[
  {
    "startMs": 12680,
    "text": "Everyone, I'm Jess. I'm from Braintrust. I'm a DevRel engineer at Braintrust, and in this talk today, I'm gonna be teaching you guys what evals are, why they're important, and then I'm going to be talking through a relatively complex real-life eval that I did to try to solidify some of the concepts I'm about to teach you. Uh, before we get started, has-- Can you just raise your hand if you've heard of evals before or, like, would confidently say that you know what they are? Okay, cool. Pretty much everyone here. Makes sense."
  },
  {
    "startMs": 42800,
    "text": "Um, cool. So I'm just gonna start off with the basis of what evals are, why they're important. Um, I've been on, um, calls with customers where they've said that they've shipped features because either the engineering team told them that it was ready, or maybe a PM said that they tried out a couple of prompts, and they said it looked good. Uh, and this is what you don't wanna hear 'cause essentially, these teams are making ship decisions based off of vibes, uh, which is not good. Uh, and essentially, what you wanna hear is more something"
  },
  {
    "startMs": 72840,
    "text": "like, \"Oh, I ran, um, two hundred different test, test cases, and ninety-four percent of them passed, and that's why we're shipping this feature.\" Or be able to say nuanced things like, um, you know, \"I shipped this feature that increased accuracy by five percent but actually decreased tone by five percent.\" So when people ask me what evals are, I think the best way to answer it is that it's just a way to answer questions about your AI system. So for example, uh, which LLM is the best choice for our needs? Especially"
  },
  {
    "startMs": 102420,
    "text": "when there's new models dropping all the time, it's good to be able to have data to back up your decisions. Uh, how does AI perform across diverse real-world scenarios? So for example, your system could do really well with English outputs but not Japanese-- or English inputs but not Japanese, or types-- uh, do really well with Python but not TypeScript. Um, a big one is cost efficiency, so how can we achieve high performance without excessive costs? This is definitely becoming more topical now. Uh, brand consistency,"
  },
  {
    "startMs": 132400,
    "text": "so does AI reliably reflect our company's voice and standards? Uh, feedback, are we learning from users and improving iteratively? And then one of the biggest ones is how do we know when something breaks or gets worse? So a real-world scenario that I like to use a lot is, uh, about a year ago, so April of twenty twenty-five, um, OpenAI actually shipped a model change and model update, um, that essentially made their model more-- was supposed to make it more helpful but actually made it sycophantic and"
  },
  {
    "startMs": 162360,
    "text": "therefore, uh, too agreeable and therefore less truthful. So this is a really good example of where evals can help you catch that, um, especially because, uh, AI systems are so complex and undeterministic that it's really important to have these eval systems in place at your company or whatever you're trying to build. So we're gonna get down into more of the specifics of what evals are comprised of. So evals comprise of four main things. Uh, the first thing is a dataset."
  },
  {
    "startMs": 192460,
    "text": "So this is the data that you're passing into your AI system, and it usually includes things like, uh, your golden standard, uh, cases, your edge cases, your failure modes. Um, then once you have your dataset, you want to write a task. So a task defines how you want your AI system to behave when it takes in the input and produces an output. So concretely, this usually includes a system prompt and then the model that you're picking to execute this task."
