Sync PostgreSQL data to a Pinecone vector knowledge base using Gemini embeddings

PostgreSQL Auto-Discovery to Pinecone Vector Knowledge Base

Turn existing PostgreSQL databases into AI-searchable Pinecone vector knowledge bases without manually defining every table and column.

This template discovers your PostgreSQL schema, selects text-rich tables, builds readable embedding documents from rows, generates Google Gemini embeddings, upserts vectors into Pinecone, and tracks per-table sync state for incremental indexing.

It also includes a companion semantic search webhook workflow that embeds a user query, queries Pinecone, and returns source row metadata for RAG and AI assistant use cases.

Features

Dynamic PostgreSQL schema discovery using information_schema. No hard-coded schema, table, or column names. Primary key detection. Timestamp column detection for incremental sync. Text-rich column detection for embedding. Default sensitive-column exclusion. Configurable allowed schemas, allowed tables, excluded tables, excluded columns, and max rows per table. Google Gemini embeddings with gemini-embedding-001. Pinecone upsert using namespaces. Per-table sync state stored in n8n workflow static data. Fail-fast error behavior so PostgreSQL, Gemini, or Pinecone errors stop execution. Companion semantic search webhook for testing RAG retrieval.

Architecture

The main workflow runs through these stages:

Manual or scheduled trigger. Configuration. PostgreSQL metadata discovery. Table analysis and filtering. Dynamic row fetching for selected tables. Row-to-document transformation. Gemini embedding batch preparation. Gemini embedding generation. Pinecone vector batch preparation. Pinecone upsert. Sync state update and summary output.

How Auto-Discovery Works

The workflow inspects PostgreSQL metadata, groups columns by table, and chooses indexable tables based on:

Allowed and excluded schema/table settings. Presence of a primary key, if required. Presence of text or JSON columns. Presence of timestamp columns, unless full-sync fallback is enabled. Sensitive-column filtering.

Security Filtering

The workflow excludes sensitive columns before documents are built. By default, it filters column names containing:

password passwd pwd token secret api_key apikey access_key refresh_token reset_token otp pin session cookie hash salt private_key client_secret

You can add more column names in the excludedColumns setting.

Important: review your database schema and compliance requirements before enabling indexing in production. This workflow provides safe defaults, but it cannot understand every business-specific sensitive field.

Supported PostgreSQL Column Types

The workflow treats these as text-rich embedding candidates:

text varchar character varying char character json jsonb

When includeScalarContextColumns is enabled, the document also includes useful scalar context from:

integers and numeric types booleans dates timestamps UUIDs

PostgreSQL Setup

Create or select a PostgreSQL credential in n8n with read-only access.

Recommended permissions:

GRANT USAGE ON SCHEMA public TO your_n8n_user; GRANT SELECT ON ALL TABLES IN SCHEMA public TO your_n8n_user; ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT SELECT ON TABLES TO your_n8n_user;

For production, prefer a dedicated read-only user and restrict schemas with the workflow allowedSchemas setting.

Pinecone Setup

Create a Pinecone index. Use the same vector dimension as the embedding model. This workflow requests 1536 dimensions from Gemini so it can use a Pinecone index configured for 1536-dimensional vectors. Copy the index host, for example your-index-abc123.svc.us-east-1-aws.pinecone.io. Create an n8n HTTP Header Auth credential: Header name: Api-Key Header value: your Pinecone API key. Set pineconeIndexHost in the workflow configuration.

The workflow uses Pinecone namespaces in this format:

postgres-{schema}.{table}

Example:

postgres-public.tickets

Gemini Setup

Create an n8n HTTP Header Auth credential:

Header name: x-goog-api-key Header value: your Gemini API key from Google AI Studio

The default embedding model is:

gemini-embedding-001

This workflow requests geminiOutputDimensionality = 1536. Pinecone index dimensions must match the vectors you upsert.

