"At the core of AI is search, and at the core of how AI does search is vectors... People who are working on agents, people are trying to build generative AI applications... may not necessarily realize at first glance that, oh, the thing that underpins emerging AI is actually search."
With massive context windows and new agent frameworks, do vector databases still matter? Ram Sriharsha, CTO at Pinecone, joins the conversation to make the definitive case that they're more critical than ever. He explains that at the core of all AI is search, and externalizing this function is non-negotiable for security, auditability, and control.
Ram offers a clear starting path for engineering leaders: begin with simple Retrieval-Augmented Generation (RAG) applications, but immediately implement a robust evaluation framework to manage hallucinations and ensure quality. He shares his perspective on the skills that matter most now, arguing that curiosity and the rise of the generalist engineer are critical in an AI-powered world. This episode is a guide to building the AI stack from the ground up, from using AI as a "good junior engineer" for testing to cultivating the engineering mindset of tomorrow.
Show Notes
- Learn more about Pinecone: pinecone.io
- Pinecone Tutorials & Blog: Explore guides on RAG, vector databases, and more
- Connect with Ram Sriharsha: LinkedIn
Transcript
(Disclaimer: may contain unintentionally confusing, inaccurate and/or amusing transcription errors)
[00:00:00] Andrew Zigler: Hello, and welcome back to Dev Interrupted. I'm your host, Andrew Ziegler, and today we're sitting down with Ram Sriharsha, the CTO at Pinecone. Prior to joining Pinecone, he was the vice president of engineering at Splunk, where he began as a senior principal scientist. And before that, he was a product and engineering lead at Databricks for their unified analytics platform for genomics.
[00:00:21] Andrew Zigler: Ram, welcome to the Dev Interrupted. We're so excited to have you here.
[00:00:24] Ram Sriharsha: Thanks, Andrew. It's great to be here.
[00:00:26] Andrew Zigler: we're here at ELC, the Engineering Leadership Conference here in San Francisco, and you gave a talk today. Yes. Uh, about vector databases. Why are vector databases still important?
[00:00:37] Ram Sriharsha: That's a great question. In fact, my talk goes into exactly this topic. Uh, the reason I wanted to give the talk in the first place was, uh, first of all, there's a conference where people are talking about AI agents are a hot thing. It's uh, it's an emerging technology that people are starting to get familiar with.
[00:00:52] Ram Sriharsha: Through all this, people may not realize that at the core of AI is search, and at the core of how AI does search is vectors. [00:01:00] So I wanted to give a talk to frame the problem from that perspective and to kind of maybe educate the audience a little bit. That if you take this to the logical conclusion and think about how do you now scale search, you have to have vector search and vector databases at the core of this.
[00:01:16] Ram Sriharsha: And it, it's a new way of motivating vector databases for a new audience. I think people who are familiar with recommendation systems, people who have done information retrie well, they kind of get it. Uh, but people are working on agents, people are trying to build generative AI applications and so on.
[00:01:31] Ram Sriharsha: May not necessarily realize, at first glance that oh, the thing that underpins ai, emerging AI is actually search. So I just want to give a new framing to that.
[00:01:41] Andrew Zigler: So would you say that it's like framing a building block that folks should be using and getting them educated about how they're used to build higher level software? Yeah. Vector databases are really important in the AI space, and I really want to understand more from your perspective of why [00:02:00] their prevailing importance in search will probably never go away.
[00:02:03] Andrew Zigler: And understanding the role that the Vector database plays within the engineering world for people that are building agents or working with AI tools and. Most importantly, what are the ways that engineering leaders can think about those kinds of tools, those primitives, those catalysts for change and use them within your own organization?
[00:02:24] Andrew Zigler: What kind of perspectives do you hear from leaders?
[00:02:26] Ram Sriharsha: Yeah, that's a great question. So, uh, first of all, the reason why vector databases don't go away is that if, if you're building agents you need knowledge. You wanna be able to externalize this knowledge, you wanna be able to audit it, you wanna be able to redact it, you want to make sure it's secure. And the moment you start thinking about that, you need vector databases and it becomes a foundational component.
