n8n vector store integrations transform documents into embeddings and power
semantic search that retrieves meaning, not just keywords. You choose between
three deployment models: pgvector self‑hosts inside your
existing PostgreSQL instance, Pinecone offers a managed API‑first
vector database with purpose‑built nodes, and Weaviate provides
either a self‑hosted or cloud‑managed option. This guide details the common RAG
blueprint applicable to all three stores, then provides store‑specific
configuration, a comparison table, and AI Agent patterns for production‑grade
retrieval‑augmented generation.

How do you build a universal RAG pipeline blueprint in n8n that works across all vector stores?

A production RAG pipeline in n8n follows a six‑stage blueprint:
Ingest (Default Data Loader reads PDFs/HTML/JSON) →
Chunk (Recursive Character Text Splitter with
chunk_size and chunk_overlap) → Embed
(Embeddings OpenAI or Cohere) → Store (any vector store node)
→ Retrieve (Vector Store Retriever or AI Agent Tool mode) →
Answer (Question & Answer Chain fed by retriever).

The “keep metadata in chunks, embed on insert, and attach source citations at
the end” pattern works universally. text-embedding-3-small (1 536
dimensions) as embedder, chunk_size: 1000 and chunk_overlap:
200 are commonly used as flexible starting points for text‑based RAG.
Attach source URLs in metadata.source, and at the final formatter
output, join citations with a simple expression that iterates over
$json.results to list each source. For real‑world applications of
this pipeline, see the
n8n AI Agents & LLM Orchestration guide.

How do you configure a Pinecone vector store node for hybrid search and metadata filtering in n8n?

Pinecone provides two integration paths in n8n: the Pinecone Vector
Store node for full pipeline control and the Pinecone
Assistant node for managed RAG with minimal setup. For production
agent flows, use the Vector Store node in “Retrieve Documents (As Tool for
AI Agent)” mode — this is the default and recommended operation since n8n
v1.33.0 added serverless index support.

Never omit the “Tool Description” field: the AI Agent reads it to decide
when to query the vector store; leaving it empty causes agent failures.
Delete any legacy “Answer questions with a vector store” node if present —
these older nodes have known bugs. For embedding dimensions, match exactly to
your model: text-embedding-ada-002 expects exactly 1 536
dimensions. Set the namespace option to true if your index uses
one, otherwise the connection will fail. Pinecone integrates at the tool level
with AI Agents; the official RAG template demonstrates the complete flow from
Google Drive to Pinecone to OpenAI Chat.
For detailed credential setup, see the
n8n Credential Security guide.

How do you set up Weaviate as a self‑hosted or cloud vector store in n8n?

The Weaviate Vector Store node supports four modes: Get Many,
Insert Documents, Retrieve Documents (As Vector
Store for Chain/Tool), and Retrieve Documents (As Tool for
AI Agent). The last mode connects directly to an AI Agent’s tool
connector.

Weaviate supports self‑hosted and cloud clusters. For self‑hosted, start a
Weaviate cluster with Docker Compose, then enter the instance URL and API
key in n8n credentials. For cloud, sign up for Weaviate Cloud Services,
create a cluster, and use the provided endpoint and key. A standard
credential panel accepts the Weaviate URL and API key for either path.
The official “Document Q&A with RAG: Query PDF content using Weaviate
and OpenAI” template demonstrates a minimal RAG implementation that uploads
a PDF, generates embeddings with OpenAI, stores them in a Weaviate
collection, and provides a chat interface for natural‑language queries.
For the complete set of node parameters and vectorizer options, see the
n8n Nodes & Techniques Hub.

How do you run pgvector locally with Docker and connect it to n8n as a vector store?

Spin up pgvector with Docker: pull pgvector/pgvector:pg16, run
a container with persistent storage, then enable the extension and create a
table. Next, connect n8n to this instance using a Postgres
node with the same credentials, then open a Canvas and add the
PGVector Vector Store node. Use the Insert Documents mode
to store embeddings and Retrieve Documents for queries.

After pulling the pgvector image and starting the container, connect with
psql as postgres to create the database and enable
the vector extension. Then create a table with a vector
column matching your embedding model’s dimensions (1 536 for
text‑embedding‑3‑small, 3 072 for text‑embedding‑3‑large). Add an HNSW
index (or IVFFlat for smaller datasets) to accelerate cosine similarity
queries over millions of vectors.
The PGVector Vector Store node in n8n’s AI package
connects to your PostgreSQL database and uses the table you created.
For production, pgvector runs inside your existing PostgreSQL backup
strategy — no new infrastructure. For scaling pgvector to high‑volume
workloads with queue mode, see the
n8n Scaling & Queue Configuration guide.

How do you connect a vector store directly to the AI Agent’s tool connector?

In the AI Agent node, locate the tool connector on the
right side of the node. Connect it to a vector store node set to
“Retrieve Documents (As Tool for AI Agent)” — this is the
correct mode for agentic RAG since n8n v1.33.0. Fill in the “Tool
Description” field with natural language so the agent knows when to search
the vector store.

The legacy mode “Retrieve Documents (As Vector Store for Chain/Tool)” works
for older chain‑based workflows but does not work with Pinecone serverless
indexes and should be avoided for new agent projects. The AI Agent decides
autonomously when to query the vector store based on the tool description.
You can connect multiple vector store tools to a single agent, each with a
different description — for example, one for policy documents and another
for product manuals. The agent selects the appropriate store at runtime
based on the user’s question. For debugging, check the agent’s execution
log to see which tool was selected for each query. For more on agent-based
workflows, see our
n8n AI Agents & LLM Orchestration guide.

How do you optimize vector store performance, cost, and retrieval quality for production RAG?

Pinecone is purpose‑built for vector search with hands‑off operations and
hybrid (semantic + lexical) search — best for teams without existing
database infrastructure. Weaviate provides built‑in vectorization modules,
flexible deployment, and multi‑tenant collections — optimal for
organizations needing on‑premises deployment. pgvector reuses existing
PostgreSQL tooling, simplifies the operational footprint to a single
database for backups, and supports both HNSW and IVFFlat indexes — ideal
for teams already running Postgres who want to add vector search at no
additional cost.

Dimension matching is the #1 pitfall: mismatched embedding dimensions cause
silent failures — always verify that text-embedding-3-large
generates 3 072‑dimensional vectors (or as downsized) and configure your
vector store table or index to match. The second most important
consideration is chunking: a chunk_size of 1 000 tokens with
200‑token overlap preserves context without generating excessive
dimensions. For massive datasets, Qdrant provides another viable option
with native batch ingestion and automatic collection creation — a
configuration that minimizes manual provisioning. The ecosystem also
includes Supabase Vector Store (managed pgvector with native n8n
integration) and Chroma for lightweight local RAG testing. For a complete
dataset generation walkthrough that integrates with vector stores, see the
n8n OpenAI Prompt Chain Tutorial.