RAG IA - tech-docs embeddings

[herbergt]

Hi,

I'm trying to set the Backstage tech docs as embeddings, I'm executing the request shown below

I have registered a component into Backstage with the Backstage tech-docs in order to have enough data to embed into the model, however, when I ask something related to backstage the response is very naive, seems that is not taking into consideration the tech-docs data.

The embeddings table seems to have just data related to the Backstage components, I cannot see any reference to tech-docs data

{"kind": "Component", "source": "catalog", "splitId": "1", "entityRef": "component:default/my-component"}

IA Config

# Roadie RAG AI configuration
ai:
  supportedSources: ['catalog']
  embeddings:
    # AWS Bedrock Embeddings configuration
    bedrock:
      # (Required) Name of the Bedrock model to use to create Embeddings.
      modelName: 'amazon.titan-embed-text-v1'`

Am I missing something? Is the embedding of tech-docs implemented or is it something that I need to implement?

[herbergt]

@Xantier do you have any idea regarding this topic? I have been looking through the source code and I see that there are references for embedding documents, just wondering if they are references for the Backstage tech-docs.

[Xantier]

Hi @herbergt ,

Tech docs retrieval or indexing isn't currently yet implemented. Would you be willing to contribute the functionality of that one? I can take a look at the pull request and get it merged in. The call sites that would need to be configured are:

• Retrieval: https://github.com/RoadieHQ/roadie-backstage-plugins/blob/main/plugins/backend/rag-ai-backend-retrieval-augmenter/src/defaultInitializer.ts#L55
• Indexing: https://github.com/RoadieHQ/roadie-backstage-plugins/blob/main/plugins/backend/rag-ai-backend-retrieval-augmenter/src/indexing/DefaultVectorAugmentationIndexer.ts#L111

[herbergt]

Thanks for your response @Xantier , I would like to contribute to this pluggin, just looking for a place to start on how create embeddings based on this kind of documents, do you any suggestion that you can share?

[Xantier]

I believe you would likely need to do few things here to initially get the data to be split and then experiment with the splitting functionality itself.

Splitting/Creating Embeddings:
For the catalog items we are for example on the default naive implementation that you are seeing in the code base now splitting the catalog YAML files simply by using the default RecursiveCharacterSplitter with some configuration options. You can see the call for that in here: https://github.com/RoadieHQ/roadie-backstage-plugins/blob/main/plugins/backend/rag-ai-backend-retrieval-augmenter/src/indexing/DefaultVectorAugmentationIndexer.ts#L90

For HTML files that the tech docs eventually end up being, it might be better to use the HTML option for code splitting as described in here: https://js.langchain.com/docs/modules/data_connection/document_transformers/code_splitter, in case you want to use the langchain library. It is probably a good idea at least to start with that and then iterate further after seeing what kind of results we are getting from it.

The additional metadata to be added to the techdocs embeddings can likely follow the structure that the catalog items get as seen on the constructCatalogEmbeddingDocuments method.

Getting the data is probably something you would need to check from the open source Backstage projects and the API endpoints the tech-docs backend exposes. You should be able to get the HTML documents from there, creating embeddings out of those and eventually using the already existing methods to store those into the vector database.

[Xantier]

We are also on Discord if you want to discuss about implementation details while working on this. You can join it from here: https://discord.com/invite/W3qEMhmx4f

[herbergt]

Thanks, @Xantier for taking the time for the thorough explanation, this guidance gives me a place to start, and I will keep the team posted about implementation details.