Pinecone

AddisDB connects to Pinecone’s data-plane API. A connection points at a single index host, and from there you can list a namespace, run a raw query body, or issue an explicit request such as describe_index_stats.

Who it is for

Pinecone is a fully managed vector database — no nodes, no index tuning, no capacity planning. It scales to very large vector counts with consistent low latency, which is the whole reason to pay for it.

Namespaces are its partitioning primitive and the feature most worth using deliberately: a query runs inside one namespace, so per-tenant or per-corpus separation is enforced by the API rather than by a metadata filter you have to remember to add.

It fits production RAG and semantic search where you would rather not own the operational burden of a vector cluster, and where the workload is large enough that "just use pgvector" has stopped being the right answer.

Set up the index

1. In the Pinecone console, create an index with the dimension and metric matching your embedding model.
2. Copy the index host — a URL ending in .svc.<environment>.pinecone.io. This is per-index, not per-project.
3. Under API Keys, copy your API key.
4. Note your namespaces if you use them; an empty namespace is the default.
5. The dimension and metric are fixed at creation — changing either means a new index and a re-upsert.

Connect from AddisDB

1. New Connection → Pinecone under Vector / AI. Port prefills to 443.
2. Paste the full index host URL into the Connection URL box, or put the hostname in Host.
3. Paste your API key into the API key / token field.
4. Namespace is optional — set one as the default or leave it blank.
5. Test, then Save.

The index host is not the console URL and not a project-wide endpoint. Taking it from the index page in the console is the difference between connecting and getting a 404 that looks like an auth failure.

How to query it

// A namespace name — lists vectors in it
production

// A raw query body
{ "topK": 20, "includeMetadata": true, "vector": [0.01, -0.2, …] }

// Or an explicit request
POST /describe_index_stats {}

describe_index_stats is the first thing to run against an index you did not build: it reports dimension, total vector count, and the count per namespace — which immediately tells you whether the data you expected actually landed where you expected.

Metadata filters go in the same body as the query and are applied during the search. Keep the filterable fields few and low-cardinality; Pinecone indexes metadata separately, and a filter on a free-text field is not what that index is for.

{ "topK": 10, "includeMetadata": true,
  "vector": [0.01, -0.2, …],
  "filter": { "tenant": { "$eq": "acme" }, "year": { "$gte": 2024 } } }

What AddisDB gives you

• Matches flattened into the grid — metadata keys become columns alongside id and score.
• The Vectors view for seeing how the returned embeddings sit relative to each other.
• The JSON tab for the raw API response.
• Read-only enforcement — DELETE is blocked, as are POSTs to upsert, delete and update endpoints.
• Vector counts per namespace feeding the size alerts.
• Query tabs and history, and ⌘K search across every connection.