Why HTTP Request Nodes Are Used

n8n has native AI/vector nodes, including embedding and Pinecone vector store nodes. This template uses HTTP Request nodes for Gemini and Pinecone because the workflow needs precise control over production indexing behavior:

Stable custom vector IDs in the format postgres::{schema}::{table}::{primary_key_value}. Exact Pinecone metadata for tracing search results back to PostgreSQL rows. Dynamic namespaces for each discovered table. Batch-level Pinecone upsert handling. Sync state updates only after successful upserts. Raw Pinecone query responses for the companion semantic search API.

The native nodes are excellent for standard document ingestion and retrieval flows. For this auto-discovery template, HTTP requests make the workflow more predictable and easier to customize for database indexing.

Main Workflow Configuration

Open the Workflow Configuration node and set:

| Field | Description | | --- | --- | | allowedSchemas | Comma-separated schemas to include. Default: public. Leave empty to allow all non-system schemas. | | allowedTables | Optional comma-separated table names or schema.table names to include. | | excludedTables | Tables to skip. Supports table names or schema.table. | | excludedColumns | Additional columns to exclude. | | maxRowsPerTable | Maximum rows fetched per table per run. Default: 10 for safe first runs. | | maxDocumentsPerRun | Hard cap on documents sent to Gemini in one execution. Default: 10 so first tests finish quickly. | | fullSyncTablesWithoutTimestamp | If true, tables without timestamp columns receive limited full syncs. | | skipTablesWithoutPrimaryKey | If true, tables without primary keys are skipped. | | includeScalarContextColumns | Adds scalar values such as status, price, priority, dates, and UUIDs to documents. | | maxDocumentCharacters | Truncates very large documents before embedding. Default: 2000. | | geminiEmbeddingModel | Gemini embedding model. Default: gemini-embedding-001. | | geminiOutputDimensionality | Output vector size requested from Gemini. Default: 1536. | | geminiBatchSize | Number of documents sent to Gemini per embeddings request. Default: 1 for safe first runs. | | geminiSmokeTestMode | Sends only one document to Gemini. Default: true. Disable after one-document embedding works. | | pineconeIndexHost | Your Pinecone index host. | | pineconeNamespacePrefix | Namespace prefix. Default: postgres-. | | pineconeBatchSize | Number of vectors per Pinecone upsert request. |

Document Format

Each PostgreSQL row becomes a readable document for embedding:

Database: postgres Schema: public Table: tickets Id: 123 Updated At: 2026-05-17T12:00:00.000Z Title: Payment failed for user Description: Customer payment failed after webhook timeout Status: open Priority: high

The goal is to preserve business meaning instead of embedding raw JSON alone.

Pinecone Vector Structure

Each vector is upserted like this:

{ "id": "postgres::public::tickets::123", "values": [0.0123, -0.0456], "metadata": { "source": "postgres", "schema": "public", "table": "tickets", "primary_key": "id", "primary_key_value": "123", "updated_at": "2026-05-17T12:00:00.000Z", "text_columns": "title,description", "document_columns": "id,updated_at,title,description,status,priority" } }

Vector ID format:

postgres::{schema}::{table}::{primary_key_value}

Workflow Setup Steps

Import workflow.json into n8n. Add your PostgreSQL credential to both PostgreSQL nodes. Add your Gemini HTTP Header Auth credential to Generate Gemini Embeddings. Add your Pinecone HTTP Header Auth credential to Upsert Vectors to Pinecone. Open Workflow Configuration. Set allowedSchemas, pineconeIndexHost, and any table or column filters. Run manually with a small maxRowsPerTable for the first test. Inspect the Return Sync Summary output. Confirm vectors appear in Pinecone. Enable the optional Cron trigger when ready.

Then import semantic-search-workflow.json and configure its Gemini and Pinecone credentials.

Troubleshooting

No tables are indexed

Check:

allowedSchemas matches your schema. Tables have at least one supported text or JSON column. Tables have a primary key, or disable skipTablesWithoutPrimaryKey. Tables have a timestamp column, or enable fullSyncTablesWithoutTimestamp.