  },
  {
    "startMs": 223080,
    "text": "Then the next thing you need to do is you n-need to write a scoring system that will tell you whether the AI's output is good or bad. Um, and so usually it's, uh, maybe the PM or engineer that is writing the criteria that comprises that goodness and badness. Uh, and you might have heard of terms like, uh, deterministic scoring, LLM as a judge, human-in-the-loop. Those are different types of scoring, um, frameworks. Uh, but yeah, that's-- So you write your scoring system. And then the last thing you do, which is the most fun part of writing evals, is"
  },
  {
    "startMs": 253060,
    "text": "you experiment. So every configuration of a specific dataset, a specific task, and a specific scoring system equals to one experiment. And as you are testing out different hypothesis for how to make your system better, you're tweaking either your dataset, your task, or your scoring system, and these all lead to different experiments. So once you have your different experiments, then you start to compare them. So you might run experiment one and experiment two, and you can see whether there's"
  },
  {
    "startMs": 283060,
    "text": "been improvements or regressions. And you can also, uh, click into specific rows to see exactly which test cases have changed. Uh, this screenshot is from the Braintrust UI. Another thing that you can do is you can use, uh, natural language to query across your traces and experiments. Um, so for example, you can ask it things like, \"Summarize my experiments for me. Um, highlight the problems in these experiments.\" Um, I find this especially useful when you're running experiment-- or you're running, like, um,"
  },
  {
    "startMs": 313760,
    "text": "tra-- you see traces with, like, long multi-turn conversations. It's good for, um, helping you catch things like hallucination or drift that's kind of hard to catch manually. Um, you can also-- If you're running different experiments, you can ask it, like, which performance or which experiments performed the best of, out of all of these experiments. Um, so this is another way you can, like, find patterns in your traces. Um, we also-- This is using, like, the UI in Braintrust, but we also have a CLI and an MCP"
  },
  {
    "startMs": 343580,
    "text": "server if you prefer to do it that way as well. Um, and then the last pillar of Braintrust is just, like, regular observability. So you can basically use the Braintrust SDK to wrap your code, um, and then, uh, have your production logs show up in Braintrust and essentially do traditional observability, where you can monitor your production logs and create dashboards and things like that. Um, so kind of to tie this together, I'm gonna walk you through kind of a flow of how you might use observability and"
  },
  {
    "startMs": 373648,
    "text": "evals to improve your AI system. So we're gonna start on the top left corner where it says app. So imagine you have some sort of AI application. Uh, in this scenario, I'll say you have a, a new docs page that you just shipped with like an AI chatbot built in so that users can ask questions to your AI chat-chatbot about your documentation. So let's say you shipped that yesterday, and then you have customers start to use your AI chatbot. So when they start to hit your AI chatbot, you're gonna start to"
  },
  {
    "startMs": 403548,
    "text": "get real production logs coming into your system. Once you get those production logs, you can, uh, sample a subset of them, maybe like ten to twenty rows or something, and create a kind of dataset. Uh, once you get that dataset, then we can run evals on that, right? So you can... You have your dataset, you can write your task, you can write your scoring system, and you can run, uh, evals on them. And you can start to tweak your dataset, you can start to tweak your prompt or s- tweak your score and run multiple different evals on that. Once you have your"
  },
  {
    "startMs": 433028,
    "text": "evals, like run a couple different evals, um, and start to see which eval does the best, you might learn something about your system, and then you're gonna go ahead and make your code change or make a change to whatever AI system that you're using, uh, and then deploy that back into your application. And then the whole flow repeats again because, uh, those changes will show up in your production logs and the whole system happens again. So that's kind of how the flow works as a whole. Um, the last slide I'll show before I get into a"
  },
  {
    "startMs": 462948,
    "text": "concrete example is... I love this slide because, uh, conceptually, I, I wanna hammer into people that evals are a team sport. Uh, there's so much that goes into creating evals, uh, from getting real world data into your platform, labeling it, developing hypothesis, things like that. So the AI engineer kind of works on obviously making the code changes, but also getting the data into the platform. Uh, the product manager is very important. I think they're, they usually spearhead evals at a company, but they're the ones that develop the"
  },
  {
    "startMs": 492948,
