# Diagnosis
## Symptom
`PineconeVectorStore(...)` builds fine, then `add_documents` or `similarity_search` fails with a not-found or connection error.
## Likely cause
PineconeVectorStore assumes the index already exists. Construction is lazy; the failure surfaces at first use, which misleads agents into debugging LangChain instead of the missing index.
## Confirm it
1. `pc.list_indexes()` with the raw Pinecone client: is the name there?
2. If present, `describe_index`: does its dimension match `len(embeddings.embed_query('test'))`?
3. Check the namespace: the store may target a namespace while you verified the default.
## Fix
1. Create the index first (serverless spec, right dimension and metric), then construct the store.
2. Or point the store at the existing correct index.
3. Make index creation part of your setup script, ordered before any LangChain code runs.
## Verify the fix
Add one document through the store, then `similarity_search` it back; assert the round trip works before bulk ingest.