Your Coding Agent Is 6 Months Out of Date — Jakub Hojsan, Exa

AI Engineer · 12 min · 136 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

1. 00:12Hello everyone.
2. 00:13Can you all hear me?
3. 00:15Good.
4. 00:16Sweet.
5. 00:17My name is Jacob Hoisan.
6. 00:18I'm a Ford deployed engineer here at Exa.
7. 00:20And uh today's presentation is on how we built search for coding and review agents and how we power most of the Bay Area.
8. 00:27Um, in this regard, um, coding agents don't really need these 10 blue links that you see when you search Google.
9. 00:33Uh, what we do is we have a semantic search engine that essentially you pass in a query and we provide you context rich highlights to pass into your LLM.
10. 00:44The core problem that we have with large language models in code review and coding agent applications is that there's a knowledge cut off to all of these models.
11. 00:53As you can see from this graph, we're typically lagging around 6 months from the knowledge cutoff date to the release date.
12. 00:59In these instances, you have this gap that exists between when the model was released and when you might have an important change log or a PR push to a repository that you do have to review.
13. 01:10Um, in this instance, I do have an example where from the uh cooter vector store PR um that was made a few months back.
14. 01:20This is within our model blind spot after GPD 5.5's cutoff date.
15. 01:25So in these months leading after the cutoff date, you have all of these changes being made to real repositories that one you can't create net new repos with, but you actually can't review at all.
16. 01:37Um, so that's kind of where Exa comes into play in this instance.
17. 01:41Uh, and we'll actually just review this diff real quick so you can kind of see what I'm getting at.
18. 01:46But before we do that, like how would a human even review this?
19. 01:50Very typically a a human would probably go on to Stack Overflow and say, "Was this like random variable removed?
20. 01:56In this case, it's called inertia check."
21. 01:58Um, and like why isn't my code compiling?
22. 02:00And uh if your code's not compiling, you'll probably just paste the error code into Stack Overflow and you'd actually get the response that you would want and then you would make a PR to the repo in order to resolve it.
23. 02:11Now in this case in with using a large language model many people without web search will go see okay here's the error the diff looks consistent and it looks like a cleanup but that's really not the case.
24. 02:24The reason why we're removing this parameter uh in this instance is because we actually need to refactor to a dependency bump.
25. 02:32So using web search, you're actually able to take a look at the repository itself, any breaking changes from this repos change log and then explain the migration in depth.
26. 02:48So what's interesting here is that we don't return the entire GitHub repository page.
27. 02:53What we'll return is a very small snippet from the page explaining the change.
28. 02:57So if the query is exactly this, we have an interpretation step which will take that entire page and distill it down into exactly what the large language model needs to answer the question.
29. 03:09So you're not feeding the model 100,000 characters anymore.
30. 03:12You're feeding it quite literally only 500 characters to answer the question.
31. 03:20Now adding a search tool to your agent is not enough.
32. 03:23There's obviously kind of a two-step process here.
33. 03:26Um, if you just bolt on a web search onto your agent, many of you have seen maybe when you're using cloud code, you'll say, "Hey, I'd like to use sonnet 46 or 47 now."
34. 03:35And the model will say it doesn't exist, right?
35. 03:39So even though it has access to a web search tool, it's not actually like doing anything.
36. 03:43Um, because natively cloud code does have a web search tool.
37. 03:46Um, and in these instances, the model simply doesn't know that you're requesting a version that does exist.
38. 03:52So you actually have to provide the model with a set of rules.
39. 03:55So when you're doing a code review agent, the rule might look like when you have a diff that bumps a dependency or version, you might want to look at the upstream source to verify that this is actually true and then ground the actual review in what you find.
40. 04:10Um, so when I work with these code review and coding agent customers, there's several instances where this might happen.
41. 04:16They might actually add the tool to their agent and then adding it to the harness is simply not enough.
42. 04:21So, it's really a two-step process.
43. 04:22It's instructing your agent when to use web search uh since oftent times it's not built in and then also actually just executing the search uh in the sentence.
44. 04:33So, there's four things that AXA kind of gives us a model.
45. 04:36Uh we have transparency.
46. 04:38When you're using a native web search tool within OpenAI or Anthropic, you actually have this black box that exists.
47. 04:45They they call a web search API.
48. 04:47They synthesize all the information.
49. 04:48It might take up to 10 seconds and then you get some of the sources and then none of the content.
50. 04:53So that's really point one.
51. 04:55When you use Exo, you get the entire span.
52. 04:57You get the entire trace of what we looked for, the exact queries, the exact sources, what highlights were passed into the model.
53. 05:03Um, this is all given to you and can be passed into any of your telemetry to go investigate when something does go wrong.
54. 05:10Um, we also provide token efficient highlights, which I'll cover on the next slide, which distills an entire page down into exactly what the model needs.
55. 05:18And then it also comes down to cost.
56. 05:21Many many of the times that cost from these model providers are actually so large that you don't notice your web search bill.
57. 05:27Um but given like sufficiently large workloads were priced actually much more effectively than native web search providers and much better quality in most instances.
58. 05:36Um and then another thing that most of our customers are very happy with is that not being tied to one model provider.
59. 05:42uh when you use a thirdparty web search tool, you have this level of standardization that exists.
60. 05:47Uh you can use GLM when it comes out.
61. 05:50You can use the new anthropic models when it comes out.
62. 05:51You can use the new OpenAI models when it comes out.
