I am trying to implement small-to-big

I am trying to implement small-to-big retrieval and in this chunk of code

Plain Text

vector_index_chunk = VectorStoreIndex(all_nodes, 
                                              service_context=self.service_context,
                                              storage_context=self.storage_context)
index = VectorStoreIndex.from_vector_store(vector_store=self.vector_store, 
                                           service_context=self.service_context)
vector_retriever_chunk = vector_index_chunk.as_retriever(similarity_top_k=2)
retriever_chunk = RecursiveRetriever("vector",
                                     retriever_dict={"vector": vector_retriever_chunk},
                                     node_dict=all_nodes_dict,
                                     verbose=True)

if rag:
    query_engine_chunk = RetrieverQueryEngine.from_args(retriever_chunk, 
                                                        service_context=self.service_context)
    response = query_engine_chunk.query(query)
    return str(response)
else:
    result = retriever_chunk.retrieve(query)
    res = []
    for node in result:
        res += [display_source_node_custom(node, 200, True)]
    return res

I want the function to store the vector_index_chunks and then to read them from the vector store. When I run this function (the whole thing not just this snippet) with rag = False, I get:

Plain Text

ValueError: Query id c092cd56-9404-43b8-84a0-d591c2cc2dc9 not found in either `retriever_dict` or `query_engine_dict`.

I'm guessing it's an error that the relevant node was not saved somehow, but I can't figure out what exactly causes the error.

11 comments

LLogan M

This won't work with the vector store alone. node/chunks can only be retrieved by ID right now using the docstore

bbixqu

Oh, darn, I was hoping it would work. Indeed, if I remove the vector store, it does run. Any idea if this will be implemented in the future?

LLogan M

Eh, for this, its kind of recommened to use a docstore. We have quite a few docstore integrations (mongodb, redis, firestore, postgres)

For this to change
a) The base vector store class needs a get() method
b) Every (or most) vector stores need to implement this
c) Logic has to be updated under the hood to optionally pull from the vector store vs. docstore

LLogan M

given the amount of work, its kind of low priority lol unless a community member tackles it

bbixqu

Sounds good, thanks 🙂

sstdweird

@Logan M ideally we can pass a docstore to the recursiveretriever, now we have to first make a huge dict and then pass it.

sstdweird

looking at the code, making a RecursiveRetriever instance and then setting the _node_dict attribute to the docstore will probably work?

sstdweird

hmmm, it's not that simple. there is no .get for the docstore, but we probably only need the .get_document for it?

LLogan M

docstore.docs is a dict of node_id -> nodes, I think that works?

sstdweird

in a kv docstore, it's a property that builds up the whole dict; and more direct method/property to get the id would be nice to have. i'll see if i can find what i want. the good thing is that the docstore has no .get, so i can implement one myself in a subclass 😉

sstdweird

actually, i passing a dict means it's not updated, so it's quite annoying if you live update the index

Add a reply

Find answers from the community

I am trying to implement small-to-big