[00:02:48] Ram Sriharsha: And, uh, now it becomes the question of how do you take this foundational vector database foundational language models and put them together to build a stack for your agent? Right? [00:03:00] what I see, people often do is they start simple. There are, uh, good recipes out there today called rag retrieval, augmented generation.
[00:03:10] Ram Sriharsha: There's a lot of, uh, knowledge Pinecone ourselves. Uh, we have lots of vlogs. We have lots of templates of how to build rag systems. We even have what we call an assistant, which allows you to just drop. Data and ask questions. So we tend to try to lower the barrier to entry to this kind of building agents as much as possible.
[00:03:31] Ram Sriharsha: So that's where I would start. If I'm, if I'm a, you know, an engineering leader who's trying to empower a team to build agent workflows, I would start with something like an assistant. Start consuming it, start building, uh, knowledgeable agents, and then slowly go deeper into the stack as you kind of gain more expertise and also start putting more guard rails.
[00:03:51] Andrew Zigler: So when you work within that iterative. It's about just starting and then measuring what works and what doesn't. What else matters? As an engineering [00:04:00] leader, how can they equip their engineers to really be thinking about unlocking more value
[00:04:04] Ram Sriharsha: It's a great question. I think, uh, to me, starting simple matters, putting evaluation frameworks matter because LLMs hallucinate, uh, no retrieval is perfect.
[00:04:16] Ram Sriharsha: You kind of need to know to. Tweak or to tune your retrieval steps, you kind of need to know what to do about hallucination with language models. You need to know when your agent is actually returning the right relevant results and so on. So there needs to be like an evaluation framework. That's probably something you think about first.
[00:04:33] Ram Sriharsha: Uh, and then kind of go deeper and deeper and start building more sophisticated agentic workflows from there. So I, I generally would suggest start simple, but start putting guardrails, start putting evaluation frameworks.
[00:04:44] Andrew Zigler: And what would you say is like a prevailing skill that leaders should be inspiring or. Otherwise teaching their engineers to pick up in today's engineering age.
[00:04:55] Ram Sriharsha: I think curiosity, Is super important. Um, being a [00:05:00] generalist is very important, right? Because now you have, you have, uh, language models and AI tools that you can use for kind of specific tasks and they're gonna just keep getting better. But you need to be a really good generalist now, uh, more so than you had to be before.
[00:05:15] Ram Sriharsha: I think so I think, uh, breadth of knowledge. Curiosity is what I would really kind of look for in engineers and kind of guide them towards that. Yeah. And to be able to ask the right questions out of your language models and them correctly, and to know when they're hallin, know when they're giving you the wrong results and not just accept what they're
[00:05:35] Andrew Zigler: Having that, that clarity of intent, knowing what you're asking for when you go into the situation, which is way harder, by the way, harder than it looks on the surface. Everyone thinks that they're coming in with the best possible way of explaining something, but until they actually try to convince another engineer to build it or an LLM to help them do it in an afternoon, you know, that's where the rubber really meets the road.
[00:05:55] Andrew Zigler: And that clarity is really important, I think. when. Teams are [00:06:00] ultimately coming together to solve problems. Right now, AI is kinda like a force multiplier and like a for all the good and all the bad, right? If you don't have process, it's gonna make things really messy. If you have a lot of process, things can get really ated and rigid.
[00:06:15] Andrew Zigler: Right? I'm curious, within Pinecone, how do y'all approach that curiosity and how do you, um, create the space for your engineers to reinvent on top of what you're building?
[00:06:26] Ram Sriharsha: It's a great question. I mean, uh, first of all, we are very heavy users of Cursor. You know, I myself use Claude, for example. Uh, so we are, we are very big users of ai. Uh, we also do kind of take advantage of Gemini and, uh, tools like that for code of use. So there's a, there's a lot of first level kind of reviewing and the first level kind of problems that AI is very good at doing today.
[00:06:54] Ram Sriharsha: You can think of it as a good junior engineer like AI tools are getting pretty good at. [00:07:00] We leverage it for that a lot. What we typically tend to do though is to take it from there and refine it. And that's, that's a process that, you know, is purely left to software engineers today. AI cannot really help you as much there.
