Any ways I can embed a directory hierarchy structure as well as the text? Or any way I can prompt a retriever?

[djsteve1223]

Hi, I'm currently trying to create a chatbot over documents that are stored in a tree-like hierarchy structure.
I've tried to search over vector databases that takes the directory structure into account as well as the text is self, but I couldn't find any. Is there any way to do so?
I'm currently using chroma and saved the persist directory on my local workspace.
I have a couple of questions.

1. Other than just ".as_retriever.get_relevant_documents()" method, is there any way I can "induce" the retriver to query over the database, taking the hierarchy structure into account? It seems that '.get_relevant_documents' just searches over the whole text corpus without taking the directory structure into account

After observing the above problem, I've tried to bypass the problem by wrapping the original query with relevant information about the hierarchy structure. I've created a new prompt template that explains the file structure and passed the whole prompt into the retriver. However, the retriever doesn't take the prompt as a "guide", but the prompt itself as a whole "question". Thus, the most relevant answer that the LLM found was the document that contained names of all categories mentiioned below.

2. how can I solve this problem? can I prompt guide the retreiever?

`retrieval_query_modification_template = """
You must query within the policy book of the firm ~~~~.

The provided documents were originally stored in a tree-structured hierarchy, and the hierarchy is labeled as "hierarchy" on the metadata.
The first branches are labeled on the "category" metadata.

Here are some brief summurizations on each categories

1. Category 1
Its explanation

2. Category 2
Its explanation

Trye to exploit the metadata "source", "hierarchy", "category" while searching.

This is the question originally asked by the consumer. 
You must find relevant documents

{question}

Relevant Documents:

"""

retrieval_query_prompt = PromptTemplate(template=retrieval_query_modification_template, input_variables=['question'])

query = retrieval_query_prompt.format(question=question)

retriever = db.as_retriever(search_type='mmr')

retriever.get_relevant_documents(query)

Any suggestions, comments would be helpful.
Thank you

[dineshresearch]

@dosu-bot could you answer this question?