Deep Dive FAQ

This article contains common questions and answers about Deep Dive, Everlaw's generative AI powered question-and-answer tool.

To learn more about how to use Deep Dive, see:

• Enroll and Get Started with Deep Dive
• Use Deep Dive 
• Deep Dive Admin Guide
• Deep Dive Ask Guide

Questions

• Understanding Deep Dive
  • Does Deep Dive derive answers from the full contents of the document(s), or does it use some other source like the document summary?
  • How many relevant docs do we have the LLM look at?
  • How does Deep Dive take into consideration repeated content (e.g., email footers)? How does this kind of content impact the output and/or is it being factored in?
  • How many documents can Deep Dive include in the corpus-wide search?
  • Does Deep Dive ever produce hallucinations? I'd understood that while they can be reduced they are an almost inevitable bi-product of any GenAI.
  • Does Deep Dive take into account sentiment or tone of the conversation?
• Using Deep Dive
  • Is the conversation threaded? In other words, is this a conversational type of question and answer, in which responses can build off of the questions and answers from a previous question?
  • Is Deep Dive available for Early Case Assessment (ECA) projects?
  • Do the same “best practices” from Coding suggestions, such as providing context for specific terms, apply to asking questions of Deep Dive?
  • Is there a way to export the answer to the questions to another part of the platform, like Storybuilder?
  • When generating facts from the dataset, does Deep Dive take everything said in the documents as true? Will it identify a statement in a document that is untrue, or will it read all statements as sources of "fact"?
  • Can Deep Dive point me toward additional data sources that aren't currently in my dataset, based on what's already in my dataset?
  • Can I turn on Deep Dive only for a partial project?
  • How are the answers to my questions affected If I have access to only a subset of documents (due to restrictions set via Document Access Management)?
  • Is Deep Dive available in my region?

Understanding Deep Dive

Does Deep Dive derive answers from the full contents of the document(s), or does it use some other source like the document summary?

Deep Dive uses Retrieval Augmented Generation (RAG) to generate answers from the full contents of the documents. In RAG, the data (your document contents) is ingested into chunks of data called embeddings that are stored in a vector database. This format and storage structure makes your data more searchable and retrievable for the LLM to answer questions.

How many relevant docs does the LLM look at?

Deep Dive searches all your documents, then retrieves up to 200 documents as possibly relevant. Of those 200, it analyzes up to 50 documents for fact relevance. You can see this process happening while Deep Dive is generating its answer.

How does Deep Dive take into consideration repeated content (e.g. email footers)? How does this kind of content impact the output and/or is it being factored in? 

Our ingestion process is designed to limit duplicate/repeated chunks of information, including handling email footers and near duplicate documents.

How many documents can Deep Dive include in the corpus-wide search? 

Deep Dive performance is dependent on the total combined text contained within your documents. So far we have tested Deep Dive with up to 10+ million documents with positive results.

Does Deep Dive ever produce hallucinations? I'd understood that while they can be reduced they are an almost inevitable bi-product of any GenAI.

No GenAI tool is 100% hallucination free. That said, Everlaw AI Deep Dive is designed to reduce hallucinations based on safeguards such as: 

• Search only within the four corners of the documents – not the web or the model’s own knowledge
• Alerts you when no promising answers are found
• The response includes citations that can be quickly verified so users can evaluate results

Does Deep Dive take into account sentiment or tone of the conversation?

Yes. Deep Dive utilizes semantic retrieval and large language models that can retrieve documents and produce answers based on a particular sentiment or tone.

My documents are not in English. Will Deep Dive work?

Deep Dive works with text across many languages.

Using Deep Dive

Is the conversation threaded? In other words, is this a conversational type of question and answer, in which responses can build off of the questions and answers from a previous question?

Not at this time. You can ask multiple questions at the same time and the responses can get generated in parallel, but the answers and questions do not build off of each other.
In other words, each query and answer is independent and not incorporated into subsequent  queries. Prior questions, answers, and supporting links are stored (at the project level) for the user and others to see, reference, and re-use.
Question threading is currently on our development roadmap.

Is Deep Dive available for Early Case Assessment (ECA) projects? 

Yes. Enrollment for ECA databases works the same as for standard databases. You can enroll the entire database, which enables Deep Dive on both the ECA project and Review project(s), or just the Review project. 

Do the same “best practices” from Coding suggestions, such as providing context for specific terms, apply to asking questions of Deep Dive?

What's unique about Deep Dive is that you can ask the question, see how it responds, and then iterate from there. 
For example, if you want to ask about suspicious activities, iterate your question and then ask follow up questions if the first ask doesn’t get at what you’re looking for.
To help give overall context, an Admin should fill out the case description on the Project Settings page. This can give background that will be useful for all questions asked in Deep Dive.
Check out our Ask Guide for more suggestions on how to ask questions

Is there a way to export the answer to the questions to another part of Everlaw, like Storybuilder? 

Yes, you can copy a response to a Storybuilder Draft or Deposition.

When generating facts from the dataset, does Deep Dive take everything said in the documents as true? Will it identify a statement in a document that is untrue, or will it read all statements as sources of "fact"?

Deep Dive’s document analysis is concerned with weighing the relevance of the document to your question. In that sense it’s not determining "truth or falsity" of the content, but whether the content is relevant to answering the question.

Can Deep Dive point me toward additional data sources that aren't currently in my dataset, based on what's already in my dataset?

Not at this time.

Can I turn on Deep Dive only for a partial project?

The projects that Deep Dive is available for is determined at the time of enrollment for the database. The Organization Admin or AI Admin who is enrolling the database can enroll the entire database (all complete and partial projects, both current and future), or a subset of one or more partial projects. Once enrollment is confirmed, there is no way to later update the enrollment for that database. If only partial projects are enrolled, additional partial projects cannot be added to enrollment later. See Enroll and Get Started with Deep Dive for more information.

How are the answers to my questions affected If I have access to only a subset of documents (due to restrictions set via Document Access Management)?

Users whose access to documents is restricted via Document Access Management cannot use Deep Dive. 

Is Deep Dive available in my region?

Deep Dive is available for clients in our Canadian, EU, UK, US, and Australia environments. It is not available in GovCloud.

• US and Australia: Deep Dive is handled by sub-processors in the US
• Canada: Deep Dive is handled by sub-processors in Canada and the US
• EU and UK: Deep Dive is handled by sub-processors in the US and EU region

I just uploaded new data into a case where I am already using Deep Dive. Will that data get included in Deep Dive responses?

Yes, the data is automatically ingested into any projects in which  Deep Dive is enabled. During ingestion, Deep Dive is still available. A warning banner within Deep Dive lets you know that documents are currently being ingested and the responses may not include the documents that were just added.

What happens if my Everlaw contract expires?

Deep Dive requires access to AI credits to enroll new documents. If the contract that enables AI credit usage is expired, new uploads will not be enrolled into any databases or projects. 

Deep Dive will remain active and functional for all documents that are already enrolled. Productions that you create on Everlaw will still be enrolled even without an active contract that enables Evelraw AI credits, as the productions you create on Everlaw don't require credits to enroll.

If you uploaded documents into a Deep Dive-enrolled database/project while the contract that enables AI credits was inactive, reach out to support@everlaw.com when your contract is active again. They can help you reactivate enrollment and get your unenrolled documents into Deep Dive. When you contact support, include your Database ID (Organization home > Projects & Users tab > three-dot menu for the relevant database > Database details) in your communication.