> ## Documentation Index
> Fetch the complete documentation index at: https://agentheya.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# External vector stores

> Connect your existing Pinecone or Weaviate cluster as a private RAG source — the agent queries it at chat time alongside built-in retrieval.

External vector stores let an agent query a Pinecone or Weaviate cluster you already own — **without ingesting anything into Agentheya**. At chat time the platform fans out to each enabled store in parallel, merges the results with Agentheya's built-in keyword and semantic retrieval, and feeds the unified context to the agent.

Use this when:

* You already have a private knowledge base indexed in Pinecone or Weaviate that you don't want to duplicate.
* Your data is too large or too frequently-updated to be a good fit for Agentheya's per-file ingest pipeline.
* You want multiple agents to share a single source of truth (one store, many agents, same connection).

External stores live on the **Internal Data** page only — alongside Files and URLs. They are deliberately not exposed on the Public Data page: owner-controlled API keys never sit on a surface that the agent could cite back to subscribers, and the runtime treats internal-scoped hits as private context.

## Per-card docs

Each backend has its own focused page with the exact connection settings, gotchas, and example payloads:

* [Pinecone](/sections/data/external-stores/pinecone)
* [Weaviate](/sections/data/external-stores/weaviate)

## How it fits into retrieval

Every turn, Agentheya combines its built-in keyword and semantic search over your internal and public knowledge with results from each enabled external store, then assembles the most relevant context for the agent.

Built-in results are intelligently ranked and combined; external-store results are **added** to the candidate set rather than re-scored — vendor relevance scores aren't comparable across systems, and each vendor already does its own ranking. You can optionally apply a reranking model before the context is given to the agent.

The retrieval trace shows one entry per external store with hit counts (or skip / error reasons), so you can diagnose live on the monitoring page.

## Secrets

The three secret fields — `pinecone_api_key`, `weaviate_api_key`, `embedding_api_key` — are AES-256-GCM-encrypted at rest before they're written to disk. They never leave the server in plaintext; the dashboard editor sees only the sentinel `***REDACTED***` until you click **Replace** to rotate.

## Schema enforcement

However you configure a store — through the dashboard, the Management MCP, or the AI Designer — the same validation applies. A misconfigured `pinecone` block missing its `pinecone_api_key`, or a `weaviate` block missing `weaviate_class`, is rejected with a clear error before anything reaches disk. The same enforcement applies to MCP `add_section_item({ section: "data_internal", array_field: "external_stores", item: { ... } })` calls.

External stores can only be added to your Internal Data; attempts to attach them to Public Data are rejected.