Inside Weaviate's MCP Ecosystem | Webinar

Author: Weaviate vector database · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Inside Weaviate’s MCP Ecosystem | Webinar · Channel: Weaviate vector database · 54:44 · Watch on YouTube

Topics: AI agents, Data, databases and embeddings, Software engineering with AI, RAG, search and information retrieval

Summary

The webinar introduces Weaviate’s MCP ecosystem, including a database MCP server built into Weaviate and a separate MCP server for the Query Agent. The presenters demonstrate how assistants can inspect, search, and write database data, and explain Query Agent features such as schema-aware retrieval, structured outputs, image generation, and research mode. Deligence Brokerage founder Alex Ledbetter describes building an AI-enabled commercial insurance business with Weaviate, using it to process policy documents and provide servicing across multiple channels. The closing Q&A covers Weaviate’s advantages, MCP availability, grounded answers, and Alex’s development tools and architecture.

Key points

MCP is an open standard that lets AI assistants connect with data and tools through servers.
Weaviate’s database MCP server is built into the database, served at /v1/mcp, and uses standard API-key authentication and RBAC.
The database MCP server provides tools to inspect collection configuration and tenants, run hybrid queries, and upsert objects; write access is configurable.
The Query Agent uses schema and data understanding to translate natural-language questions into Weaviate queries and support grounded answers.
Query Agent modes include ask, search, suggest queries, and research; the webinar also highlights structured outputs and retrieval-augmented image generation.
Alex Ledbetter used Weaviate as part of building Deligence Brokerage, an AI-enabled commercial insurance platform, despite having no prior coding experience.
Deligence processes policy documents using images, OCR, classification, and structured extraction, and makes insurance information accessible across channels.
Alex highlights Weaviate’s flexibility, developer experience, self-hosting options, and technical support as reasons for choosing it.
In the Q&A, presenters explain that Query Agent result evaluation can check whether answers are partial, missing information, or supported by sources.
Alex says he would simplify his architecture by using Weaviate for both structured and unstructured data.

Chapters

Start	Chapter	Summary
00:03	Welcome and webinar agenda	Andy introduces the speakers, outlines the webinar agenda, and explains that audience questions will be answered at the end.
01:36	Weaviate overview	Andy describes Weaviate as an open-source vector database for AI applications and briefly introduces related platform components.
03:10	MCP ecosystem and database server	Yevon introduces MCP and explains how Weaviate’s documentation, database, and Query Agent servers fit into the ecosystem.
06:15	Database MCP architecture, tools, and permissions	Yevon explains that the database MCP server runs inside Weaviate, authenticates through API keys, and provides four tools with permission controls.
08:50	Database MCP live demo	The demo connects an assistant to a Weaviate Cloud cluster and shows it inspecting collections, searching products, and adding a product.
13:43	Query Agent capabilities	Connor explains how the Query Agent supports RAG through schema understanding, query generation, retrieval, and grounded answers.
17:18	Query Agent modes and new features	Connor introduces the agent’s four modes and highlights structured outputs, retrieval-augmented image generation, effort tiers, and research mode.
20:59	Query Agent MCP integration	Connor describes connecting the Query Agent MCP server to assistants and agent frameworks, and discusses its tools and trade-offs.
25:08	Alex’s background and Deligence	Andy introduces Alex Ledbetter, who recounts starting Deligence after recognizing an opportunity to apply AI to commercial insurance.
28:15	Insurance onboarding and document processing	Alex demonstrates the platform’s onboarding and explains how it builds insured-specific knowledge bases from policy documents.
33:32	Insurance servicing across channels	Alex shows how extracted information supports quote requests, policy servicing, customer interactions, and multi-entity knowledge bases.
40:16	Why Weaviate and customer experience	The Q&A begins with a comparison of Weaviate to other database options, followed by Alex’s reasons for selecting it.
46:25	MCP availability and grounded answer evaluation	The speakers clarify which MCP services are open source or managed and explain how Query Agent evaluates answers against retrieved context.
48:58	Alex’s AI coding stack and closing	Alex discusses his tools and agent-first approach, then the hosts thank the speakers and audience and promise follow-up resources.

Quotes

There is no package, no extra container, nothing to install. — 06:15

So, that’s the whole loop. — 13:11

Weaviate’s mCP server was critical part of of me building out um our infrastructure. — 28:15

we have 142 customer-facing MCP tools. — 35:35

If I were to do it again, I would simplify and I would I would use Weaviate for structured data and unstructured data. — 53:43

Concepts and entities

Weaviate, MCP, Weaviate Database MCP server, Query Agent, Query Agent MCP server, RAG, vibe coding, Claude Code, Cursor, Pydantic AI, LangChain, Alex Ledbetter, Delegance Brokerage, AWS, Supabase

Full transcript (original language)

00:03 Right. Hi everybody and welcome to today’s webinar inside the Weaviate MCP ecosystem. I’m Andy Ellicott and my co-presenters today include Ivan and Connor from Weaviate engineering and we also have Alex Ledbetter, the CEO and founder of Deligence brokerage who’s going to talk about his experience using vibe coding with Weaviate. Our agenda today is starts with a quick introduction of overview of Weaviate in case you’re new to it and following that

00:35 Ivan will talk about the new MC server MCP server for the Weaviate database followed by Connor who’s going to explain the MCP server for the Weaviate query agent and we’ll wrap up with Alex talking about how he I think single-handedly and without much if any computer science experience vibe coded a commercial insurance brokerage business into existence in under a year and I think yesterday

01:06 just announced his pre-seed round of funding. So he’s off to a great start and his his story is really interesting and very relevant to anybody here who’s has come to this event hoping to learn more about AI assisted coding and operations of Weaviate. We’ll wrap up with Q&A and if you have any questions at any time during the webinar use the Q&A button on your Zoom screen and we’ll go through those at the end. We’ll get through as many as time

01:36 allows. And since one of the most common questions we get is will this be recorded and shared? Yes, it will. The event today is being recorded and we’ll share a video recording of this after the event in case you’d like to revisit any of it or share it with others. All right, so let’s start with a quick introduction to Weaviate for those of you who are new to it. Uh Weaviate is an open source vector database. It’s built specifically for AI use cases like semantic search, generative AI, rag, AI inference, agentic AI, and many others.

