Adding proposal on Generative AI Feature Pack #580
Conversation
github-actions
bot
added
the
invalid-categories
Signed-off-by: Emmanuel Hugonnet <ehugonne@redhat.com>
github-actions
bot
removed
the
invalid-categories
| * model-name: the name of the embedding model served by Ollama. | ||
|
|
||
| ---- | ||
| subsystem=ai/ollama-embedding-model=myembedding:add(base-url="http://192.168.1.11:11434", model-name="llama3:8b") |
There was a problem hiding this comment.
I will pick ollama-embedding-model, but this is general objection. I am asking myself how good idea is to map LangChain4J model API to Wildfly model 1:1. LangChain4j API is very broad and will change. Wildfly model will always be lagging. Aren't we misusing wildfly model as Inversion of Control system here?
This is different to Hibernate (jpa) subsystem or datasource (jdbc) subsystem which are small. If this approach will be taken ai subsystem can be over time very big (hard to maintain) subsystem.
But I am not sure what could be a good alternative approach :) Hiding implementation details? Expose in Wildfly model just what is common to any implementation; url modelname in this case?
There was a problem hiding this comment.
Well, I wondered myself. langchain4j is already an abstraction over the llms and other elements of the rag chain so adding one more is maybe too much. Also you don't have to match 1-1, you only need to expose those attributes you think make sense. Would having a map of attributes to a generic type that would somehow do the matching be more to your liking ?
| ---- | ||
|
|
||
| It should also support vector database backed embedding store like for Weaviate. | ||
| It should expose a simple `weaviate-embedding-store` resource with the following attributes: |
There was a problem hiding this comment.
Followup on my concern expressed above. For example there is a lot of relevant vector dbs out there (chroma, qdrant, milvus, pinecone, elasticsearch, ...) https://lakefs.io/blog/12-vector-databases-2023/ I am questioning if providing resources for all of them does scale?
There was a problem hiding this comment.
We don't have to support everyone of them, all the more so as each bring their own dependency tree
No description provided.