Learning Library

← Back to Library

Mastering Perplexity AI Search Prompting

Key Points

  • Perplexity is an AI‑native search engine that uses retrieval‑augmented generation, pulling and embedding external web documents to craft answers with citations.
  • Its “research mode” (a genetic RAG system) performs dozens of searches, reads hundreds of sources, and makes multiple passes to deliver highly thorough results.
  • Unlike Google, which simply returns web links, and ChatGPT/Claude/Gemini, which rely on internal model weights (parametric answers), Perplexity looks outward at the live internet for every query.
  • This fundamental difference means Perplexity excels at up‑to‑date factual retrieval, while parametric models may provide outdated or inaccurate information about recent topics.
  • Effective prompting for Perplexity requires only a few precise keywords or context additions, as even minimal specific wording can dramatically boost the relevance and quality of the results.

Sections

Full Transcript

# Mastering Perplexity AI Search Prompting **Source:** [https://www.youtube.com/watch?v=05RRGiF7QC0](https://www.youtube.com/watch?v=05RRGiF7QC0) **Duration:** 00:20:23 ## Summary - Perplexity is an AI‑native search engine that uses retrieval‑augmented generation, pulling and embedding external web documents to craft answers with citations. - Its “research mode” (a genetic RAG system) performs dozens of searches, reads hundreds of sources, and makes multiple passes to deliver highly thorough results. - Unlike Google, which simply returns web links, and ChatGPT/Claude/Gemini, which rely on internal model weights (parametric answers), Perplexity looks outward at the live internet for every query. - This fundamental difference means Perplexity excels at up‑to‑date factual retrieval, while parametric models may provide outdated or inaccurate information about recent topics. - Effective prompting for Perplexity requires only a few precise keywords or context additions, as even minimal specific wording can dramatically boost the relevance and quality of the results. ## Sections - [00:00:00](https://www.youtube.com/watch?v=05RRGiF7QC0&t=0s) **Prompting AI Search with Perplexity** - An overview of how to craft effective prompts for Perplexity’s AI‑native search engine, explaining its retrieval‑augmented generation architecture and the intensive “research mode” that pulls and synthesizes multiple sources for higher‑quality answers. - [00:03:24](https://www.youtube.com/watch?v=05RRGiF7QC0&t=204s) **Avoid Few-Shot Prompting** - The speaker explains that few‑shot prompting skews Perplexity’s results and recommends explicitly using API parameters—such as source limits, date filters, and search depth—to steer searches effectively. - [00:06:54](https://www.youtube.com/watch?v=05RRGiF7QC0&t=414s) **Advanced Prompting: Constraints and Focus Mode** - The speaker explains how specifying output constraints and leveraging Perplexity’s focus modes can reduce hallucinations and shift the model’s perspective mid‑conversation without clearing the context. - [00:10:58](https://www.youtube.com/watch?v=05RRGiF7QC0&t=658s) **Recent AI Updates for Builders** - A request for a curated, well‑grounded roundup of the latest AI developments over the past two weeks that impact developers and builders, highlighting items such as Agent Kit, GPT‑5 Pro, Anthropic’s agentic coding tools, rising cloud‑code usage in Korea, security‑scan IDE partnerships, open‑source model convergence, and Perplexity’s free browser offering. - [00:14:19](https://www.youtube.com/watch?v=05RRGiF7QC0&t=859s) **Double-Check AI Search Results** - The speaker advises using a two‑tool verification loop—pairing Perplexity with another LLM like ChatGPT or Claude—to confirm citations, scrutinize quoted material, and prioritize academic databases for high‑precision queries. - [00:18:00](https://www.youtube.com/watch?v=05RRGiF7QC0&t=1080s) **LLMs vs RAG: Fact vs Fluency** - The speaker contrasts ChatGPT’s pattern‑based, confident language generation with Perplexity’s retrieval‑augmented approach that sources facts, arguing that as models grow more fluent the need for fact‑checking architectures like RAG becomes increasingly critical. ## Full Transcript
0:00How do you search with AI and make it 0:02good? That's what we're going to look at 0:03today. We're going to look at prompting 0:04for searching on the internet. We're 0:06going to look at the best tool for that, 0:08which is perplexity. I'm going to give 0:09you a guide. It's very different from 0:10traditional prompting. So, let's hop in. 0:12First, how does Perplexity work? This is 0:15often misunderstood, so I want to 0:16actually explain it clearly. Perplexity 0:19is a search engine like Google, but it's 0:22AI native. It specifically uses 0:25retrieval augmented generation as its 0:27fundamental architecture. That means it 0:29retrieves relevant documents, extracts 0:32paragraphs, and uses this information to 0:34craft answers with citations. So the 0:37pipeline looks like external documents 0:39across the internet are embedded. 