02:08 It makes building AI applications exceptionally easy uh because it’s purpose-built for indexing and uh uh very, very quickly searching vector data for those types of applications. Uh it’s also uh very capable of scaling and performing fast enough to support uh really large-scale AI initiatives as it does at Weaviate customers like Booking and Cisco and and many others. Weaviate is uh part of a platform that includes a

02:38 couple of other components including the query agent, which Connor’s going to talk about today. Uh the query agent facilitates natural language querying of your your data. Uh there’s also uh Weaviate embeddings, which is a capability that allows you to plug uh AI models into your um into the database to facilitate and simplify application architectures by facilitating the generation of vector data within the database itself. And finally, Ingram, which is a a newer

03:10 product which gives you a jump start in using Weaviate for AI memory in personalization, other AI applications. Okay. So, that’s uh that’s Weaviate in a minute. And uh without further ado, let me turn it over to Yevon who will dive into details on the new MCP server for the Weaviate database. >> Thank you for the introduction. So, let’s head straight in. Hi, everyone. I’m Yevon. Um I work on developer

03:40 experience here at Weaviate, and I help develop the database MCP server. Today, I’ll give you a short overview of the MPC ecosystem we use at Weaviate, uh show you the server that’s built into the database and spend some time in a live demo with V8 Cloud showcasing how to actually use the MCP tools. So, first off, the ecosystem map and a very short intro about MCP for anyone new to it. So, MCP, the Model Context Protocol, is an open standard that lets

04:11 AI assistants and tools, um, read data and call other tools from outside its system. It started at Anthropic like about 2 years ago and now lives under the Linux Foundation. Um, tools like Claude, ChatGPT, Cursor, all use it. It’s basically a protocol for AI-powered tools to directly communicate with various servers, applications, um, anything you can think of essentially. At the top of the slide, you can see some of the clients you might already be

04:42 using and underneath are the servers in the V8 ecosystem. Uh, for example, on the left is the V8 Docs MCP server. It’s hosted and powered by Kappa AI and it serves the V8 documentation to your assistant directly. So, for example, if you’re using an AI assistant to, uh, develop some code with V8 and have questions about the API or some configuration values inside V8, instead of a performing a web search, you can ask the Docs MCP server directly and it’s probably be more accurate than trying to

05:14 fetch the data yourself. Next is the V8 Database MCP server. It’s built directly into the V8 database itself and and it gives your assistant direct access to your own cluster, so you can inspect the schema, uh, search your data, write objects into it, and I’m going to cover that one in more depth in the, uh, next few slides and through the demo. Next to it is the Query Agent MCP. So, instead of querying the database directly using the Database

05:45 MCP server, you use a dedicated uh, AI-powered agent or assistant that already knows how to turn your questions into the right queries and with much more precision than your AI tool will do it. Conor will talk more about that one later. And you can, of course, also create custom MCP servers if you have a very specific use case that’s not covered by any of the aforementioned ones.

06:15 So, now let’s talk about the V8 database database MCP server. You can see the architecture of our built-in MCP server. What does built-in mean? There is no package, no extra container, nothing to install. So, when you run V8, the MCP server is already running inside of it and it’s served at /v1/mcp on the same port as the REST API. So, if you follow the arrow from left to right, you can see that your client, for

06:45 example, Cloud Code or VS Code or an agent, connects over streamable HTTP and sets a regular V8 API key to authenticate against the cluster. Every call goes through the same authentication and Rback as the rest of the API. Every tool is also mapped to a specific permissions. So, if, for example, you give your assistant viewer only key, it only has read-only access to your cluster. So, the

07:18 AI tool cannot do any destructive operations. Behind the gate, you can see the four currently available tools that are available. And the write tool has a lock on it because by default in the self-hosted V8 version, write access is disabled. You need to enable it yourself. On V8 Cloud, which I will show the demo on. The MCP server is on by default and write access is also on by default, but you can

07:48 disable it. Or you can provide granular permissions to an API key that you’re using, so the MCP has only access to some restricted data. The four tools that are available are the get collection config tool, so your assistant can learn about what data is inside your cluster, like the properties, which vectorizer you’re using, if multi-tenancy is turned on. The list tenants tool is used for inspecting multi-tenant collections,

08:20 uh, seeing which tenants are actually part of it. The hybrid query tool combines vector and keyword search, so it’s the main search tool you’re using to retrieve data from Weaviate, and the upsert tool is the main write tool, so you can use it to either insert new data or as an update tool to replace existing data with it. Uh, on the right you can also see the Rback permissions that map to the tools, so you can also provide granular access to what an agent can use.

