GroundX Search is a custom fine-tuned reranker model that selects the most relevant results from text, vector, and micro graph searches. This ensures that your queries consistently return the best possible answers, even in the most challenging datasets.
GroundX offers a unique approach to evaluation. Most systems only look at the end of a RAG pipeline, comparing completions to retrievals or ground truth. But experienced RAG practitioners know RAG errors start at the beginning of the pipeline not the end. GroundX evaluation tools help you see errors in document ingest, search retrievals and chat completions. Here's how.
The Power of X-Ray
Gain unparalleled visibility into your data processing with X-Ray, our groundbreaking visual tool. X-Ray shows exactly how your documents are parsed and transformed into LLM-ready data. Click on any object in your source document and instantly review the semantic object GroundX has generated (JSON, narrative text, suggested text, extracted text, file summary, keywords and more).
This is an absolutely critical step for catching ingestion issues at the front of the pipeline, before it's too late. X-Ray is a unique feature of GroundX you won't find in any other RAG platform.
See X-ray in action on your docs
Retrieval Viewer
Transparency is key to trust and performance. With the Retrieval Viewer, you can inspect your search results and their metadata before they are sent to the LLM for completion. This lets you verify the accuracy and relevance of the retrieved content, ensuring reliable outputs every time.
Because GroundX is natively multimodal, with every retrieval you get the suggested text and an image of the object (if it's graphics or a table), the document URL, the page number and the bounding box coordinates for where that object lives on the page. You also get the ranked score for each object.
Completions + Sources
GroundX's instant chat tool, lets you "talk" to your documents in a single click. You can click on sources for each answer and see the x-ray of each object. This takes you all through the RAG chain from completion to retrieval to source document to extraction.
Conclusion
If you're building RAG applications or features and need enterprise-grade accuracy, security and scalability without building a whole RAG team, GroundX delivers what you need in just a few lines of code with a totally unique tool set. Talk to us.