0:40They're stored. Every query triggers a 0:42fresh retrieval of relevant documents. 0:45But there's an important nuance here. If 0:47you are using Perplexity's research 0:49mode, which we will see in a moment, 0:51I'll show you it. Then you have a new 0:54approach using the same architecture. 0:56And I want to explain it sort of in 0:57layman's terms. It's called a gentic 0:59rag. And what it means is research mode 1:02will perform dozens of searches, read 1:04hundreds of sources, and do multiple 1:06passes across the rag architecture to 1:09ensure it finds the best possible 1:12answer. It basically takes the effort 1:14level on perplexity and turns it up to 1:1711. That's how perplexity works. It's 1:19very different from Google, right? 1:20Because Google just finds you an answer. 1:22But what is less understood is that it's 1:24also very different from chat GPT. Chat 1:27GPT is fundamentally a parametric answer 1:31engine, which is a fancy way of saying 1:34chat GPT's default is to go and look 1:36inside its own training data and its 1:39weights in the model for an answer for 1:42your question. It does not go out and 1:44look at the internet by default. And by 1:46the way, that is why chat GPT doesn't 1:49know about new chat GPT instances. 1:51Right? If you ask Chad GPT, it will 1:54often give you the wrong answer when you 1:57ask it what the current Chad GPT model 1:59is. It's not just Chad GPT that does 2:00this. Claude does the same thing. Gemini 2:03has done this. The reason why it's not 2:05some diabolical plot. It is that they 2:07are parametric answer engines and they 2:09look inside their weights and perplexity 2:12looks outside. It looks at the internet 2:14as a whole by default. It's like imagine 2:16a world where you have an answer engine 2:19in chat GPT that looks inside the house 2:21first inside your own weights or you 2:23have a choice like perplexity that looks 2:25at the whole world first and isn't 2:27necessarily focused on reasoning first. 2:30That's the difference. And so that 2:31shapes how and where we use it. And it 2:34also profoundly shapes our prompt 2:36strategy. Let's get into the prompt 2:37strategy piece. First, you need to think 2:39of prompting with perplexity 2:42as as a little bit goes a long way. Just 2:45adding two to three words of critical 2:48context can dramatically improve the 2:51value of relevant results. I'm going to 2:52show you an example here in a moment. 2:54Basically, if you have a search like 2:56climate models, you're going to get all 2:58the semantic results from the entire 3:00internet associated with climate models 3:02in whatever order Proplexity is able to 3:05find it. If you say climate prediction 3:07models for urban planning, you're going 3:08to get a very precise pull. The thing 3:11that I want you to remember is that that 3:13doesn't mean you have to use a long 3:16prompt. In fact, on average, perplexity 3:18prompts are much shorter than chat GPT 3:21prompts. And I'll show you that as well. 3:22Principle number two, this is this is 3:24another non-obvious prompting strategy. 3:27You want to avoid what is called fhot 3:29prompting. So fshot prompting gives the 3:32model examples and I encourage it often 3:35when you are using chat GPT but don't do 3:38this when you're using perplexity and 3:40the reason why is that perplexity will 3:42overindex on those examples and dredge 3:46up only things related to those examples 3:49from your fshot prompt. So if you say me 3:52examples of French architecture like the 3:55Louvre, you're only going to get museums 3:58like the Louvre. You're not going to get 4:00anything else about French architecture 4:02because of how fot prompting works with 4:05Perplexity's architecture. Another 4:07non-obvious prompt strategy, you want to 4:09use the exact parameters for search 4:13behavior control that are embedded in 4:16the API. And I realize that that can be 4:18a lot if you're not a technical person. 