63. 05:54And then you have like one fixed API that you're calling instead of everything that the other model providers are using.
64. 06:00And you have a fixed set of parameters that are extremely flexible um to what you would need across all model applications.
65. 06:08So, here's a quick gif of kind of how this works since I'm more of a visual learner myself.
66. 06:12Um, but in this instance, if you're asking about the access search API, we're quite literally just pulling certain lines at runtime, not using an LLM.
67. 06:21This is completely computational.
68. 06:23So, we actually cut zero, we add zero extra latency to our search call by providing you the contents of pretty much any page on the internet.
69. 06:32Um, and I do have an example of how this works.
70. 06:37So, let's say you're looking up.
71. 06:42Hopefully, you can see that all right on the query there.
72. 06:44If you're looking up and then a biography, you'll quite quickly get a result that I love photography and then what I do in my free time.
73. 06:54So, I like driving motorcycles.
74. 06:56As you can see, it's right here.
75. 06:57But then from that same website, I could actually ask for my phone number.
76. 07:02And then at runtime, we'll actually just get a completely different set of information.
77. 07:07So you can see this is generally applicable to pretty much any use case.
78. 07:10In a coding use case, you might have an entire GitHub repo that has like a bajillion lines of code, but you don't want the bajillion lines of code to be passed into your model.
79. 07:18And you don't want to use a model to synthesize that information either.
80. 07:21Um, so this works across any type of query.
81. 07:24If you want the latest news about Nvidia and it has to mention Jensen, then you could also get that.
82. 07:29Um but yeah, this is like one of the many applications that we have um in these instances and coding is really just the beginning.
83. 07:40Um this is one of the reasons why I joined Exa in the first place.
84. 07:44We have we had such a large swath of people that wanted to use Exa as a search API to get coding docs.
85. 07:52Um and I was super excited for the application.
86. 07:54Um but as we went forward and we took on many popular providers and they loved using us and we provided them a great experience.
87. 08:01Cursor, Cognition, Warp, Code Rabbit, all of these companies now use Exa to power their web search.
88. 08:06Um but now we have a new offering.
89. 08:09Uh it's called Exa Agent.
90. 08:11Now there's many instances where you don't want to orchestrate your own search, but you need a good search experience.
91. 08:16Uh and that's kind of where it comes in.
92. 08:18We partnered with several data providers to not just surface from the web but then also surface highlights from many different places such as similar web for web analytics, particle for podcast intelligence and crunchbas for private markets.
93. 08:31So we're powering the biggest financial firms and hedge funds uh with this information.
94. 08:36Um and if you want to learn more about this you can go to exit connect.
95. 08:39And then I do have a small little message here as well.
96. 08:45I ran a quick query.
97. 08:46If I wanted to find all of the people that are attending the AI engineer conference, you can quite literally find anyone that has ever mentioned this conference at this location.
98. 08:56And it's kind of crazy because I see people that I know and have emailed me here.
99. 09:00So, uh, these are most of the speakers that are attending the conference.
100. 09:04Um, and this applies broadly across every single use case.
101. 09:07If you wanted to find all the employees that work at Exa, you can also do that.
102. 09:11And you could also get all their emails and LinkedIns for outbound.
103. 09:15But yeah, um I think this is quite exciting and we wanted to package our search in the best way possible um because we can orchestrate our search quite well.
104. 09:27But yeah, we're you if you want to use Access today, you could simply call us via our MCP and then we're on basically every single provider that supports it.
105. 09:35Um and if you'd like any help in setting it up, that's my job.
106. 09:39So you can feel free to send me an email.
107. 09:41Um, but yeah, there's also a QR code at the top.
108. 09:44Um, if you want free credits, you could also talk to me after the talk.
109. 09:47Um, but yeah, if there's any questions, happy to take them.
110. 09:51I think I could take one question or two questions.
111. 09:58Sorry, I'll come down real quick.
112. 10:20I'll get it.
113. 10:23Hello.
114. 10:24Hello.
115. 10:24Uh he asked how we manage the reranking and organization of like billions of documents on the internet basically.
116. 10:31Um and it's a multi-stage process.
117. 10:33Um so essentially we pass in your query, we turn it into a query embedding and then once we've turned it into a query embedding.
118. 10:39There's several steps in between for keyword filtering uh alongside semantic search.
119. 10:43But um there's uh many different methods of doing this.
120. 10:47Uh but we do employ reranker ranking steps during our search uh in order to give you the best information.
121. 10:52Um we drop results that are not relevant.
122. 10:54Our index is highly curated.
123. 10:56So, we don't have as many documents as Google, but I bet you that we have very high quality documents in our like tens of billions of documents index.
124. 11:05Uh, but yeah, if you're interested about any specifics, feel free to talk to me after.
125. 11:09But yeah, cool.
126. 11:20Yeah.
127. 11:25Yeah.
128. 11:27Yeah.
129. 11:27So, we actually have this feature that's quite useful.
130. 11:30Um, so you could actually generate a schema.
131. 11:33So, you up to 10 fields on deep and then I I believe it's up to 100 on agents.
132. 11:37Uh, but you basically type whatever you want in natural language and you can generate it and then in this instance it'll generate and then you'll just click confirm and there you go.
133. 11:50So, and then we adhere to it and then if you want additional properties, you just set this to true and then you can get whatever you'd like.
134. 11:56You could also, my favorite demo is hobbies.
135. 11:59But yeah, um, great.
136. 12:01I'll take some questions off the stage, but uh, thank you all for listening today.