All you need to know about Context Engineering

Key Takeaways

Stop wipe coding and start context engineering. Even entre karpathy prefer context engineering over prompt engineering. Vibe coding is good but without context it results in errors. So to build high quality application you need context engineering. Context engineering is providing relevant information such as retrieved knowledge, tool outputs, prior conversation, user input to the large language model so that it can complete the task more effectively. So in summary, context engineering includes rag, state and history, memory, prompt engineering, and structured output. Cognition called out context engineering is effectively the number one job of engineers building AI agents. Anthropic mentioned clearly that agents often engage in conversations spanning hundreds of turns requiring careful context management strategies. This involves adding memory to AI agents such as short-term memory, long-term memory, saving those data in vector databases and retrieving when required. It also includes providing knowledge to AI agents for example your documentation and also providing output from tools. When the user provide a input or a task we need to prepare the context. So this preparation is happening before running AI agents. Then the prepared context is passed to AI agent and the AI agent started running. During that time we need to provide memory and also knowledge. As a result, you get a high quality output. So we are going to focus on all these three and mainly we are going to use MongoDB to store the memory and also to store the knowledge. So the AI agent can retrieve the data whenever required from its own memory and also from its own knowledge. In just three lines of code, you are able to create that context which is required to provide to an AI agent before running it. While running the agent, we can retrieve relevant information from the memory, from its past conversation, and also from its knowledge. So, I'm going to take you through step by step and also explain you what all this about. That's exactly what we're going to see today. Let's get started. Prompt engineering is for one-off task, content generation, and format specific output. and context engineering is for conversational AI document analysis tools and coding assistant. By adding both we can provide production ready AI application. These are the four general categories for context engineering. First write context. So that's what we saw now long-term memories writing that in a scratch pad that is in the session also at a state. So all this data gets retrieved in the runtime when the agents are running and then select context. So we don't need all the data included in context. So we need to retrieve only relevant tools. Retrieve from scratchpad that is session and then retrieve from long-term memory and then retrieve relevant knowledge. And the third step compress context. By compressing the context we make it fit into the context window. So if it's a large code base, you can summarize context to retain relevant tokens. Trim context to remove irrelevant tokens. And finally, the last step is to isolate context. You don't need all the context for all the agents. You can partition context in state, hold an environment or sandbox, partition across multi- aent. I'll put the information about all the links which I'm sharing here in the description below. So in this we are going to use multiple agents such as codebased reader agent, review agent, requirements preparation agent and implementation steps preparation agent. So when we provide with a URL or a goal, it'll automatically go through all these agents and finally give the context document. This includes all the context required to achieve the goal provided and the codebase. So this context is nothing but a detailed prompt which clearly explains the steps required to achieve the goal. So first let's prepare the content that is the context required to run the goal and while the AI agents run we going to provide memory and knowledge to get the final output. So we have a context agent which will automatically go through different agents and prepare the final context document. So we are just providing the URL of the GitHub repo and then I'm going to say need to add authentication that is the goal. So literally three lines of code also I'm going to add GitHub token so that we can access the repo. So you can generate GitHub token from github.com. Even if you want to provide a path to a folder in your computer, this agent should be able to access that folder and prepare the context based on the goal. I'll provide all the code in the description below. So step number one, preparing the context. So just three lines. Now open your terminal. Export your open AI API key like this and then click enter. And then export your GitHub token like this and then click enter. If you're using your local folder, you don't need this GitHub token. Next, pip install praise MongoDB like this and then click enter. That is now installing. Now I'm going to run the code. And now first it went to the structured input passer. So it's automatically extracting the URL and the goal from the input. And you can see here the code base and the goal. Next, it's extracting detail information. It's analyzing individual file and analyzing the code. That's the first agent. Next, enhanced file selection agent. So choosing only relevant files. Not everything is required. Now the relevant files been chosen. Focus repository analyst for each individual file. It's giving a description aligned to the goal that is we need authentication. It's giving the architecture patterns dependencies testing integration. Finally, we got the document here adding authentication to praise AI and it's clearly listing the steps and everything is completed. The document is also stored here. So I can just open this and here it is implementation add authentication the URL and executive summary repository analysis goal relevant components dependencies and is more detailed list of files to use implementation blueprint implementation checklist and that is really nice and now if I have a cursor I can just add the documentation by dragging it here and I can just say implement this feature and you can see it's going through the requirements And it's automatically adding the to-do list which we gave and is now fulfilling the request going through one by one. That is really nice. See how good this context engineering uh providing the context beforehand is useful. Even in win you can do the same. So you got the document here. I provide that document here and I can just say implement this. Now I just need to wait for another 15 minutes or 20 minutes for this implementation to be completed. Next I was telling you about memory and knowledge. So in regards to cursor, cursor has its own inbuilt memory and knowledge. But we can customize this further. So if you are building an AI application, you want this memory and knowledge to be customizable. I'm going to show you how you can do that. For that I'm going to use MongoDB. This can be used as a vector database and also as a key value pair database. And it's available in the cloud version. And you can get started for free. Just choose the free tier and the location where you want to deploy. Then add your username and the password. And this is how you get the connection string username M and get the password. And in the connection string to store your knowledge, I'm adding the username here and the password here. And the remaining connection string you can get it from here. When you go there to the dashboard, click connect and the drivers. There you got the full connection string. So you can just copy this and paste that here with the with your password. So here is the code. So same as before we are providing the context agent which is capable of generating the context and we are taking the output from that and we need to pass that to the goes to the implementation agent. So this agent is going to implement the context which is provided that is the output from the previous agent that is a context agent and we are connecting that to the database. Here for implementation you might need to provide other tools such as code execution, file management, web scraping but for now I'm going to keep it simple just to show you the demo but you can extend this from here. You can include as many tools as you want based on your project requirements. Now going to run and show you. So in your terminal mentioning the file name and clicking enter. This will automatically go through the previous step collecting all the required context and then it goes to the implementation agent and you can see that the implementation agent is writing the code and all the relevant information will get stored in here in database. Similarly, if you want to feed knowledge data to an agent, we can provide to which database the knowledge can be stored. In our case is a MongoDB. So, I've got a document called documentation.md. So, that's where I've got the documentation stored. So I can provide that as a knowledge to this agent. So now you get the point. Giving the right information to the agent to execute its task. This is MongoDB canvas. This is where you can connect to your database with the right information. That is your database information. Connect. And then you can see the database details here. So we got knowledge and memory stored. So if I see memory, you can see all the data here with the embeddings stored in this way. For knowledge, you got the details here. So when the agent start implementing, it'll retrieve relevant information from the memory and knowledge to give a high quality output. That is context engineering. And I'm super excited about this. Do try and let me know in the comments below what you think about this. And thanks to MongoDB for sponsoring this