4:20So, I'm just going to tell you there are 4:21a few that are pretty obvious that you 4:23can use without being a technical 4:24person. Like limit your sources and say 4:27what they are. Filter by date is 4:29something you can do in plain language. 4:30Adjusting search depth is something you 4:33can do directly in the prompt as well. 4:35The idea is don't be vague about matters 4:39that are in the API. So if you say only 4:42search recent sources, that's going to 4:44be much less helpful than using a date 4:48filter. And you can use the date filter 4:49in text. It's even stronger to do it in 4:51the API if you happen to be a developer. 4:54But regardless, in practical terms, you 4:56see a huge jump in quality when you're 4:59more specific about things that 5:00perplexity is wired to care about, like 5:03exact dates. A fourth non-obvious choice 5:06is to demand multiple perspectives on 5:09the thing you're looking for very 5:11explicitly. So instead of saying, "What 5:13are the health benefits of X?" say, 5:15"Compare findings from at least three 5:17peer-reviewed studies on X and ensure 5:20that you note conflicts in conclusions 5:23that are relevant for understanding X's 5:26effects." You see how I'm much more 5:28specific there? How I demand a degree of 5:30disagreement in the findings. This 5:32focuses the model on finding 5:34triangulation rather than just 5:36converging on a single source synthesis 5:39and just paring that. It ensures that 5:42you get a wide enough search parameter 5:44or a wide enough search scope that it's 5:46actually useful. Another non-obvious 5:49strategy, progressively deepen. This is 5:52not something that you really get to do 5:53in chat GPT or in Google the same way. 5:56Treat perplexity like a conversation 5:58where you are starting with a root 6:01question to explore and every answer 6:04opens up new questions that you can 6:05thread. So you want to intentionally if 6:08you're exploring a space start broader 6:10than you would necessarily with chat GPT 6:12and then you want to iteratively drill 6:14down with increasingly specific and 6:17actionable follow-up. So the first query 6:19kind of maps the territory and then you 6:22want to get into something that is like 6:23a promising path that is useful for you. 6:25This is a very different approach than I 6:28find ch sort of prompting chat GPT or 6:30claude where you want to bring the 6:31intent into a very structured initial 6:33prompt and really drive the entire 6:36conversation. It's not that way in 6:38perplexity. You have room to evolve 6:40because you're essentially threading the 6:43search engine through the rag 6:45architecture to find a particular area 6:47that's interesting to you as you 6:49discover the conversation together. 6:52Another non-obvious technique, specify 6:54output constraints. If you specify 6:57output constraints, you are more likely 6:58to reduce hallucinations. So, as an 7:00example, please provide evidence. For 7:03every claim you make here, please list 7:05specific section references or page 7:07numbers so I can check your work. This 7:09forces perplexity to verify claims at a 7:13granular level rather than assuming it 7:15can make broad attributions if it finds 7:17two or three different sources and just 7:18gloms onto them. Last but not least, 7:21actually we have two more two more 7:22non-obvious prompt techniques. Use focus 7:25mode really strategically. So for 7:27example, if you are an academic for 7:30peer-reviewed sources or looking for 7:32social sources, those are things that 7:34you can turn on as particular modes in 7:38perplexity. I'll show you in a moment. 7:40You want to use that in the middle of 7:42the conversation to force a reset of the 7:45model's thinking when you are trying to 7:47get it out of a rut. So if you're in the 7:49middle of a conversation, you're talking 7:51about French architecture and you feel 7:52like the model isn't taking a 7:54historian's perspective, you could go to 7:57academic mode in the middle of that 7:59conversation without resetting and it 8:00would force the model to jump and reset 8:02a bit. And that is actually very 8:04different from chat GPT because 8:05typically you would want to wipe the 8:07context window. But in this case you are 8:09just shifting the approach in the rag 8:11structure that perplexity is navigating 8:13and that's different from wiping the 8:15context window and starting over with a 8:17parametric answer engine like chat GPT. 8:20They work differently underneath. So 8:22your techniques are different. Okay. The 8:23really the last one for a non-obvious 8:25technique create spaces