No intermediate calls. It just looks at the
df.head()
output and the user query, generates some python code that is evaluated, and then it returns the output π
If you need embeddings, I suggesting moving the dataframe to a sqlite database and using the sql + vector query engine, where the vector engine is just each row embeded?
https://github.com/jerryjliu/llama_index/blob/main/docs/examples/query_engine/SQLJoinQueryEngine.ipynb