08:50 Uh, and now let’s show all of this in action. First, we need a Weaviate cloud cluster. I have already created one, as you can see in the main dashboard. This is the REST endpoint that we’re going to use to connect with the MCP. You can also click on the how to connect button, and here in the MCP tab, you have a command that you can directly run in the tool you are using. This is for cloud code. You also have other options

09:23 available. I already connected mine locally, so you don’t have to see the API key. Let’s check it out. Now, let’s first check if the connection has been established. I’m going to do it with the MC P command and here I can see that we are connected to V8 cloud. Let’s also check out. This is the rest URL I have showed earlier of the cluster. And here we can also see the tools that are available. This isn’t

09:55 really of any importance to you as a user. This is what the agent is going to use in the end. For example, this is the description and the parameters that the agent will see and be able to configure on its own. So now let’s actually try a query. Let’s first ask it what is in the V8 cloud cluster. And if everything goes right,

10:27 as it always does with such live demos, it’s going to use the correct tool. Yes, it’s the V8 collection get config so it can check the schema of the cluster. And now it’s also going to run the V8 query hybrid tool to also check out some of the objects that are stored in the clothing collection. It also found a second collection, a brands collection. It also wants to query that one. Let’s let it.

11:00 And in just a few short seconds, it should make sense of the data that it retrieved from the cluster. And here we have it. It’s an e-commerce data set with two collections, a clothing one, essentially a product catalog with different clothing items and brands that are associated with those clothing items. So let’s now for example do a realistic query that someone in an e-commerce web application would use. Find me a summer dress for a wedding

11:35 under 200 bucks. And let’s see how it’s going to interpret this plain language request into a query. Here we see that it’s searching the clothing collection. The query is elegant summer dress for a wedding guest. Uh it’s limiting it to eight results. Here it’s applying specific filters because I said dress. Uh it’s searching in the dress category. The price needs to be less than 200. And

12:06 because I didn’t uh specify what to return, it’s going to just return all of the properties. Right, let’s run it. Nice, it worked. And I’m going to show you another query. Let’s for example add a product. So, add a new product to the clothing collection with some more details.

12:39 And it’s again going to construct the tool call. It’s using the objects upserch tool. It populated all of the fields correctly as I can see from a glance. Now, let’s run it. I’m interested to see if it’s going to perform a hybrid search just to check if it worked. And yes, it’s going to do it. Let’s see if it’s going to retrieve it.

13:11 And I assume it worked. Yes, it added the harbor cashmere crewneck to the clothing collection. It also created the brand. So, here you can see that everything we have tried in plain language has so far succeeded. So, that’s the whole loop. The assistant discovered the cluster on its own. It searched it. It uh looked up the database schema, fetched a few objects from it to understand the context. Um these four tools are currently

13:43 available, but you can expect us to extend the MCP with additional tools in the future as well. That’s it for me. >> Thanks, Yevhen. Uh next uh if we can uh we’ll switch over to Connor, who’s going to uh dive into detail on the uh MCP server for the Weaviate query agent. >> Awesome. Well, uh thanks so much for the introduction, Andy, Yevhen, Alex, Laura for putting together things. Everyone for being here. I’m super excited to

14:13 dive into Weaviate’s query agent. Uh talking about some things that are what’s new and then anchoring with our MCP server, so one of the newer additions to Weaviate’s query agent. So, quick highlight of what’s new, structured outputs, retrieval augmented image generation, effort here’s research mode in our query agent MCP server. So many exciting things. It’s just been so fun to work on. I’m just so excited about this. So, before diving into what’s new, let’s maybe do a quick background for people who aren’t already familiar with the query agent. Uh retrieval augmented generation is about grounding the generations from AI

14:45 systems in your particular data. Imagine asking a question like, “Which waterproof jackets under $200 are in stock?” That’s like a completely nonsensical question to just ask an LLM. The answer is completely about being grounded in your data. And there’s all sorts of examples of having private data, mixing your internal data with public knowledge. Just all these things of having retrieval augmented generation to supplement these generations and things like hallucination verification, making sure that the outputs coming from the systems are actually supported by the data in your Weaviate database.

15:16 So, the query agent is built for RAG with Weaviate. It’s a agent that’s co-designed with the database. And there are all sorts of interesting ways that that uh manifests itself. Some of the biggest areas are schema understanding, query writing, and then these AI layers that we integrate into these applications like arbitrary generation or say chatbots or search or research, which is say re-ranking, long context handling, and then grounded generation and hallucination checking. So, I’d quickly go through some of these core features of the query agent.

15:47 Schema understanding is maybe the best example of this co-design in addition to the kind of like text SQL part of it. But, um we have so we look at the collection schemas, we know how to parse what comes back from Weaviate’s meta APIs to turn it into a prompt as well as a structured output for writing Weaviate queries. We have the data sampler, how we inspect some of the data. Then we have the filter inspector where we look at the distribution of values of these filters. So, for example, if you have I love this example of having say Pokémon games where you might have a property on color, and just color alone the LM would

16:18 have no idea that the particular values it takes on are ruby, sapphire, and emerald. So, it’s an example of how we can co-design the agent to give it information about the data in your data to fit in your database to facilitate these AI applications. So then, uh translating user questions to Weaviate queries is kind of translation of your natural language questions into structured Weaviate queries, being able to write search queries, uh use filters to retrieve objects, combine searches and filters, and then apply aggregations to your data. So, say you have a numeric

16:48 property on price, you can calculate the average price of that in Weaviate like you would be able to in a relational database. Weaviate also has very strong support for structured data. So then, we have some AI layers on top like re-ranking, handling long context. So, say some of the objects that come back from Weaviate are just these massive objects or you’re retrieving images from the database, the query agent is able to read your images, whether that’s like brand assets, graphics, or say PDF pages, you can read these in the query

