Thanks again for this Logan M I started

Thanks again for this @Logan M. I started investigating Pinecone as a vector store. I'm having trouble finding:

How do you delete documents from an index.
How to add a namespace.
How to add documents under a namespace.
How do you add metadata for filtering for a query.
How to update metadata for a filtering on existing docs.

10 comments

When you insert documents, each Document has a doc_id. You can eithe record this ID or set it to something else before inserting. Then you can use index.delete(doc_id). This will delete all nodes created from this document

If you need to, you can also check source nodes from a response to get the doc id

response.source_nodes[0].node.ref_doc_id

LLogan M

namespace is configured in the pinecone vector store constructor

PineconeVectorStore(..., namespace="my_namespace")

LLogan M

Updating metdata is not really supported for any index right now -- need to delete and re-add

LLogan M

you can filter like shown in this notebook
https://github.com/jerryjliu/llama_index/blob/main/docs/examples/vector_stores/pinecone_metadata_filter.ipynb

OOrion Pax

Thanks again. How does one add a single document to an existing index? Does VectorStoreIndex.from_documents() add to an existing index or is there another function to insert a single document to an existing index? I'm only see tutorials in the context of creating the index from scratch.

LLogan M

index.insert(document)

OOrion Pax

So, something like this?

Plain Text

index = VectorStoreIndex(
  vector_store=vector_store,
  storage_context=storage_context,
  service_context=service_context,
)
[index.insert(doc) for doc in documents]

LLogan M

uhhh not quite

This will create an empty index, that you are then inserting into

Plain Text

index = VectorStoreIndex(
  [],
  storage_context=storage_context,
  service_context=service_context,
)
[index.insert(doc) for doc in documents]

LLogan M

If you have an existing vector index, you can load it from disk using load_index_from_storage()

If you are using a vector db integration, you can connect back to it by setting up the vector store object to point to the existing index, and then do

Plain Text

index = VectorStoreIndex.from_vector_store(
  vector_store,
  service_context=service_context,
)

OOrion Pax

Perfect. Thanks

Add a reply

Find answers from the community

Thanks again for this Logan M I started