    "text": "hypothesis of what success looks like, um, and help, uh, tweak, tweak the prompts and things like that. Uh, subject matter experts are really important, especially in like more niche, uh, areas like, um, insurance, medicine, law, and things like that. Um, they can help, uh, label the data or provide a ground truth for what, what excellence looks like. Uh, and then data analysts obviously to analyze the data. So I always think that this slide is, is really useful to look at. Okay. So let's"
  },
  {
    "startMs": 522868,
    "text": "go into a like interesting real world example of how you might use an eval. Um, in this one, I'm comparing agentic search to vector search. And the story behind this specific eval is, um, is that my CEO, Ankur, sometimes goes on Twitter and sends me articles of things he wants me to eval. Uh, so he sent me this, uh, tweet from Cursor, where Cursor talks about how they used semantic search, also known as agentic search, um, to significantly improve their coding"
  },
  {
    "startMs": 552748,
    "text": "agent ex-- uh, performance. And he was like, \"Jess, can you eval this for me?\" And I was like, \"Okay. I'm busy, but okay.\" Um,  and so I kinda went on Twitter and I saw that there was a lot of, uh, discourse or discussion around this topic, so I thought it might be interesting to eval. So, uh, we're gonna eval a vector search and agentic search, so I just wanna give a quick, uh, summary of what each one is just so everyone's on the same page. So vector search essentially is the process of taking, uh, data like, uh, code or"
  },
  {
    "startMs": 582108,
    "text": "text and turning them into embeddings, where embeddings basically represent the semantic meaning of that chunk of text or code. So, uh, if you look at the graph here, it's a very good visual example where, um, the embeddings, uh, wolf, dog, and cat are very close to each other 'cause semantically they all are animals. And then, uh, the embeddings banana and apple are also close to each other 'cause they're fruit, but then, uh, they're far apart from each other because fruit and animals are different things. Um, and so what happens"
  },
  {
    "startMs": 612148,
    "text": "is in, you know, um, in a real vector search implementation, you would take these embeddings, uh, and store them in a vector database, so something like Qdrant or Pinecone or something like that. And if you come up with a query like, uh, for example, where does the database logic in my code base live? Uh, what would happen is the vector database would return back the chunks of, um, embeddings that most closely semantically match the words in your query. So the, the chunks that are closest to things"
  },
  {
    "startMs": 642148,
    "text": "like database or logic. So that's how vector search works. Agentic search is slightly different. Agentic search works more like a human would. So essentially with agentic search, uh, you give an LLM tools like, uh, grep, find, ls, cat, um, bash commands, and it basically explores, explores your code base like a human would. So it would, uh, look up a function name, open a file, read the file, follow that function call to another file, and then open that file and read that file, and then continue"
  },
  {
    "startMs": 672108,
    "text": "until it's found what it's ne- what it needs to. So this is, uh, much more similar to how a human would kind of go through a code base and, and look up-- look for certain pieces of logic. Okay. So then how did we eval this? So the first part of creating eval is creating your dataset, right? So to do this, we basically used, uh, Microsoft's TypeScript Go repo, uh, which is open source. And what we did is we basically found all the PRs, um, the merged PRs with the title fix in"
  },
  {
    "startMs": 702148,
    "text": "them. And, uh, we basically checked out the parent commit. So essentially we checked out the commit to the point where it still has the bug in it. And then we used Claude to generate a, like a bug task or like a linear task based on the diff, so based on the diff between the buggy code and the fixed code. And then we fed that into an LLM, uh, and had that LLM either use RAG search or agentic search to find the buggy code, to locate where the logic, the buggy logic lives."
  },
  {
    "startMs": 732488,
    "text": "So that's how we created our dataset. We had like maybe We, we found like 20 different PRs, um, and, and ran the eval on that. Uh, in terms of our task, uh, we actually had to implement vector search and agentic search obviously. Uh, implementing agentic search was really easy because Claude Code uses agentic search by default, so we just set Claude Code on these tasks. Um, implementing vector search was slightly more complicated because we still wanted to use the Claude Code"
  },
  {
    "startMs": 762428,
    "text": "harness, um, to execute the vector search just to keep the experiment consistent, but we had to basically prevent, uh, Claude Code from defaulting to agentic search. So we did two things. Uh, the first thing is that we explicitly said in the prompt instructions not to use agentic search. Um, but if you've ever done this before, you know that prompting is never enough. So what we also had to do is use the, uh, disallowed tools flag to prevent, uh, Claude Code from calling certain things, uh, that it would normally use in agentic search."
