AI Engineer Code 2025
Enterprise Deep Research: The Next Killer App for Enterprise AI — Ofer Mendelevitch, Vectara
Read the talk
Enterprise Deep Research: Investigating Private Data
Enterprise deep research extends autonomous investigation to private documents, combining retrieval, parallel agents and synthesis to produce grounded answers across an organization’s data.
From a talk by Ofer Mendelevitch
The platform beneath the research
What does an enterprise agent platform need to support document generation, internal and external chatbots, and deeper research? Ofer Mendelevitch opens with these adjacent applications of Vectara, then turns to the infrastructure they share. Its agent operating system is available as SaaS, in a customer’s own VPC, or on-premise in a customer’s data center.
The first requirement is making the organization’s information retrievable. Multimodal ingestion brings images and tables into retrieval-augmented generation (RAG) and agentic RAG workflows, so useful evidence need not be confined to document prose. Hybrid retrieval, metadata features and re-ranking help find and prioritize relevant material. Hallucination detection and correction address what happens when that material becomes a generated answer.
Vectara’s HHEM evaluates whether generated text is consistent with supplied source material; that does not establish whether the source itself is true. Mendelevitch reports that HHEM had passed five million downloads a couple of months earlier and tentatively estimates about 5.5 million at presentation time. Those are adoption figures, not accuracy measurements.
The surrounding enterprise requirements include security, role-based access controls, bring-your-own-model support, custom prompts, observability and monitoring. These capabilities determine where the application can run, who can use it, and how its behavior can be configured and observed.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Research needs reliable evidence
A comprehensive report still needs a reliable factual foundation. Hallucinations remain a problem across generative AI applications, including enterprise research. The retrieval and mitigation capabilities matter because the research process must base its answers on robust information.
Mendelevitch cites approximately 73% of LLM customers implementing use cases as naming factual accuracy their top challenge. The statistic’s survey source and methodology are not identified in the supplied material, so it serves here as his motivation for emphasizing hallucination mitigation, rather than a measured result for Vectara.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From a question to a cited report
Deep research begins with a question that requires investigation. An agent browses or searches the web autonomously, gathers results across multiple steps, and synthesizes them into a comprehensive report with citations. The output brings the findings together to answer the original question. Mendelevitch names Google’s Gemini, ChatGPT deep research, Anthropic and Perplexity as implementations of this web-based pattern.
The recording then shows where to select deep research in Gemini and ChatGPT. These are feature-selection screenshots, rather than a completed research run. Mendelevitch estimates that a deep research task typically takes 20–30 minutes because of the work performed underneath. That is his general estimate for the workflow being discussed, not a cross-provider timing benchmark.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Turning the investigation toward private data
Enterprise Deep Research applies the same investigative pattern to an organization’s private information. Mendelevitch describes multiple agents executing in parallel, reflection during the research process, and synthesis of the final results. The agents query enterprise data through Vectara’s agentic RAG capabilities, carrying retrieval accuracy and hallucination mitigation into the investigation.
The additional planning capability is corpus understanding: understanding the available collection well enough to plan research around the data it actually contains. Parallel execution provides a way to pursue multiple parts of an investigation, while corpus understanding helps determine what those parts should be. The overview names these capabilities without specifying the reflection algorithm or agent orchestration details.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
RFP answers, onboarding guides and investment memos
RFP responses. Mendelevitch’s RFP example involves answering 150 questions. The difficulty is assembling the right information across enterprise datasets. Enterprise research would search those datasets, select the documents relevant to each question, and use them to prepare answers. This is a proposed application of the research workflow, not a timed demonstration of a completed response.
Employee onboarding. A new employee joining a team or company may find no onboarding guide—or one last generated three years earlier. The proposed alternative is an on-demand guide assembled from documentation in Jira, Notion, Google Drive and SharePoint. Here, research turns scattered organizational information into a document intended to help someone get started.
Industry-specific outputs. In financial services, the resulting document could be an investment memo. Mendelevitch also points to healthcare and insurance as possible domains, without developing specific examples. The substantive ending broadens the output from answers and guides to documents shaped around an industry’s research needs.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Open hallucination-evaluation model with examples for checking generated text against source material.
OpenAI's launch explanation of ChatGPT deep research, including its workflow, citations and limitations.
Further reading
Explains corpus introspection, research planning and parallel agents for investigating private enterprise documents.