Gemini embedding request fails

Check:

HTTP Header Auth uses x-goog-api-key. The value is your Gemini API key from Google AI Studio. Your Google AI Studio project has access to gemini-embedding-001.

If you receive status 429, Gemini is rate limiting the workflow or the Google AI Studio project does not have enough available quota. n8n may show a generic batching/rate-limit message for both cases.

First run with smoke-test settings:

geminiBatchSize: 1 geminiSmokeTestMode: true maxRowsPerTable: 1 maxDocumentsPerRun: 1 maxDocumentCharacters: 1000

The Generate Gemini Embeddings HTTP Request node should also have Options > Batching enabled with:

Items per Batch: 1 Batch Interval: 5000 or higher

This spaces out the HTTP requests after the documents have already been grouped into embedding payloads.

If the workflow still returns 429 with smoke-test mode enabled, check your Google AI Studio project:

Billing is active. The project has remaining credits or monthly spend limit available. The API key belongs to the project you are checking. The project has access to gemini-embedding-001. No other workflow or app is using the same project limits at the same time.

Gemini embeddings are documented here: https://ai.google.dev/api/embeddings.

Gemini embedding node keeps loading

This usually means the workflow is processing more documents than expected while also spacing requests out to avoid rate limits. For example, 100 documents with a 10 second batch interval can take more than 16 minutes before retries.

For a quick smoke test, use:

maxRowsPerTable: 1 maxDocumentsPerRun: 1 geminiBatchSize: 1 Generate Gemini Embeddings > Options > Batching > Batch Interval: 1000

After the smoke test succeeds, increase the limits gradually.

Pinecone upsert fails

Check:

HTTP Header Auth uses Api-Key. pineconeIndexHost is the host only, without https://. The Pinecone index dimension is 1024. Your Pinecone project and API key match the index.

Incremental sync does not pick up rows

Check:

Your table has one of the supported timestamp columns. The timestamp column is updated when rows change. The stored static data lastSync is not newer than your test row.

Too many rows are embedded

Lower maxRowsPerTable, restrict allowedTables, or keep fullSyncTablesWithoutTimestamp disabled.

Performance Considerations

Start with a low maxRowsPerTable such as 1 or 10. Use allowedSchemas and allowedTables to limit the first run. Add indexes on timestamp columns used for incremental sync. Keep pineconeBatchSize between 25 and 100. Consider separate workflows for very large schemas or high-volume tables. Avoid embedding huge JSON blobs unless they are meaningful for search.

Cost Considerations

Costs come primarily from:

Gemini embedding tokens. Pinecone vector storage. Pinecone query and write operations. n8n execution time.

To control costs:

Exclude low-value tables. Exclude noisy or sensitive columns. Reduce maxDocumentCharacters. Use incremental sync. Test with small row limits before enabling scheduled runs.

Customization Guide

Common customizations:

Add more timestamp candidates in Filter Indexable Tables. Add business-specific sensitive column names. Change namespace format. Include only selected scalar columns. Add tenant or environment metadata. Add a notification step after Return Sync Summary. Route failed batches to Slack, email, or an incident workflow. Add a second Pinecone query workflow for internal tools or chatbots.

Use Cases

AI customer support over tickets and helpdesk records. RAG systems over business databases. Internal company knowledge bases. CRM AI assistants. Support ticket semantic search. E-commerce product discovery. Real estate listing search. AI copilots over operational databases. Semantic search for audits, notes, cases, articles, and documentation.

Assumptions

PostgreSQL tables use primary keys for stable vector IDs. Best incremental sync results come from updated_at, modified_at, or last_modified columns. Pinecone index dimension matches geminiOutputDimensionality. Users review security settings before indexing production data. n8n can reach PostgreSQL, Gemini, and Pinecone from its runtime environment.