Local Deep Research - Local AI Agent Full Privacy with This 1-Click Setup

Ajouté : 2026-05-09

[00:00:00]Let's talk about AI agent and productivity.

[00:00:03]So today, we're going to check out something really cool. These local AI tools are able to help you do deep research, analyze data, and build your own knowledge locally on your machine.

[00:00:15]This is called local deep research. Now, this GitHub repository, I was checking it out a few days ago. Since I've been working on AI agents pipelines and building apps related to these topics, the interesting thing I'm seeing is that it has similarities to what I was doing.

[00:00:31]This GitHub repo has a similar approach, but it's focused entirely on local environments, which is convenient for a lot of individuals who just need their own tools having their private knowledge base built with specific domain expertise, rather than having an enterprise-grade app with hybrid cloud and local combining team collaborative work for research. So this one is suitable for a lot of audience in here where you can build it by yourself. Now, I saw an example from another great YouTube channel, Fahad. He's all focusing on the AI stuff, not just image or video types. He did an example installation in Linux. And yes, it's quite easy with this installation as well, even in Linux. And what I'm going to try this time is an even easier way, the pip install. Just one command line.

[00:01:19]This is compatible with Windows, macOS, and Linux operating systems. And the setup guide showing in here for Windows, one thing is that if you need to export the research results to PDF, you will need to set up another component called a Pango. This allows Windows to create PDF files in the back end of this system. So how does this research work?

[00:01:40]First of all, this is built based on the LangGraph agent. And LangGraph, as I've shown in previous videos as well, in the video where I talked about the LangGraph AI agents with ComfyUI and also the MCP, that is using the Nginx workflow. It has the similarity of using the LangGraph or LangChain framework.

[00:02:00]Especially the Muse Studio that I built previously, which had some features and some simple workflow that is based on using the LangGraph framework to build the agentic types of features. So, building the knowledge base using deep research, or in general, a lot of deep research features nowadays with AI agents have the similarity of you doing research, downloading those sources, or just temporarily grabbing those source context, indexing that, or embedding that into your RAG system.

[00:02:26]Those are very typical for enterprise grade as well.

[00:02:29]Now, as we can see here, it uses the local SQLite DB to store those data, as well as having indexing to do the semantic, building your own semantic data basically, and synthesize every single downloaded source that has those keywords, etc. are going to enable you to build your knowledge base. And eventually, in the future, when you use this again, this local deep research agent is going to help you get smarter and smarter in building your reports.

[00:02:55]So, I'm going to try this out. First of all, you'll need two components to get started. The first one is SearXNG.

[00:03:02]This is a local search aggregator. I would say it's not a search engine, but it's needed for aggregating different search engine results, and helping you get all those search data for your own local usage.

[00:03:13]And then I've built a simple plan in here that is easier to follow. We use one command, pip install local-deep-research.

[00:03:21]This is all you need to grab this local deep research repo on your OS system.

[00:03:27]The SearXNG, as I just mentioned, is going to run within your Docker container. So, you'll need Docker Desktop or your Docker client app on other OS to run this search aggregator.

[00:03:38]The third thing you'll need is Ollama.

[00:03:40]This is an easier way to set up your own local language models, providing the runtime for you.

[00:03:46]Then the fourth step is to get started with the web UI, and then we can go to our URL to process all the deep research tasks.

[00:03:54]So, through all these steps, I've built an even simpler way. For Windows users, you can use a BAT file. Just one click to install. And here, I'm showing that I've set up this BAT file. So, go to the command prompt window, and you can start with this. So, basically, copy and paste that link and press enter, and it will start going through the process, checking you have Python, pip, Docker, Ollama, is it available, etc. to continue. So, press any key to continue.

[00:04:22]It will go to step two.

[00:04:24]Now, basically, it will download that local deep research repo for you, and it's automatically run.

[00:04:30]It will take some time because there's a lot of files for the Python library dependencies of this deep research repo.

[00:04:37]Okay, so after the first time installation, it also goes through the steps showing local deep research installation successful. And now, it starts with the SearXNG search engine.

[00:04:46]If you have already downloaded Docker Desktop, it will help you set it up for the first time. Otherwise, if you already have it in the Docker container, it will detect the container and help you link it up with your runtime.

[00:04:57]Ollama models. You can choose whichever one you want. By default, these are just some small models for easy getting started, but I chose my own models in this one since I already downloaded the newer Qwen 3.6 model.

[00:05:10]So, after you choose the Ollama model type, it will go to starting local deep research. It will start these three links here. So, basically, you have your SearXNG search engine. You don't have to access that, but if you access this link, you'll see something like this one showing. So, it's just like a typical search engine. For example, a cat.

[00:05:28]And all of this is aggregated from different search engine sources, putting it together into your own local aggregator, which is this local app in your Docker. You can see all the suggestions with tags, etc. The response time here shows you where the sources are getting from in this search inquiry.

[00:05:44]So, So, you don't need to use that in your AI runtime workflow. It runs in the back end of the system. But, if you want to use it like a traditional search engine, you can use it as well. Just type in this interface. So, as you can see, the link is localhost:8080 port number. This is the settings of what it is for this app.

[00:06:05]And the second one is Ollama. We have that installed already. And the third one, of course, is local deep research.

[00:06:11]Now, you can click this link, localhost with port number 5000.

[00:06:15]The first time you load this page, it has local deep research with a login here. Now, all these login things are going to save in your local SQLite database, which is just a small piece of files storing the SQL data. You won't have any demo account in here. So, it's better just creating your new account for whatever purpose you want. So, let's say I'm going to do a test one, and also create a password.

[00:06:39]And then we can create an account locally. So, after you create an account here, that you'll see lots of features.