  },
  {
    "startMs": 792488,
    "text": "So that's how we implemented, uh, vector search and agentic search, both using the Claude Code harness. Um, another kind of interesting thing that we ran into is, um, when we were trace, uh, running the traces, we ran the Claude Code, uh, runs as subprocesses, but it essentially mean, meant that the traces were orphaned. Um, and, uh, our solution was basically passing the parent span IDs as environment variables. So if that doesn't make sense, that's fine because I'll show"
  },
  {
    "startMs": 822528,
    "text": "you via pictures. So this picture that you see here is what the trace looks like when it's running as a orphaned subprocess. So you can see in the highlighted on like the left panel where it says, \"Run Claude Agent,\" um, that's basically in the trace saying, \"Okay, it's running the Claude subprocess,\" but you can't... You don't actually get any visibility into what happens when it's running that Claude process. With the fix that we did, uh, this is what you see now in the trace. So you can see that it's, uh,"
  },
  {
    "startMs": 852668,
    "text": "calling the Claude Code subprocess, but you can also see everything that's being run in that subprocess. So you can see all the turns, all the LLM calls, uh, every single terminal command that's being called, like grep, bash, ls, find, things like that. Uh, and it gives you visibility into what's happening in this actual run. Uh, and implementing this fix was really important because being able to have transparency into what's happening, um, is, is important for debugging, uh, later on. And scoring was really simple. Uh,"
  },
  {
    "startMs": 882588,
    "text": "for our scoring, we basically said, uh, because we're using Microsoft- Microsoft's TypeScript Go repo, we basically said if the test, uh, passed the, uh, t- uh, Microsoft TypeScript Go test suite, then it gets 100%, and if it fails, it gets a 0%. So it was very binary scoring here. Uh, okay, so in terms of our results, uh, the summary is basically vector search and agen- agentic search both got the same level of accuracy, but essentially vector search cost, uh, four"
  },
  {
    "startMs": 912348,
    "text": "times more than agentic search. So when we looked at the actual traces and, uh, kind of manually like went through to try to understand what happened, we had a couple of learnings. So the first thing is that vector search, uh, returns back chunks of code, but these chunks of code often miss things like, um, imports or tool calls or the calling code above it. Um, and so as a result, it's not enough. It doesn't give enough context for how the agent should actually solve the bug. Uh, so one example is in"
  },
  {
    "startMs": 942348,
    "text": "one of the rows, we saw that the vector agent made 26 searches and still couldn't piece together how three functions interacted across the file. In contrast, agentic search was actually able to like, as I said, work more like a human would. So grep the function name, read the full function, uh, spot the flaw in the logic, and then follow the calling, um, calling logic, uh, into another file. Um, and the way that I kind of like to think about it when I was just going through the traces is it felt like,"
  },
  {
    "startMs": 972748,
    "text": "um, vector search gave a lot of proximity to the correct code, but agentic search had more of the connective tissue to connect the actual logic of those chunks of code. Uh, and then tying that into the third learning I had is, um, the reason why vector search was so much more expensive than agentic search is because it made more, um, searches. 'Cause it was o- um, c- constantly searching, like returning a chunk of code and then searching again, returning a chunk of code, searching again. And as a result, um, all"
  },
  {
    "startMs": 1002708,
    "text": "of those LLM calls, uh, piled up and made it a lot more expensive, uh, to execute. Um, and this is the last slide, but essentially, uh, you might be sitting here thinking, \"Oh, there's so many holes in this eval,\" which is very fair. Like, there's lots of things I could've done. Uh, I, uh, you know, there's lots of best practices when it comes to running evals, like running multiple trials per task. Uh, especially because LLMs are so non-deterministic, it's good to run multiple trials to be able to trust the score that you're getting back a lot"
  },
  {
    "startMs": 1032368,
    "text": "better. Um, my vector search implementation was relatively weak. Uh, I could've done a better implementation. Um, also expanding the data set. Um, I think we only ran this on like 20 different rows, so it would've been better to expand this, uh, to even different, um, different repos or things like that. Um, but the, the point of this talk is not to create the best, uh, off- like the best eval ever, but it's more to illustrate to you the concept of h- the kinda how to create a eval end to end."
  },
  {
    "startMs": 1062508,
    "text": "So hopefully I've done that, um, and that is the end of my talk. Um, so hopefully you learned a little bit of what evals are, why they're important, and how you might do one in real life. Um, and yeah, I think that's it. So thank you for listening."
  }
]