17:18 agent as well as handling long conversations. And then a huge part, just critical part of these AI applications is grounded generation, making sure every answer is checked against your data, and you don’t have to worry about hallucinations with the query agent. So with this core, we then have four modes, four applications that we deliver on top of this functionality. We have ask mode, search mode, suggest queries mode, and research mode. So these are different ways to use the query agent. Ask mode is sort of a a a pretty general, like the responses API from

17:50 OpenAI’s LLM SDKs, where you can do a lot of things with it. And search mode is us unpacking the under the hood search layer and offering that natively to just use it for search. Suggest queries is a helper with building chatbots where we it’s like a helpful thing with the developer with the with the experience of someone going through your chatbot to suggest some queries to them that your data can answer, as well as suggesting follow-up queries when you’re building chatbots and you have these long conversations. And then research mode, I’m just so excited about research mode. I think it’s just such an incredible technology.

18:22 So the query agent, we built it with customization in mind. We have a lot of features to help you adjust it to what you need. You can import a custom system prompt. You can set your effort tiers. And then we have particular things like if you want to really fine-grain control over the filtering that the query agent is going to be using as it’s looking through the data in Weaviate. So awesome, now diving into some of the new things we’ve been up to. I think all these things are just extremely cool. So starting off with structured outputs. Structured outputs are this really cool way when you have an LLM generating

18:54 anything, you can particularly type the outputs that it’s going to be producing. So say it’s going to be picking a jacket, and it’s going to have a name, a price, and whether it’s waterproof or not. We’ve now added support for structured outputs with ask mode. And a super super awesome part of this is how we’ve added retrieval augmented image generation. So, one of these structured outputs can be an image. So, this is the most popular example internally that we use, creating an advertisement of our mascot wearing our best-selling sweater, and you get this

19:25 awesome kangaroo in a orange cardigan. So, the query agent is able to look at your data to find out who our mascot is. It could be images of this kangaroo guy, and then it could be and then it looks through your images to find what our best-selling sweater is. So, our best-selling sweater is demonstrating this capability of sorting the objects by how many sales they have to retrieve that kind of information, combining this for retrieval augmented image generation. I’m just so excited about the potential of this. So, then we have the effort tiers that have firstly landed in search mode. To

19:57 give you a sense of what this means, we benchmark this on a search benchmark where basically you’re evaluating how accurate the search results are given your collection of say, in this particular case it’s about 500,000 documents, is the gold or the human annotated correct document in the say top 1, 5, or 20 results? Hybrid search by itself produces these numbers. Using the AI search layer in the query agent, we’re able to boost this up, nearly doubling all these scores. And then using the newly introduced high and ultra-high effort

20:28 tiers, we’re able to just get incredibly accurate search by using the query agent. And then just something I think is just amazing. I’m so excited about the future of this is deep research with your data. Who isn’t excited about these long horizon agents, that work say for 2 hours, 10 hours on a task, and putting that pointing that at the data that’s stored in your database, and having it answer these really hard questions, spending a lot of time and compute, and using sophisticated systems. I just think this is awesome. So, if you this is currently in a product preview stage, please there’s a form on

20:59 weaviate.io/productpreviews if you want to join the testing with this and work with us closely on the development of this. So, now diving into the MCP server, a core focus of us with the Query Agent is integrating this wherever you work. We have You can use this in the Weaviate Cloud Console. Personally, I’m a huge user of that. I love being able to talk to my data directly in the console where I’m looking at my collections and looking through my data. So, I think the cloud console is just incredible. As well as this Python client and TypeScript client. And now MC Query Agent MCP server.

21:31 So, this makes it really easy to connect with as Ivan, I think, already gave an incredible explanation of this with Claude Code, ChatGPT, VS Code, Notion, these kind of GUI tools. As well as really, I think, what pushed us to get on get in on this as well with the Query Agent team is the how people build and build agents with LLM SDKs, frameworks like Pydantic AI, LangChain, and DSPy. So, again, kind of maybe maybe a little redundant. I think Ivan already Ivan already gave a great explanation of this, but I’m a pre-built MCP server

22:03 means that you don’t need to describe the tool, deploy it, scale it, and probably most annoyingly is maintaining it. We just have this URL api.agents.weaviate.io/mcp, and you can just instantly plug this into your application. So, that makes it just this one line of code to just copy and paste this from our docs and add in your the Query Agent accessing data in Weaviate using our AI layer around it right into Claude Code. You can also easily integrate this with, again, these frameworks like Pydantic

22:33 AI. has this MCP tool set abstraction. So, this is like how you give them your list of tools and the super powerful thing that we’re seeing people doing with this is they say combine the Query Agent Weaviate MCPs with, say, their data in Slack as well as, you know, read write access. So, it might be able to write messages in your Slack as well as Notion, say, Google Calendar. And so, you have this incredibly simple interface for how you would define these agents in Python if if you want lower level control over this, this is just another example of

23:05 just showing what the syntax looks like for another framework, another popular framework, which is LangChain. So, now going a little deeper into the hood of how we build the MCP. So, we’ll be maintaining this on our docs, and you can find the particular ways that we’re describing the tools, say how many tokens we’re using to describe what the query agent’s ask mode and search mode are doing to your LLM. As you can see for your LLM, I’m sorry. So, we have this tool description, ask mode transforms your query into actionable searches or aggregations, and then provides a final answer to the question. So, we’re telling the LLM what