[00:07:13] Ram Sriharsha: Uh, you need to know exactly what to take out of what it gives you, how to iterate on it and when do you think it's good enough. So I, I think, I think that's where we do find really good use of it though, is in testing. So, uh, you know, particularly Pinecone is building vector databases, is building kind of AI infrastructure and so on.
[00:07:32] Ram Sriharsha: And if you think about traditional databases, they're very complex, uh, systems that need to be accurate and across a spectrum of things, right? Yeah. So, uh, correctness is a very challenging problem in database literature. Typically when you have to build correctness testing and things like this, it'll take months and, and teams of engineers to build it.
[00:07:54] Ram Sriharsha: Yeah. Today with ai, I think it acts as a huge force multiplier in the testing [00:08:00] and the first testing process of databases. So we use it extensively for that, and we found, and it's also a good problem because that's a problem that's in the wheelhouse of a AI today, which is well specified. It's code, it has structure to where you, you, you're asking it to write tests and verification stuff.
[00:08:18] Ram Sriharsha: It's really good at that. And, and, uh, the risk to having a test that is maybe not correct is less than the risk to having a production code that's not correct. Yeah. so we tend to leverage it a bit more, uh, I would say with a, with a bit more, uh, flexibility in testing. Yeah. Obviously with production code, we have, have far more guardrails and we, we really use
[00:08:41] Andrew Zigler: we're just talking about the space we are playing and creating and trying things out because we're all trying to figure these things out right now. There's a lot to
[00:08:48] Ram Sriharsha: Yeah. What I would do is, uh, you know, it's a, it's a great way to start by having AI write your tests or, in particular what's called property testing, which is, it's like writing a test that [00:09:00] captures a scenario of testing. Uh, it was complex to write before, but it's one of these things that you can easily teach an LLM how to do and it can do it for you.
[00:09:09] Ram Sriharsha: So these things can unlock, uh, for example, things that used to take days, take minutes now.
[00:09:14] Andrew Zigler: That's like a really interesting unlock too, because you get this world where you're able to like. Nail down the determinism while you're working within a probabilistic space like you're engineering in this like latent area where nothing is yet defined until you or the LLM talk about it or write it down.
[00:09:32] Andrew Zigler: And so by working with a really rigid structure of like. Having that conversation first, everyone's talking about like the spec based coding, right? Getting that really good markdown document, creating these artifacts of the, of that coding process that we didn't have a few years ago. Now, whenever I, I start a new project on Git, it's often something I spun up in an afternoon, maybe while maybe cursors walk, working on it while I'm eating lunch, sometimes, generally.
[00:09:56] Andrew Zigler: And so you come back and then you, you push it and you want others to try it out and play with [00:10:00] it and see what it is. And I find myself constantly. Checking in those artifacts, those documents, so that, that, that, that spec md, that agent MD that we wrote together. and that becomes just as important as the code that came with it.
[00:10:12] Andrew Zigler: And so it's like a new way of working with engineering knowledge work and sharing it with others. And I wanna, I wanna double click on something you said about. The rise of the generalist. I think that's a really powerful message. I think we've talked about a bit before on Dev Interrupted, we had Lee Robinson here, uh, from Vercel, and he talked about not so much the end of specialization, but rather the rise of these hybridized engineers that just don't accept roadblocks when they have a problem.
[00:10:42] Andrew Zigler: As a team in the past where, oh, you know, in order to figure this out, I need to go talk to the UX team. I need to set up a meeting. I need to go do whatever. You get these fragmented bursts of coding and working. But now with ai, you can encapsulate all that in an afternoon. The things that usually help make an engineer get stuck where they'd have to swivel and ask that backend engineer, [00:11:00] they can now ask the LLM.
[00:11:01] Andrew Zigler: So yeah. So how do you see that force multiplier affecting folks? Even within your own organization?
[00:11:06] Ram Sriharsha: Yeah. Yeah, I think, I think, uh, the reason I think generalists are very, you know, uh, I think being a generalist is. Particularly, uh, force multiplier right now is because you can think of AI as kind of having like a small team right next to you. Right? Uh, so you could be a team of size one, but with language models and with agents and with ai, you suddenly have a team of, of fairly substantial size.