[00:06:45]The warning here doesn't matter. If you first try it out, you can do those settings later.

[00:06:51]But, in here, we can search our own models, such as, let's say, I use the Qwen 3.6 27B BF16 model using Ollama.

[00:07:00]So, now it's connected to this model in Ollama.

[00:07:03]The strategy in here, one more thing worth mentioning. As you can see, the LangGraph agents, this is using LangGraph multi-step nodes to create the processing steps for the agents. So, what it means is that if you take a look at some multi-step agents, they have a plan, prepare the source from research, aggregate those sources, maybe have a human loop into the steps to approve or not. Then, the next thing is actually creating the report for the research.

[00:07:29]Those are typical ways of how you use AI agents.

[00:07:33]So, let's go back in here using the autonomous agentic research. This is the most hands-off way to run deep research.

[00:07:39]The quick start, we can just click here.

[00:07:41]Let's say yes, we forgot to put our topic. So, let's put our research topic in here.

[00:07:47]And then, here's the fun part. I asked my Open Claw, my AI assistant, what kind of deep research topic it wants me to do. And I said, let's switch the roles.

[00:07:57]This time, I become its AI agent. So, Open Claw gives me this research topic, local AI agent frameworks, what's available beyond LangChain. Yeah, the concept is kind of crazy, right? My AI asked me to do research, and now I'm running that research through another AI. It's like AI seption. But, I try to mimic that as I am the AI agent running this local deep research. So, let's try it.

[00:08:22]And without any quick option settings, I'm going to just start the research here and see what we get in the progress. So, in the meantime, as you can see, like a progress animation popping up here showing you the status of this research. And right now, it's going on. Wait until some moment, it will come back with some research result.

[00:08:40]So, in the meantime, I'm also going to record the progress of how it's going to pop up. Each call getting all this data with different search paths. As you can see, there's some going out on the web doing web search, having the research subtopics. So, in the meanwhile of research, as you can see, there's a progress where you can keep track of the agent's reasoning.

[00:09:00]It has the web search, so it's using SearXNG to gather some links and building the agent with some basic information first, and then now it's using another research subtopic action, which is a tool calling to create another content of the subtopic of what I just wrote.

[00:09:16]Down here, we have all the research logs that you can keep track of. What kind of extracted data it has.

[00:09:23]As well as going through each extracted search result, seeing what kind of content is relevant for our research topics.

[00:09:29]And then lastly, it will create a final research report for you. This is basically the basic workflow of how the AI agents work. And honestly, watching this unfold in real time, seeing the agent think, search, verify, and synthesize, it's kind of mesmerizing.

[00:09:47]Like, this is your PC doing PhD-level research. And we'll wait for a little while to go through, and it does seem to be working. So, yeah, check this out if you're interested in AI agents deep research.

[00:10:00]And deep research, this is not only for scientific topics or academic. You can research for something related to your work. Whether you're creating a video for your content, you can do a research before you create a video about whatever topic you're doing. So, this is a really cool AI agents workflow.

[00:10:18]My experience with this is not really a complete AI tool that you can use in your enterprise or company, but this is a good open source that demonstrates it could use the agentic workflow with some open-source framework as well as local models to create such things.

[00:10:33]So, the generate result here, you'll see the progress bar going to 100%. Status complete. The research task has been ended.

[00:10:41]Which has the view result button here. A very typical UI flow for this kind of deep research result. For the progress, this is the important thing you have to look at, not just the output of the content. Because you have to check this progress here. You'll see some research IDs of my queries with the summaries that have relatively content, not just generic content generated from the language models. For instance, if you scroll up to the status here, you'll see that it's crawling some external content for reference.

[00:11:12]It will be verified if this is relevant content as well with the language models, and then analyze the content into each milestone. So, when you look at the milestones here, it's going to make like a to-do list, step by step going through each of the research steps. So, each research step in here, it will show you the progress. Finish the research.

[00:11:32]Now, you see in the last few steps, it's generating the content through saving research information, as well as putting that into the database. So, all of that has been generated, and you can click view result. And in the view result here, it will show the inquiries of my research topic. It has very in-depth crawled content that's relevant to what I was searching and looking for, and the AI agent also puts all the source references. It feels like academic research where you put all those references as well in the appendix. The same concept goes here. You got 1,393 searches, which picked up 20 relevant results and made it into really precise content for your research. That's how we got our summarized content for this research. On the top here, where you see all the language references, some of the similar content in a few reference links, and that's how it got merged into one reference in the sentence like this.

[00:12:25]So, at the end, you can download as PDF if you install the Pandoc extension, or you can export it as a markdown file.

[00:12:33]And this is able to read even by other AI agents, so you can have further reference. And also, this is easier to ingest into the RAG if you have one for your knowledge base. You can save that to your collections and reuse this knowledge further in the future for another research.

[00:12:49]So, that's why you have the knowledge base here as well. It's like a library where you can sync it up with your content, as well as collections. What I just did is save it into the history here. So, we have one document here that was created. It shows in here that we have this research report.

[00:13:05]So, what you can do is index this research in the system as well. As you can see, scroll down to here. You got the RAG and indexing state. So, in here, in the history tab that I just found out, once you create your research and the AI finishes the research paper, it shows the document. You can click index all, and it will process your content right into your knowledge base. But, it will take a while to generate since this is not just normal content indexing.

[00:13:30]It's actually putting those into your RAG as well.

[00:13:33]So, yeah, that is it for this video.

[00:13:36]Really practical AI tools that we can run locally. And the AI agents in recent months have been improved a lot. That causes a lot of tedious work to be doable with AI agents nowadays. So, yep, we'll check it out more in the future video. See you guys in the next one.

[00:13:51]Bye.