23:35 the tool does. We tell the LLM what its input arguments are for this tool, whether they’re required or not, and particularly the result evaluation. And this is it to LLM, then and it will get this expanded response that we’re using to fight hallucinations and support verified claims. So, then this is what we use for search mode to explain to the to your LLM, your agent that’s consuming the query agent API, what it is about search searching with the query agent, how it’s different from say hybrid search or normal search, this idea that we’re adding these extra reranking and

24:06 query writing layers to get super accurate search at the cost of a bit more latency and inference cost. So, then again, these are the input arguments to search mode, different parameters the LLM the agent has control over, and then we tell the the consumer of the of the MCP what response format to expect. So, concluding with some thoughts on building with MCPs or using the Python and TypeScript clients, I think the MCP server you definitely give up a bit of control and customization when you’re

24:37 using the MCP server, but then it just makes it so easy to integrate. A lot of people are already building this way, so it makes sense to just meet them where they’re at. But, if you do want more control, I do recommend also to seeing the Python and TypeScript clients. There are a lot of things in there that extra control over with using the native clients instead of using the MCP. So, thank you so much for listening to this talk on the Curry Agent MCP overview the Curry Agent. We’d love to connect with people to understand what kind of software you’re building, your use cases, and see how we can help you use the Curry Agent. So, please just

25:08 email us if you’re interested in that at agentsteam@weaviate.io. >> Connor, thanks. Um and uh before we uh as we switch over to Alex, uh I just want to let everybody know that uh if you have questions, please enter them into the Q&A box or in the chat. We’ll get to those after uh Alex finishes uh his part and um we’ve already got some some good ones coming in. But, with uh further ado, why don’t we uh switch over to Alex Ledbetter, who is

25:40 the uh founder and CEO of Delegance Brokerage. Such a great story and very relevant to um building with AI. Alex, uh I think was uh had a uh job in a commercial insurance brokerage and what struck him above all else was how little how underutilized like like zero utilization >> [laughter] >> of AI at the time and saw an opportunity to transform commercial insurance um by uh through the application of AI and

26:11 it led him to uh down a path that uh saw him uh vibe code a Agentech uh commercial insurance brokerage uh business into existence um last year and uh launched it and um you know, single-handedly and uh as of yesterday announced uh a nice uh pre-seed funding round and is uh off and growing. So, uh very thankful to have uh Alex here today to talk about his story and uh without

26:42 further ado, let’s let’s hear that. Thank you, Andy. No, and thank you, Connor and Ivan for doing such a great job tying all this up and uh thank you guys so much for having me. It’s it’s awesome to be back. I was on the podcast a few weeks ago, so it’s It’s great to make another appearance here at this webinar. Um like Andy was saying, I started off at a commercial insurance company in the summer of 2023. I I was actually heading into the summer before my senior year of college. I went to a liberal arts school and studied

27:12 economics and English. Um I had never written a line of code or read a line of code um prior to March 2025. And I ended up raising about $200,000 from angel investors in February uh 2025 um to push this forward. I realized I did not have enough resources to bring in any external support, so I was going to need to get deep in the weeds and build the product out um the best I could myself. So, it wasn’t uh it wasn’t ideal. It

27:43 was, you know, early days of of vibe coding. Um you know, back then, I don’t think, you know, it’s funny to talk about some of this stuff now, but we didn’t even have plan mode until the fall of that year. Um we didn’t even have um good context windows in any of these models um until about the summer time. So, it was a grind trying to just maintain all of context and um and push across, and that was how I stumbled across Weaviate. It became really important that the model had immediate

28:15 access to all of the different systems and tools I was trying to use. So, Weaviate’s mCP server was critical part of of me building out um our infrastructure. This is our site. So, I’m going to take you guys through the onboarding process, and I’ll talk a little bit a little bit about how we’re using Weaviate, where we’re using Weaviate. Um and ultimately, what I found out was that leveraging this type of vector database architecture unlocks more

28:46 sophisticated servicing capabilities than what a lot of the brokerages and insurance companies are using today. So, we’ll just go through here. Um there’s a lot of information you need to provide to insurance companies to get quotes. Um but I ultimately found out that these insurance brokerages are charging anywhere between 15 to 35% annual commissions. And with

29:21 this underlying servicing capabilities that um WeViate helps unlock, it can reduce the cost to serve these accounts in a way that allows me to charge less for the cost of brokerage services um and help these businesses save on their insurance premiums just through reduced commissions. So, I’m just going through here. But we take all of this information and we are leveraging it to build a comprehensive

29:53 knowledge base for every single insured. So, um uh we won’t do a website this go around. But uh once I do this OTP here, we’re about to upload our documents. Um so, a big part of of the advanced servicing capability and and building out this knowledge base is that all of these businesses have documents. They have these binders. They have these policies.

30:24 I’m sure many of you, your companies have these 200-page policies, maybe you know, 50 to 200-page policies that nobody’s read before. There’s a lot of fine print nobody understands. And I think that’s a that is a compelling reason to leverage this type of architecture because you can get real-time answers to um all of your all of your policy-related questions. And one thing I’ll also say is uh we don’t This isn’t the only way to engage with our platform. We have uh

30:55 servicing capabilities available over email, over We have a full SDK, we have a Slack app, an iOS app. And one of the great things about Weaviate is that with the um custom data schemas that we’re able to create, it was really easy for us to scale our memory system across different channels. So, I was able to tag portal activity versus email activity versus phone activity. And so, having kind of a unified uh memory