[00:11:30] Ram Sriharsha: Now, if you are a generalist, you can fully utilize them as a specialist. You're limited in some sense from utilizing that. Now, that doesn't mean specialists go away. I think specialists are extremely important, but being a generalist has like a special power now, which, in some sense wasn't there before.
[00:11:46] Ram Sriharsha: I think. I think there's a big unlock from AI from that perspective. So we do encourage our, first of all, encourage our engineers to tinker. All of our engineers tinker. we have them work on very different aspects of the systems. [00:12:00] So we generally, in fact, in the database team for example, we don't really have special, I mean, we have some specialists, but we have more generalists than specialists and generalists tend to work on very different parts of the stack.
[00:12:11] Ram Sriharsha: And, uh, in some sense, AI is helping them really unlock and kind of bring them all together in a way that's just harder to scale before. So, uh, I find that as a pretty big multiplier for the team itself.
[00:12:24] Andrew Zigler: There's also this new experience happening for engineers. This is something I lived. Very vividly recently I sat down for a, a Vibe, coding Hackathon with block, open source, uh, block, uh, goose platform. Uh, and I had a partner. So he and I were working together in this hackathon and we're using exclusively Goose to basically command and write our project for together.
[00:12:47] Andrew Zigler: And it was, it exceptionally hard to bring our worlds together in a way that I had never encountered before in a hackathon, because I've done lots of them. And when you're all working with people, you're just shouting things out. It's like line chefs in a, [00:13:00] in a, in like a, a kitchen, right? You just keep everyone updated.
[00:13:02] Andrew Zigler: But. In this world, me and my partner were both commanding our own siloed teams of AI engineers that were all doing all sorts of stuff, and then we had to find a way to marry those worlds together and it was exceptionally hard. Uh, I'm wondering how do you think about that problem for engineers that work on separate problems and then bringing their AI collaboration together?
[00:13:22] Ram Sriharsha: I, I think. Being a really good generalist helps, right? Because you kind of need to know how to combine the two kind of AI teams, so to speak, uh, which, which is more natural if you're kind of used to putting these parts together and being a really good generalist and so on. Uh, the other thing we do is, other thing that I think is helpful is if you're kind of dabbled in those areas to begin with, right?
[00:13:44] Ram Sriharsha: it's much harder to put things together if you have actually not had some amount of expertise in those areas. So it's like building a agent end-to-end, but I haven't really built a UI before, so I'm far more trusting of that language model and I'm far less capable of, [00:14:00] you know, maybe course correcting it or
[00:14:03] Andrew Zigler: you lean on it
[00:14:04] Ram Sriharsha: exactly.
[00:14:04] Ram Sriharsha: I lean, I lean on it far more, uh, than I would if I was a generalist. Right. So I myself am not a true generalist in that sense because user interfaces are some, an area I've not worked with.
[00:14:15] Andrew Zigler: Yeah, exactly. You're using it like a fill in the gaps, like, like we talked about a moment ago of just not getting stuck, not accepting that you're gonna get stuck and keep going. Right.
[00:14:23] Ram Sriharsha: Yeah.
[00:14:23] Andrew Zigler: I'm, I'm wondering about your, your own background from other engineering roles that you've had, and you have a science background, right?
[00:14:30] Andrew Zigler: You ap you approach this from a very methodological way. How does that influence your, your, your role in your leadership as the CTO?
[00:14:39] Ram Sriharsha: It's a good question. I find that I think my, you know, in general, a kind of a physics or a scientific background, I actually find it really helpful because it makes things a little bit simpler for me in, in how I think about problems, which is instead of, you know, trying to remember a lot of facts and sort of trying to kind of keep a lot of things in my [00:15:00] head, I try to build a model of the system.
[00:15:02] Ram Sriharsha: I try to think about the model of this thing that I'm building, and then I use that model to kind of infer about, oh, if this is the momentum model of the system, then this is how it should be behave. And if it's not behaving like that, then I can go back to the basics to figure out what's happening.
[00:15:15] Ram Sriharsha: That's something you are trained as a physicist, right? Like in physics for example, you cannot really understand the universe, so you really boil things down to very simple components that you can piece together and build models out of. Uh, and that's stayed with me for forever and I find that's extremely powerful and useful, even in computer science, even in AI and every other field that I worked on.
