Knowledge Stack
Basic troubleshooting steps for Knowledge Stack
Having an issue with a knowledge stack? Let's first make sure that the basics are working. Here are some steps to troubleshoot:
Check if Knowledge Stack is working
Download MixedBread Embed model
If you haven't already done so, download the MixedBread Embed model from the Msty app. We'll use this to check if the Knowledge Stack is working.
Create a new Knowledge Stack
Select MixedBread as the embedding model.
Use the following content for the Knowledge Stack
Copy the following text to a .txt
file on your device. Then, add it to the Files section of your new Knowledge Stack.
Compose stack
After adding the .txt
file, select Compose.
Select option to view chunks
After composing is complete, select the icon next to the green 'Composed & Ready to Chat' message. This button opens the Knowledge Stack Chunks Console where you can see what chunks are passed along when chatting with a model.
Insert 'test' into the input field and click 'Run'
This will run a query and display results in the console.
Verify
If you see the chunks in the console output, then it means that the Knowledge Stack you created can be used when chatting with a model and that Knowledge Stacks are working properly.
If you can see outputs for this test but the prior one is not working as expected, continue to the next section to view the possible issues.
Possible Issues
Incompatible file type
Knowledge Stacks work best with text files (.md, .txt, etc) and pdfs.
File size too small
If the text content in the individual files are less than 50 characters, they will most likely be ignored.
Tiny files aren't designed to be used as a RAG source.
Verify the model is compatible with RAG
The model you chose for embedding
is indeed an embedding model and not a chat model. This is unlikely but may be possible.
You restarted Msty
If you've restarted Msty, check that there are no rogue Msty-App processes running that may be interfering with Msty and the Knowledge Stack.
Understanding the limitations of RAG
Possibly the main thing to understand about RAG is that it doesn't work like a traditional search engine - it doesn't have the full context of the whole complete files but rather 'chunks', or 'vectors', created as a result of vectorization.
With RAG, the model retrieves relevant chunks of information based on the query. So, for example, if you ask to summarize a file, it may not be able to do that properly because only the most relevant chunks are provided. Meaning, the model doesn't have the full context of the document, and it may not be able to provide a coherent summary.
Some models may infer the context of the chunks and provide a more coherent response, but this is not guaranteed and may even lead to incorrect or misleading information - a.k.a "hallucinations".
This may not present as an issue for most use-cases; however, for a few, such as with summarizations, it is an important limitation to be aware of.