31:27 system across all of the different channels that clients engage with our platform, um that was really important. So, because, you know, all of these clients have different ways that they’re used to talking to their broker. Um and so, I wanted to make whatever was minimal friction. Um and so, as you see here, some of these are already starting to come through. Effectively, what we’re doing I talked a little bit about this more in in the podcast episode. Um but effectively, what we’re doing is uh we’re converting all of these PDF pages um into images

31:58 and leveraging multiple types of OCR. I don’t like flat OCR for um these types of documents. You end up with a lot of graphs and charts. And to Connor’s point about you know, Weaviate being able to index images and and query against images, that’s also really important for our use cases. Um so, as you see here, uh we pull out a lot of information from this policy. So, there’s a lot of different documents that we can actually classify against. So, there’s a lot of different agents um doing different things for

32:30 each document type. So, the first one is kind of clarifying like, all right, which of these document types did they just upload because the whole goal here is I don’t want any of my clients to be experts on their commercial insurance. I don’t want them to know what all of these things are, what they have. Most of them don’t. So, I want my clients to be able to just dump in their binders. We can classify everything for them. If they know that something’s wrong, they can override it. But based on the document that we classify, um there’s certain types of structured

33:01 data that we look for based on the classification. And we’re vectorizing everything while also building a table of contents for every single document type. So, um all of this is cooking. We pull out key information from all the policies. So, all of the sub coverages you have, the limits for each sub coverage type, the deductible for each sub coverage type. Um so, all of this gets pulled out. You guys saw it was pretty fast. Um so, trying to get people in through quickly. Um we won’t do this, but there is an

33:32 option to go in and do a full a few additional questions, get some context about or provide some context about your portfolio, um and then get a full audit on all right, maybe you’re over covered here, there’s some optimizations here, some redundancies here. Um and so, it provides that level of analysis, but you can also jump straight into a quote request. So, we’ve got about 25 carriers that we work with today. Um and the goal

34:03 is to make this as seamless as possible for our insured to um answer the questions they need to get quotes. So, that’s also a big part of that extraction process is we can pull out all of those details, all of the information that um you might already have in a corpus of documents. Um and so, you’re not needing to re-key information um because we’re able to leverage what we’ve extracted, what you’ve provided to help um um that information transfer between um

34:33 between insured businesses and and carriers. I I see my job as a broker, it’s really a glorified information transfer um with some servicing capabilities baked on top. Um but we’ve built kind of a full-fledged platform, so um all don’t try to call this stuff. It It won’t work unless you have a real account. Um we we whitelist all this stuff, but you’re able to go in and and um text an AI agent, call an AI agent,

35:04 email an AI agent. You can schedule time with me. We have all of our integrations here, so um our MCP servers are our PWA. We have an iOS app. Um I need to figure out how to put it in here. We’ve got our Slack app. Um and then we have a full SDK, so I’ve actually been working on a CLI product as well. So um I don’t know about you guys, but my experience getting commercial insurance um before this, it was kind of ridiculous. I had to go in, I would fill out a form, and

35:35 they’d expect me to go into a 30-minute call with an underwriter, answering more questions. Um and so I wanted to make an experience where you could do all of the self-servicing capabilities, everything that you would ask your broker to do for you. Instead, you can just drive it in real time because um we have 142 customer-facing MCP tools. Um so any any piece of qualitative information that we get from an insured might trigger um an

36:06 automated policy update, a new COI, proof of insurance, um really anything that a client might need. Um instead of having to talk to someone or wait on them, and I tell people this is so much better than the average commercial insurance broker is 60 years old. So these guys are um not always the most tech-savvy um and they’re not always the fastest. So, having something that can service you in real time built in with a memory system that, you know, is just as good as human

36:37 photographic memory, uh it’s an easier sell to businesses that one want to save, want a better value, and just want better servicing capabilities. So, I think it raises the bar for what businesses should expect. Um this is my DocuSign rip-off. Uh um you’re able to sign broker of record agreements within our platform, um which trigger off downstream. We have a whole agency management system that I also built integrated with this. Um so, that

37:08 gets sent off. Um the broker of record agreement will allow you to um through natural language request COIs across all of the different mediums. You can add more people to your company, um and we also have multi-entity support. So, if you have multiple entities, um you’re able to leverage we built kind of a nested knowledge base infrastructure. So, that’s another uh great use case for Weave AI is if you have um kind of complex multi-entities or or um

37:40 requirements for nested logic, uh that’s really useful for for building out those types of parameters for different tenants. Um and then you can go in at really any time. All those documents that we just uploaded, um you can pull up, ask any questions about them. Um you know, leverage the table of contents to filter out to different things. Um we have a full conversation history, so you can filter by when queries were sent, where queries were sent. So, which medium were you using? Um we can trace

38:10 everything back except for Open AI and Anthropic do not let us uh take the transcripts of chats. They let us take the tool calls used. Um so, that’s less good, but um the point of this is no matter where you are, um, you can you can leverage all of the same servicing capabilities and uh, I’ll just talk about this. Like, what can you do for me? This will probably list most of them. Um, but yeah, I mean, just analyzing your coverage. So, hypothetical claim

38:40 support, identifying structured data within your within your documents. Um, and so all all of this is available for um, 36 different coverage types, um, and also bond markets as well. So, we’ve been we’ve been seeing a a lot of demand from businesses that that feel like um, this is a better value than than what’s being offered today, especially for businesses. I’ve talked to some businesses where it’s I call my broker once a year. I pay a few thousand dollars in in commercial insurance, um,

39:12 every year and I don’t, you know, I I’m not uh, I I talk to him once a year. That’s the most expensive phone call I’m making every single year. It’s tens of thousands of dollars. So, being able to save money and and drive uh, things end to end, it’s it’s a bit more it it’s a it’s more powerful, it’s faster and um, I think that there’s a lot of businesses, especially as more businesses start leveraging agents. Like, I personally, I would never want to handle my own commercial insurance.