with custom 8:28instructions where you have repeated 8:30workflows that touch the internet. So 8:32for example, if you upload reference 8:35files on competitor intelligence, you 8:37can have a space with a standing 8:39instruction that says structure all 8:41responses as current state competitive 8:44positioning, emerging threats and 8:46strategic implications because that 8:47space is your competitive intelligence 8:49headquarters. That's an example of the 8:51kind of internet first project space 8:54that perplexity excels at. Another 8:56example of something like that. This 8:58gets into using labs, which is an a way 9:01of using perplexity to construct 9:03reports. You want to focus perplexity on 9:06internet first use cases where doing a 9:10lot of research is going to enable 9:12perplexity to come up with the kinds of 9:15information that you only get if you are 9:18leaning in to publicly available 9:21documents on the internet. And so 9:23competitive intelligence is a good 9:24example. Stocks are a good example. 9:26Equity and financial analysis, news is a 9:29great example. And the whole product of 9:31labs and and spaces, which are two 9:33separate ways to organize information. 9:35Labs is more focused on creating a nicel 9:37looking report. Spaces is more focused 9:39on giving you a standing spot for your 9:41instructions and a continual workflow. 9:44But they're both internet native and 9:45that's what you have to keep in mind and 9:47that's what diff differentiates them 9:48from Chad GPT. Let's have a look at an 9:51actual perplexity search result. Okay, 9:53this is the first example I want to show 9:55you. This was a very simple, I would 9:57call it an unhelpful search in 9:59Perplexity. Find me recent news on AI. I 10:01give it no constraints. I just tell it 10:03to go find things. It's very vague. It 10:06gives me a lot, right? It talks about 10:08major product launches. It mentions 10:10things from Sora uh to apparently an 10:12update to Chrome, which is kind of 10:14random. We can already see the quality 10:16decaying here. It mentions nine billion 10:19dollars to build energy efficient AI 10:21data centers in Oklahoma, which is 10:23perhaps not the top infrastructure news 10:25I would have picked out given anthropics 10:27deal this week with Google. Um, it gets 10:30into healthcare and science advances, 10:32which are not necessarily super related 10:35to what I was asking for, but I didn't 10:36communicate my intent. Um, and then it 10:39gets into really vague stuff that isn't 10:40date specific, like AI investing and 10:42spending. Overall, this is exactly what 10:45we would expect given the level of 10:48specificity we gave the model. Like we 10:49we were not helpful and so we kind of 10:52get what we pay for there. Now let's 10:54look at a much more specific query. 10:56Please find me a diverse set of 10:58well-grounded novel updates on AI within 11:01the last couple of weeks, i.e. since a 11:03specific date that are specifically 11:05focused on the build use case. In other 11:07words, what has happened in AI for 11:08builders in the last two weeks or so? 11:10Surprise me. Right off the bat, we get 11:12more useful answers. We get a note on 11:14agent kit, which is absolutely apppropo, 11:16but it notes that it was before October 11:1910th. It is paying attention and trying 11:21to be helpful, but it's noting that this 11:23might be on the edge. And I love that 11:24specificity. GPT5 Pro becoming available 11:27is a great one. Sor 2 API access is 11:30relevant. Enthropic's agentic coding 11:32push, so it catches clouded code on the 11:34web. It catches claude for life sciences 11:37and claude memory. Those are both 11:38relevant. It has a slightly weird one, 11:41anthropic opening a soul office. Not 11:43sure why it matters. Um, but then it 11:45makes a case, right? It says it's the 11:46number of weekly cloud code users in 11:48Korea is up. I didn't know that. That's 11:49super cool. Um, talks about Google, 11:53talks about Microsoft. This is a much 11:55more detailed response. And then it gets 11:56into stuff that I never would have found 11:58with the other search. It talks about 12:00SNIK and Windsurf in Devon um, and sort 12:03of how they're partnering together on 12:04security scanning in the Windsurf IDE. 12:07It talks about open source convergence 12:09and how uh we're starting to see near 12:12parody with Cloud Sonnet 4.5 and open 12:14source models. Uh we're talking about 12:16Perplexity's browser and how it went 12:18free, but it notes it was outside the 12:20window. Full