[00:15:36] Ram Sriharsha: So I try to instill that in engineers when I'm mentoring them is. Always form a mental model of the system. And systems have, uh, a good mental model if they're simple. Right. The moment your system becomes complex, your mental model is no longer a mental model. So simplicity is in some sense forced on you if you try to want to keep a simple mental model.[00:16:00]
[00:16:00] Ram Sriharsha: So this is what I try to inculcate in engineers. Yeah. And it's, it's worked. I think, uh, when it works, it's very satisfying because you can keep a lot of things in your head. While keeping just a simple mental model.
[00:16:12] Andrew Zigler: Yeah. Simplicity in engineering is, is always best if it's easy for people to understand, especially because we're experiencing right now, this democratization of engineering and people getting access to building that before, never really had ability to do so. And in that world, you know, you see a lot of stuff and you read a lot of stories about people repeating or learning really fundamental engineering mistakes that are really, that we use to like.
[00:16:40] Andrew Zigler: Have that we have used to ship and create really great software in the last 10 years. Right. And then it's like all of a sudden AI's on the scene, everyone threw all that away and then they're making all those same mistakes again. And then we're relearning Yes. How those primitives build and, and go together with ai.
[00:16:56] Andrew Zigler: And in that world, do you see a a, a place where, you know, [00:17:00] non-engineers are. agentically coding going something like super base or V zero and spinning something up. And, and, and when they do that, is there a, a companion path for things like vector databases to be so simplistic? Like what you just said, that even they could pick, even, even grandma could provide code, that app that has a rag database in it.
[00:17:19] Andrew Zigler: Yeah,
[00:17:19] Ram Sriharsha: that's, that's a great question. So I think that's a, one of the things we are working on is to make, uh, vector databases so simple. So easily integrated with, uh, the kind of general tooling and framework that people are using for agents and whatever they're building, but they don't have to think twice about it, which is you can start small scale up and never have to think about the databases at all.
[00:17:44] Ram Sriharsha: That is one way to do it. The other approach of, you know, I just pick the tools that I have and I kind of just build something and go with it, I think that's fine as obviously we. Use Pinecone and kind of scale up, uh, [00:18:00] easily. I think the other approach also what happens is that people kind of do that. You get to some traction and you realize that, oh, now I have to scale and I'm meeting scale challenges. And then at that point we want to be seamless to be able to switch. Right? So there is the other aspect of it as well that we care a lot about, which is, yes, we would like to catch everybody very early, but we also seamless to kind of help people transition to scale.
[00:18:25] Andrew Zigler: And when you take a system like that to scale, a really important part of that that we dug into recently on this podcast is evals. And, and you mentioned that earlier about making sure you're evaluating things as you iterate. Because when you're working in this like latent space of, of AI right, uh, it can be really easy to miss those small details that are gonna compound and compound and compound.
[00:18:47] Andrew Zigler: And that's what evals are really great at catching. when we sat down with Andrew McNamara, uh, he's their head of a applied ML at, uh, Shopify, and he talked about his, their eval framework that they created for their agents. Yeah. And about how it was so [00:19:00] fundamentally important in order to take any kind of agentic system to production.
[00:19:03] Andrew Zigler: You know, what are your own perspectives on, on evals and how engineers should be using.
[00:19:06] Ram Sriharsha: That's a good question. So, uh, in the case of Pinecone, we are like once one level below in the stack, right? So, uh, when we think about evals, we think about evals for the vectors database itself. that's usually called recall. It's You know, things like that. And we care about that. We measure that We have benchmarking systems that, uh, tools of the open source that allow you to measure it against any database, including us, right?
[00:19:29] Ram Sriharsha: But then one level higher, we build these, uh, things called assistant, which is, uh, kind of a layer on top of the database that access a knowledge system, which is you throw all of your unstructured carpa and we can search it. Now at this layer, your search quality has a different meaning, right? The question is am I returning the relevant results, right? Is my search returning the relevant results? Secondly, if I'm taking this information and feeding it into a language model, does the language model find it useful? Right? And we have, uh, evals that [00:20:00] allow you to measure that as well. So for us, that's the level at which we operate because we are like one level below, like where a Shopify or any customer is gonna be using their evals.