39:44 I’d rather pass it off to to Claude or Open AI and and never have to think about it and, you know, leverage all of the contacts I have locally. So, that’s that’s kind of uh, how I’ve been trying to build things out. Um, we have built-in premium financing solutions as well. So, if you need financing, um, at point of sale, um, we have that as well. We can help uh, streamline the claim submission process. I don’t know how deep we want to go on all of this, but um, happy to also talk about the differences between MCP and APIs and and where MCP

40:16 wins compared to other um, solutions. Yeah, but >> Let’s start talking about the the building of it. And this is incredibly impressive and and such a great observation. I mean, insurance and commercial insurance such a document heavy process and business and such a natural fit for AI. It’s a really about extracting information and answers out of um big insurance documents to um answer customer questions and make decisions about who to um they finally

40:46 get your insurance with. But yeah, so let’s let’s start the Q&A a little bit. We’ll start with questions for you, Alex. We had a couple of questions about you know, why Weaviate? Why would I use Weaviate instead of PG Vector? Or why would I use it instead of other vector databases? I’ll give you a quick answer and then I’ll let Alex give you his answer since he he made that decision uh for real. And I think uh with respect to

41:17 uh something like PG Vector, it it really comes down to the best tool for the job. You know, Weaviate was built specifically for use cases like Alex’s. You know, what you’ll find with respect to PG Vector is that number one, developer experience. You know, because Weaviate can integrate your models right into the database, it means a lot less uh code moving data, a lot less pipeline development, uh moving data between the database and and models. Uh like all vectorization, uh generation

41:47 of embeddings can happen within the database, so it simplifies development. And then in terms of runtime and querying, as the database grows, uh you know, indexing your vector data in PG Vector is going to um become painful, painfully slow. And with Weaviate because of uh the wider variety of indexing, compression, and query options that you have, the it it’s able to scale and perform better at really any level. And we’ll share as part of the follow-up,

42:18 I’ll share a link to a blog post that booking.com wrote just a few weeks ago explaining exactly why they switched from why they decided to centralize all of their vector applications on Weaviate moving many of them off of elastic or open search to do that including some benchmark results they did. So I’ll share that with the follow-up. And then with respect to vector databases like Pinecone, again the developer experience is usually what

42:48 um distinguishes Weaviate from others. Also the deployment options features like multi-tenancy is kind of a native thing that Weaviate does quite well and also the deployment options if you’re talking about Pinecone specifically, they are a cloud-only service whereas Weaviate can be installed and run anywhere you’d like. But Alex, what was your you know, you had to make this decision once. What was your thought process? >> So I will say um I did have some experience like in

43:19 between 2025 and uh when I had graduated, we were built I was working as a an AI product manager for a consulting company that was building uh custom knowledge bases. It’s funny that this was something that like a lot of the big businesses were paying for just a few years ago, but um they didn’t know how to make their own custom knowledge bases. Um so we were leveraging like the engineers came in immediately recommending Weaviate. We

43:49 had tried to use other solutions and the setup was just way too complex and we needed something that ultimately we could pass off to this larger enterprise. So they needed something um that they could self-host uh ultimately and the fact that Weaviate um is technically, you know, open source and that that was available, we were able to build all of the infrastructure needed on our side prior to handing it off to this larger

44:20 client. So, that was really important. At the time, I think this is still probably the case, but we had looked at like it was technically possible. I think it still is, but and maybe it’s gotten a little easier, but we had looked at RDS as a solution with an AWS and it was a nightmare getting all of the metadata tagging lined up how we needed it to be. So, something like Weaviate is just I mean, for me a lot of this stuff is how quickly can you go? How easy is it to set up? Can your agent set it up for

44:51 you? And Weaviate can do all could do all of those things before a lot of the competitors and and they’ve stayed really far on the on the cutting edge, I would say in terms of like, you know, the query agents. There’s There’s not a lot of platforms with with built-in querying out of the box, especially to this level of quality. So, and I I think I know I had a dedicated account executive, which was really important for me at the time. Like, I don’t even

45:23 get account executives for my bank right now. So, having that kind of hands-on technical support as a Vibe Coder was really important to me as well. There were times when like my account executive was able to provide me with resources, with documentation for different problems I was running into, and I was able to just inject that context straight into my agent to help me resolve those issues. So, I was able to get good quality of like level of support from Weaviate as well.

45:55 I’m not sure other companies are are providing those types of resources or support um to all of their clients. So, that’s just my experience and there’s a lot of other technical things as well. I put something in the chat about this as well, but the schema flexibility is also really important. When you have a lot of these out-of-the-box tools that are really simple, it can actually be somewhat limiting because of how templated they are. Like, I never entertained N8N for

46:25 that reason. It’s Yeah, like I I think it’s important that you have this flexibility to This is my two cents. >> Thank you. Let’s see another question here for Yvonne. The The MCP servers Can you talk a little bit about the availability of of those? >> Of course. If it’s about the open-source question, then um the Weaviate database MCP server is of course open source, so it’s part of the

46:56 database itself. If you spin up your own cluster or even if you use the Weaviate Cloud, you have it by default. So, no setting up, no kind of fees or something like that. The query agent is a separate service that we offer. At Weaviate as part of Weaviate Cloud, you do have a free number of requests, queries that you can send to the query agent, but that one is not open source. It’s a managed service.

