Metadata

ss1l3nt

team, i've built a basic RAG using (SimpleDirectoryReader, VectorStoreIndex, index.as_query_engine()) and querying using query_engine.query().
the LLM prompt gets appended with the context returned by retriever. i found that the context text contains file_path which is reducing the output for my LLM.
how can i specify that i don't want to add file_path to the context of the LLM, but only the retrieved text?

4 comments

LLogan M

When ingesting data, you can specify metada to exclude from the llm, from the embedding model, or botv

LLogan M

https://docs.llamaindex.ai/en/stable/module_guides/loading/documents_and_nodes/usage_documents/#advanced-metadata-customization

ss1l3nt

for the popular examples in llamaindex repo, we usually do the following step:

Plain Text

documents = SimpleDirectoryReader("data/pg_essage.txt").load_data()

which makes documents a list of Document objects.
to exclude the file_path metadata from llm/embedding, i'll have to call the method on each and every Document object in documents list using something like:

Plain Text

for doc in documents: doc.excluded_llm_metadata_keys.append("file_path")

i'm wondering if there's a better approach than this for loop.

LLogan M

Nah that's the best approach. Pretty straightforward

Add a reply

Find answers from the community

Metadata