[00:20:11] Ram Sriharsha: Now, of course, if you are somebody who's building, Shopify, say Shopify, on top of using Pinco or using assistant and so on,
[00:20:18] Andrew Zigler: Then they have
[00:20:18] Ram Sriharsha: you're able Exactly.
[00:20:20] Andrew Zigler: of, of how those agents are performing
[00:20:22] Ram Sriharsha: Exactly. So
[00:20:23] Andrew Zigler: just as important.
[00:20:24] Ram Sriharsha: is is exactly
[00:20:25] Andrew Zigler: one layer down of the actual data
[00:20:27] Ram Sriharsha: Because that tells you whether the search system you're using or director database you're using is doing its job.
[00:20:32] Andrew Zigler: Right. Because you have to evaluate where the problems are. Right.
[00:20:35] Andrew Zigler: And the, that's like the great thing about using evals is that you're, it's catching blind spots and things that you can't really write tests for, that you're only gonna be able to see by iterating and then seeing what it throws back at
[00:20:44] Ram Sriharsha: Hundred percent
[00:20:45] Andrew Zigler: It's like a game of catch. Right.
[00:20:46] Ram Sriharsha: Hundred percent. And, and what's, what machine learning and generative AI has done over the last several years is made us from having to be machine learning experts to becoming eval experts. Yeah. See what I'm saying? I think that's where, that's where [00:21:00] we need to be spending more time.
[00:21:01] Andrew Zigler: Another thing I wanna ask about is, um, understanding intent of your users and what they're trying to do, and then connecting the dots with what's available for the LLM. I think we're in this scenario where it's really easy to overwhelm your LLM or your tools or your systems with. Way too much stuff.
[00:21:20] Andrew Zigler: Like we've heard about, like in the MCP world, like if you hook up a hundred MCP servers to your LLM, it's gonna get confused. It's gonna use the wrong tools. There's 10 tools called update, how's it gonna know what to pick? And it fills the context window, all of these problems, right? And so it's about being precise and about understanding the, the intent of the user and meeting them in their place every single time.
[00:21:41] Andrew Zigler: Uh, and this is something we talk about a lot on Dev Interrupted, and I'm curious. What, how you think about that intent problem and is it different for people that aren't of an engineering background versus are, how do you think about it?
[00:21:52] Ram Sriharsha: Yeah, I mean. I think when I think about intent, for example, uh, I go back to kind of why we built a system in [00:22:00] the first place, which is, this is about a year back, slightly more than a year back when, uh, people started building gen gen AI applications and they started using vector databases and so on. And then there were all these questions.
[00:22:14] Ram Sriharsha: Should I use a just a vector database? Should I use some other document database together? Should I use a graph database? How do I chunk. What do I do?
[00:22:22] Andrew Zigler: Right. What do I chunk, how do I summarize it? Like there's a whole bunch of little steps. It's not just turn on
[00:22:27] Ram Sriharsha: exactly. And when, when we looked at that, we, in fact, we looked at a stack that had hundreds of components, right?
[00:22:36] Ram Sriharsha: But then we were starting to ask the question, what do people really care about? Right? The problem that they were trying to solve at that point is hallucination, which is, I have a corpus, and I'm asking a question. I'm getting an answer. I just need people to attribute this correctly and to know that I'm not giving wrong answers.
[00:22:52] Ram Sriharsha: Right? So we focused on that problem. We're like, okay, how does Vector databases help us solve this problem? If you need to augment it with some [00:23:00] other things, what does that argumented stuff look like? Right? And I talked about some of that in the talk today, which is vector databases are evolving from what they used to be a few years back to the search system that's needed, To be able to answer this question. And that's, that's the way we started looking at this problem. We did just the bare minimum required at that point to be able to problem that really mattered without having to deal with all of the complexity and start kind of going down the rabbit hole of all the things we needed to tweak and tune. So that's an example of kind of how we approach this.
[00:23:33] Andrew Zigler: If you were to ask or rather talk to an engineering leader right now about what you think is the most important thing for them to focus on in the next few months as we enter into a new year. The new cycle and the development cycle around AI is so fast paced that any kind of prediction you'd have about a month from now would be wrong by the end of the week.