47:26 >> And I think another question is it uh are you able to connect to it from Cloud Desktop? >> Yes, of course. You can also configure it through I’m not sure where settings preferences in the Cloud Desktop app. I have tried it out myself. And it should work the same way as with Cloud Code. >> Great. And question for you, Connor. I think you answered this in um in the box, but just for the recording. Uh somebody wanted to know more about

47:56 the evaluation outputs for final answer is partial answer missing information sources. >> Oh, yeah, awesome. >> Clarify the question and and answer it. Yeah. >> Oh, yeah, cool. So, yeah, having grounded generation and making sure that your responses aren’t being hallucinated, they’re actually based on the data in your database is one of our core focuses with the query agent and the way we solve this so far is you have an optional argument that’s on the result evaluation. You can turn it off

48:28 if you don’t want to spend the extra inference cost for this, but when it’s enabled, what we do is from the final answer, we apply another LLM inference that we’ve tuned and optimized to assess that final answer with the retrieved context, evaluate, and populate fields like if it’s a partial answer, if it’s missing information, and then just validate the sources that it’s returning in that structured response attribute. So, those are some of the things that we’re doing towards fighting hallucinations and making sure this is a really reliable retrieval augmented generation API. And

48:58 I and we’d love to work with you. If you have you know, if you have if you’re not finding it if there’s any particular additions you want there, any tests you have, please contact us. We’d love to see how we can help further. >> Great, thanks. Alex, question about your stack. Like what is your What else is in your vibe your uh AI coding stack besides Weaviate and the gazillion MCP servers here? >> Yeah, I’d say I’m always looking to optimize. Like even now, I

49:30 um I basic I found out 2 weeks ago after our podcast, Connor, that one thing I was describing to you, there’s actually like a technical term for it. Like coming from a non-technical background, um I’ve realized a lot of the stuff that I’ve created or just kind of intuited as something I would need or would want um got kind of just like first principles building. A lot of it already exists or has like technical names behind it or so I’ve I’ve been kind of like rediscovering things and

50:02 I I want to talk about our software factory that I built. So something I mentioned this in the in the comments but when I was vibe coding I was spending seven I spent seven to eight months building before I had the first company come on to my platform. I went through my own user experience flow over 700 times and so a lot of that iteration it was critical that I was using tools or infrastructure that I could control or

50:33 feed context easily into my agents via terminal. I was building last year I did everything in cursor. I was a top 5% cursor user last year but I switched over to Claude code in the new year I think like a lot of other people. Um So oh I just got to ask is my platform agentic first? Yes, we are. I I integrated web MCP. So I did want to mention that as well. Web MCP we’ve got I know that there’s a lot of strong thoughts about MCP versus

51:04 API versus CLI. One thing that I know for certain web MCP is critical infrastructure to have if you want things to be agent first. It makes it so browser agents that visit your platform or visit your site are able to use all of the functionality easier than if you had them just doing like a screenshot loop with you know uh just continuously trying to find the right fields within a form or or find the right functionality within your

51:35 page. Web MCP is great for that. So I that’s how I’ve been trying to build. I think we’re heading toward a world I was reading where by 2030 we should be seeing three to five trillion of payments processed through a gen tech commerce. So, I’m anticipating a world I think personal agents obviously a lot of movement there over the past few weeks, but I bet within the next few years, we’re going to see business adoption and enterprise adoption trailing soon and

52:08 there’s a lot of back office operations or administrative work that people would rather leverage agents for and agents are already pretty good at at you know, especially multi-agent architectures. I think you can reduce the risk of hallucinations or things going wrong if you have multiple agents that are looking at outputs and kind of collaborating together. It It makes it harder to prompt inject. It makes it harder for outputs to look bad and right

52:40 now costs are going down. So, I think long-term there’s going to be a lot of multi-agent architecture. Like the way I see it, I see it as like a test and you have two students that both average like an 80% on a test. You know, we’ll we’ll keep it conservative. So, like OpenAI and and Claude or Anthropic, they’re both 80% level students and they take a test together. Together, there’s not one that went over lap. So, they actually

53:12 score higher on the test when they work together. So, I’ve been I see someone someone’s talking in the chat a lot to me also. But yeah, I I Superbase was great. AWS was great. The CLI for AWS was great. Superbase, that was really useful. Um I have my [snorts] my task definitions are exploding with different tools that I’m using. Um But those are probably like the the

53:43 core ones. I think, you know, Weaviate AWS and Superbase was another platform I used. But, if I were to do it again, I didn’t realize that you could use Weaviate for structured data. Connor and I talked a little bit about that as well. If I were to do it again, I would simplify and I would I would use Weaviate for structured data and unstructured data. And that would be even better because they’re in the same systems and it’s easier to retrieve and keep in context. One one less tool call that could go

54:14 wrong, one less query that could go wrong, one less dependency. >> Cool. All right, we’re at the top of the hour. I want to thank Connor, Yvan, Alex for taking time out today. Really enjoyed the information. Hopefully, everybody in the audience did too. We will follow up with a link to the recording of this event for anybody who missed any of it. Also, link to some other resources like the the podcast interview between Connor and Alex where

54:44 Alex goes into more detail on the the building of Delegates and I’ll share the the link to the booking.com blog post blog post where they talk about sort of standardizing on on Weaviate for their AI applications. And with that, we hope to see you again at a at event in the future. Thanks very much for attending today, everybody. Bye-bye. >> Thanks, everyone. >> Bye. >> Thanks, everyone.