Hello, I've been using the

Whats the issue?

I am using the code provided by llamaindex

and getting an error only sometime bc a filter operator is being selected that does not exist within the vectordb

I'll send screen shots

so its the same query for every response

sometimes it works and uses only the sql table

sometimes it works and uses both sql table and vectordb

but then it also fails bc sometimes it uses this "Filter operator function" and none of the filter operators exist

the code itself is from here: https://docs.llamaindex.ai/en/stable/examples/query_engine/SQLAutoVectorQueryEngine/

Right -- not every vector db supports every filter operation, so if the LLM decides to write an unsupported opreation, there's not much that can be done about it

Maybe it should be retrying, or the prompt should be customized for the db you are using

so this may be a novice question but the code itself doesn't have any prompts to the llm only descriptions of the data contained within the individual engines

do you have a recommendation of where I could place a prompt like this

or have an idea on which vectordb i could use that would support all these filter operators

i tried chroma and pinecone already but both have the same issue

You can override the prompt here
https://github.com/run-llama/llama_index/blob/7b52057b717451a801c583fae7efe4c4ad167455/llama-index-core/llama_index/core/indices/vector_store/retrievers/auto_retriever/auto_retriever.py#L80

The default is here (I would modify the prefix and leave the rest)
https://github.com/run-llama/llama_index/blob/7b52057b717451a801c583fae7efe4c4ad167455/llama-index-core/llama_index/core/indices/vector_store/retrievers/auto_retriever/prompts.py#L150

Optionally you can just subclass the retriever to post-process generated filters

so is there a way to just remove the llm's ability/decision path to use filter operators?

as in it will never choose to use filter operators

You could do that if you just override the class yea