[00:23:56] Andrew Zigler: Just at the speed of which things come out. Are some big rocks [00:24:00] that you just keep coming back to as you evolve in the AI space and you lead your organization to it that you think everyone should like be sailing towards.
[00:24:08] Ram Sriharsha: Yeah, I would say that, uh, I still find a lot of organizations hesitating to build, uh, gene applications, for example, because of the kind of probabilistic nature of this stuff, because there's still wrinkles, there's still unknowns, and so on. What I generally advise people is that. Start building, start, start kind of putting these things together, even in its current form, it is, uh, massively useful.
[00:24:36] Ram Sriharsha: And once you start, you'll get familiar and get comfortable with this kind of thing that gives you probabilistic answers. Things that that gives can potentially hallucinate sometimes, but it's still very useful in applications that can kind of tolerate it. And then start putting guardrails together and just start kind of working away from there.
[00:24:55] Ram Sriharsha: I think it's, uh, it's time to start using it and time to start, time to start kind of [00:25:00] getting familiar with it, rather than waiting in the sidelines and hoping that this thing gets perfect. Right. Yeah.
[00:25:06] Andrew Zigler: And so like there's a lot of opportunities in the immediate future, even just right now for folks.
[00:25:16] Andrew Zigler: As you talk, as engineering leaders, is there a recurring question that they constantly, you, you find them constantly asking you about what they should be doing or thinking?
[00:25:25] Ram Sriharsha: I think it's mainly about how do you deal with the fact that these things don't give you, Answers that are, uh, you know,
[00:25:33] Andrew Zigler: How do you deal with the fact that they're not deterministic, which is their, by the way, their superpower, which is why, by the way, we turn to them, is because it unlocks an ability to build things that before were un commercially available to do.
[00:25:45] Ram Sriharsha: to, to me, that's, that's what it often keeps coming back to and some applications it just doesn't work, In some applications it's fine as long as you kind of prompt things correctly and you kind of learn how to deal with that and so on. It, it does work. So [00:26:00] this is what keeps coming back.
[00:26:01] Ram Sriharsha: And I don't, I don't know if I have a great answer here except to point out that that's a very common
[00:26:05] Andrew Zigler: It's a, it's a common problem that people encounter and it really kind of like highlights again, the importance of like context engineering and understanding what goes into the system because.
[00:26:13] Ram Sriharsha: Exactly,
[00:26:14] Andrew Zigler: Garbage in, garbage out. Yes. You know, and, and you wanna understand and iterate on the inputs to your AI systems versus the outputs.
[00:26:20] Andrew Zigler: Because the outputs, you're on the wrong end of the assembly line. You need to be at the beginning and understanding the source and the root and using search in your database exactly is a core component of that.
[00:26:30] Ram Sriharsha: Yes, absolutely.
[00:26:31] Andrew Zigler: And you know. This has been a really enlightening conversation about how you think about vector databases, but also the opportunities you see for engineering leaders.
[00:26:40] Andrew Zigler: And you gave a talk here today about why vector databases are this important primitive building layer for AI and why folks should be using and building with it. Where can folks go to learn more about what's coming next for you and Pinecone and, and, and get plugged in?
[00:26:53] Ram Sriharsha: Yeah, the pinecone.io. So, uh, our website has a lot of information. We have very good tutorials [00:27:00] on pretty much all parts of the stack, whether it is assistance and higher level parts of building agent workflows or deep dives into vector databases. Uh, and feel free to reach out to me if anything, any of the topics I talked about today is interesting.
[00:27:14] Andrew Zigler: Yeah, no, we'll, we'll include all of that in the show notes so that folks can go check it out and we'll put it in our newsletter as well. I'm a daily practitioner of all these tools, so I'm gonna, I'm gonna go in deeper on some of the stuff that you resource for us. And and see how I can use it. So thank you so much for sitting down with us today.
[00:27:28] Andrew Zigler: It's been really great to have you here on Dev Interrupted.
[00:27:29] Ram Sriharsha: So fun. Thank you.
[00:27:31] Andrew Zigler: Yeah. Thank you.