MCP support for Chat GPT 12:22developer mode, which I knew about, but 12:24has really gotten slept on. It's a big 12:25deal. Um and then overall it gives me an 12:28assessment. I love this. There's so much 12:30to dive in because now I can say I'm 12:33really curious to learn more about AI 12:36build culture in Korea, especially 12:40around claude code. Can you please 12:44summarize a diverse set of perspectives 12:47around Korea cla code usage? And I'm 12:50going to stick with research because it 12:52will think hard. Um, and I can just tell 12:55it to go. 12:57And that's an example of how you can 12:59start to really kind of dive in and get 13:01farther. Now, one of the beautiful 13:03things about Perplexity is how flexible 13:05it is. So, while this is working, I can 13:07show you other ways to use Perplexity 13:10that are super useful. So, for example, 13:12we can choose, I know I promised to show 13:14you, we can choose to move this to 13:15academic or social or finance. We can 13:18choose to connect to other sources. So, 13:20it will search across these other 13:21sources. We can upload a file here if we 13:23want to or a Google Drive. We can speak 13:26our search. We can also get into finance 13:29and there's a whole finance product 13:30that's been built. Uh we can get into 13:33spaces. We can get into discovery for 13:35sports and culture. I think most people 13:38do not realize how effectively 13:41perplexity is owning the rich experience 13:45on the web. In the meantime, I want to 13:47talk about how we avoid hallucinations 13:49with perplexity because I get that 13:51question a lot. If it's the internet, 13:52how do we talk about avoiding 13:53hallucinations? Number one, never trust 13:56single source answers. Perplexity will 13:58site AI generated spam because it cannot 14:01tell the difference between an AI 14:02generated source and a real source. And 14:05sometimes the AI generated source is 14:06correct and sometimes it's wrong. But 14:08perplexity can't tell either way. If 14:10perplexity is only citing one source and 14:12it's an unfamiliar blog or a random 14:14LinkedIn post, treat it with skepticism. 14:16You want to be in a position where you 14:19can verify the claim with a wellsourced 14:24article for a real publication of some 14:27sort. I would also suggest if you are 14:29interested in authoritative sources, 14:32which you should be if you're using 14:33perplexity, use another LLM as a tool. 14:35So, I think I'm going to build a 14:38cross-checking hallucination prompt 14:41intended for chat GPT or Claude to go 14:44with this post because I want you to 14:46have tools to basically say, "Here's a 14:48perplexity search result. I'm not sure I 14:50believe it. Let's go to an LLM and ask 14:53the LLM to do thinking critical thinking 14:56on the post and also internet searching 14:59so that I can get a second perspective 15:01here because that two tool verification 15:03loops do work and you can use chat GPT 15:07to check perplexity's work and you can 15:08also use perplexity to check chat GPT's 15:11work. I've done that both ways. One of 15:13the things you have to be especially 15:15careful about is how perplexity 15:18attributes quotes. So, perplexity 15:21describes a quote. Please make sure you 15:23go to the cited source and search for 15:26the phrase. It is often there, but it 15:29may not be there verbatim. It may be in 15:31a different format, and it may not have 15:33the connotation in context that 15:36perplexity is suggesting in its 15:38synthesis. You have to be careful. 15:40Finally, if you have very precision 15:42critical queries, I would encourage you 15:43to select academic focus, which 15:46prioritizes peer-reviewed sources like 15:48PubMed or Semantic Scholar, and that is 15:50because that reduces the probability 15:52that you're going to get AI generated 15:54spam that's in the rag architecture that 15:56perplexity can access that sort of 15:58creeps in to the answer set. If it's 16:00focusing on academic peer-reviewed 16:02journals, it's less likely to get stuck 16:05in AI slop. The reality is hallucination 16:08is absolutely an issue with perplexity. 16:10If you ask it for verified links and you 16:12go back and check the verified links, 16:14many of them will work, but not all of 16:17them. And so there is really no 16:18substitute for that double LLM check. 