AWS Bedrock Knowledge Base Tutorial: RAG with OpenSearch & Titan Embeddings

Key Takeaways

Let's go to creating the second thing which is knowledge basis. So on the search I would mention knowledge basis. We can see Amazon bedrock feature knowledge basis. We'll click on it. every time we just have to ensure that we are in the Mumbai region just to make sure everything works otherwise uh we might see errors in our application. All right. So we'll create a knowledge base. I will select the first option which is knowledge base with vector store which takes us up here. We can give it a nice name. Now the knowledge base itself needs permission to access different services and for that it wants to create a service role. We can mention the data source type which is Amazon S3 that that is where all our documents are. But there are different source types that knowledge bases also supports from which it can fetch these documents and work with them. All right. Click on next. Now because we selected Amazon S3, it is asking where the documents are there in Amazon S3. So we'll just give it a nice name. We'll mention our S3 bucket AB demo geni bucket and we have another option called passing strategy. So this is about how these documents when converted after processing into text should actually be split. And so we'd be using the default parser and then we'd also be using the default chunking strategy. So the parsing of the documents would be done by default Amazon Bedrock service. The chunking strategy is about once the documents are passed, how we'd like to break it into. So we'd be breaking these documents into 300 token in size and then every chunk would then be given to embeddings to be created into numerical format. All right. So what embedding model we would be using to convert this text into numerical format? We'd be using one of the Amazon's default Titan embeddings, Titan text embeddings v2 and then vector store which is finally where to store these embeddings. We'll be using one of the service Amazon open search serverless. So bedrock is going to create that as well for us and simply click on next. We can review everything over here and we'll click on create knowledge base. So this is going to take around 5 minutes to create multiple services and knowledge base is now creating the open search vector database and would then also be creating multiple services at the back. And this is going to simplify our rag processing really easy because knowledge basis takes care of both putting these documents into open search and also whenever we ask question it will do all the embedding all the searching and would simply give us the final results. So we'll wait 5 minutes or we'll wait until this is complete. All right, our knowledge base is finally ready. We can have a look at it. Now we'll go into data source. The S3 bucket that we connected have it actually synced which means there is data inside it but nothing is actually created in terms of vector embeddings. So as we discussed this knowledge base is going to create an open search vector DB. So let's have a look. We'll search open search. We're opening it in a new tab. On the left, we can click on collections and we see there's a bedrock knowledge base collection created. We click on it and indexes. We can see that the whole vector DB is empty. Which means we have to sync these documents and this is very easy in knowledge bases. So if you're in the knowledge base in data source we can select multiple data sources. In this one we just have S3. We click on it and we simply have to click on sync. And now these documents are being fetched. So the PDF that we had is being fetched converted into text using parser. The text is then split by the chunking strategy which is 300 tokens each. Then Titan embeddings are called into every chunk. They are conver converted into numerical formats and along with the numerical format and the text for which those numerical format is generated is then saved into AWS open search. So all of this process is happening at the back. This shows sync complete. Let's refresh the page. And as we can see that the sync history shows that the syncing was done at this time and it is complete. We can also go to the open search vector and we can see that there is index and multiple documents which are present. All right. So this means that all of the right portion over here which is for the rack pipeline is complete. The ingestion of documents is also done and then we have a filled vector DB. We can now start using it into our application. So let's copy the knowledge base ID. So the knowledge base ID is not the name uh but the ID that we can see in the dashboard. So if we click on our knowledge base, we have the knowledge base name and then the knowledge base ID. So I'll just copy the knowledge base ID over here. All right. So now I want to connect to this knowledge base and ask it questions so it can give me back some related relevant documents based on that question. So let's head on to our code and in ENV we'll add this missing piece of knowledge base ID which completes our ENV file end to end. All right.