Walks through Vectara's use of internal documentation to draft responses to a 152-question SOC2 questionnaire.
Read the complete timestamped transcript
- 0:00
[instrumental music] Hi, I'm Ofer from Vectara. At Vectara, we developed a trustworthy agent operating system, and there's a lot of really cool use cases with this, like document generation, conversational AI or chatbots, either internal or external, and Enterprise Deep Research, which I'm gonna talk about today.
- 0:23
But before I jump into Enterprise Deep Research, let me tell you a little bit more about our operating system for agents. First of all, it's a SaaS platform, but also runs on your own VPC or on-premise in your own data center.
- 0:38
And here are some of the main features that we really are proud of. We have very advanced multimodal ingest to support images, tables, in a way that makes them, you know, findable and retrieved to be able to make sense of them in a, in a RAG or agentic RAG workflow.
- 0:56
Very strong focus on, again, retrieval accuracy with hybrid retrieval, lots of features around metadata, re-ranking, et cetera. And then we're, we're kinda known for a lot of work around hallucination mitigation, both hallucination detection and correction.
- 1:12
In fact, our hallucination detection model, also called HHEM, has just passed five mil downloads about couple months ago. I think it's at 5.5 right now or something like that.
- 1:25
And generally, our, our operating system platform is what you would need for enterprise-grade deployment, so security, role-based access controls, bring your own model, custom prompts, observability, monitoring, everything you would need.
- 1:41
So why does this matter? Well, in any generative AI application, and Enterprise Deep Research is no different, hallucinations are still a problem, and you wanna base your applications on really robust information.
- 1:57
In fact, this is, uh, statistics that shows that about 73% of LLM customers implementing use cases say that factual accuracy is their top challenge right now. So that's why we spend so much time in hallucination mitigation, which enables Enterprise Deep Research at really high quality.
- 2:17
Okay. So with that in mind, let me jump into what is deep research. Deep research itself is something many of you probably know already. It's when an AI agent conducts in-depth multi-step investigation, usually by autonomously browsing or searching the web in some way, getting results, synthesizing all these results together to generate a comprehensive report for you with
- 2:41
citations and all the information you need to answer a particular question. Many have implemented this sort of web-based deep research. Gemini or Google has that, ChatGPT, Anthropic, Perplexity, et cetera.
- 2:54
Here's an example of what it looks like. If you haven't used it, I highly encourage you to use it. It's a very powerful tool, and I use this all the time.
- 3:00
This is the screenshot of how you choose it in Gemini, for example, and here's how you choose it in ChatGPT. And this is usually something that takes about, you know, twenty, thirty minutes to complete because it does a lot of work underneath the covers.
- 3:14
Now, Enterprise Deep Research, think about it as exactly the same idea, only now it goes to your private data. So again, the same process, multi-agent with reflection, with synthesis of the final results, parallel execution of, of agents underneath, and it queries your enterprise data, of course, using, in this case, Vectara agentic RAG capabilities with all the bells
- 3:38
and whistles of high accuracy and hallucination mitigation. And then we have corpus understanding, which allows you to plan properly based on your data.
- 3:49
There are many really amazing use cases for this, and I'm gonna just mention a few here that I like. One is responding to an RFP. I've had this in my career multiple times when you have to respond to an RFP, and getting the answers to 150 questions is really difficult.
- 4:06
So having-- being able to use Enterprise Deep Research to go through all of your enterprise datasets, picking up the right documents and answering those questions is a really cool use case.
- 4:17
Employee onboarding is this idea of if you come to a new team or you join a new company and a company wants to onboard you quickly, it's usually very difficult.
- 4:26
Nobody knows what's going on. There's no, like, onboarding guide. The last one was generated three years ago, and it's not up to date, et cetera. So again, being able to generate a on-demand onboarding guide using all the documentation we have on Jira or on Notion or Google Drive or SharePoint, very, very powerful.
- 4:48
And then in different industries, there's also different use cases. For example, in financial services, you might have the generation of an investment memo as a really cool use case, and you can imagine the same thing in healthcare and insurance and other industries as well.
- 5:02
So those are some of the use cases. There's a lot more. I'm Ofer. You can connect with me here in the links I'm showing below. And if you are interested to learn more about Vectara and Enterprise Deep Research, please contact us and we'll happy to, to do a demo for you.
- 5:17
Thanks very much.