16:21And finally, for you as a human owning 16:23the results. Last but not least, I want 16:25to leave you with a few thoughts on why 16:27why we use perplexity doll. Why does 16:29something like this matter in a world 16:31where we have Google and we have Chad 16:32GPT? Isn't this just the awkward in 16:35between space? The answer is no. I think 16:38perplexity is relevant because of the 16:40knowledge recency problem. LLM training 16:43data gets out of date too fast. AI 16:46knowledge is adding to our understanding 16:48of the world very quickly. Humans are 16:50writing very quickly on the internet 16:52despite the issues with hallucination 16:54and the risks with searching on the open 16:56internet. There is no substitute if you 16:59want recent information. You can 17:01actually update a rag knowledge base 17:02like perplexity has multiple times a day 17:05and perplexity has gotten much better at 17:07that in the last few months. Whereas 17:08Chad GPT treats current information as 17:11not a part of its core parametric model. 17:13That's one of the fundamental 17:15limitations of current large language 17:17models. It does not update. But 17:19perplexity it's like you can update the 17:21foundation every day. The other thing 17:22that I think really matters as we talk 17:25about hallucinations and the importance 17:27of good information in the age of AI 17:30perplexity may not be perfect but it has 17:32an accountability architecture. Rag 17:35allows you to create verifiable chains 17:38of reasoning through transparent 17:39sourcing and everything you see on 17:41perplexity is sourced and you may 17:44disagree with the source. You may have 17:46concerns about the source but you can 17:48see it. That is not always true with 17:50LLMs and that's a big deal. Finally, I 17:52want to call out that this gets a tiny 17:54bit philosophical, but stick with me. 17:56Chat GPT and perplexity have different 18:00epistemological architectures. Big 18:02words, but really what it means is LLMs 18:04will excel in cognitive intelligence 18:06like reasoning and language generation. 18:08And a rag architecture is actually 18:10focused more on fetching facts and doing 18:13so precisely. So, chat GPT will say, I 18:15believe this is true based on patterns. 18:17That is one of the roots of 18:18hallucination in LLMs. They want to be 18:21helpful. They have parametric patterns 18:22in their data and they just do that 18:24instead of searching or using tools. 18:26Perplexity says these sources claim 18:28this. I found the sources. Here are the 18:30sources. You figure it out. As LLM get 18:33better at sounding confident, we need 18:35something like perplexity more because 18:37the gap between fluency and factuality 18:40widens. Shad GPT sounds more and more 18:42fluent, but it may not be factual and we 18:44may not be able to tell. So, I think 18:46that sounds philosophical, but I think 18:47perplexity occupies a really important 18:49place culturally as AI continues to get 18:52smarter because it allows us to actually 18:55have an AI native approach to looking at 18:59facts, not just patterns. And I think 19:01that's a really big deal. Let's go back 19:03and check on our perplexity search. Here 19:07we are. I never would have found this. I 19:09did not plan this. I'm discovering 19:11Korea's clawed code culture. I get lots 19:13of facts on this and I can see at a 19:15glance that they're useful, right? 19:17Anthropic is a reputable source. I can 19:19go through, I can see Reuters. This 19:21looks like a pretty well sourced 19:22approach. I'd have to dig in, but like 19:24it looks super interesting. I'm looking 19:26at uh the interaction between Korea's 19:28work culture and claude code and how 19:30that works. This is a super fascinating 19:32example of something that you would 19:34never ever ever get to in JPT. I could 19:37not have gotten this report no matter 19:39how good my prompting was because this 19:41report depends so heavily on finding 19:45facts on the internet. And this is why 19:48perplexity is such a joy to use and why 19:50I use it so much. It's just fantastic 19:52for discovering corners of the world 19:55that you didn't expect. I hope that this 19:57has helped you to understand why 19:59perplexity matters, why we should have 20:01it. I'll capture up those nonobvious 20:04prompting techniques. I'll suggest some 20:05specific starter prompts for you. My 20:07goal here is for you to feel like the 20:09world is your oyster with perplexity and 20:11to have a sense of how important it is 20:13and how you can use it to be more 20:15effective in your search. It is not at 20:16all the same as chat GPT search and I 20:18hope that you can see that. Best of luck 20:20with uh search in the