databricks-vector-search

name	databricks-vector-search
description	Databricks Vector Search endpoints and indexes for RAG and semantic search; covers index types, search modes, end-to-end RAG patterns
metadata	{"version":"0.1.0"}
parent	databricks-core

Databricks Vector Search

FIRST: Use the parent databricks-core skill for CLI basics, authentication, and profile selection.

Patterns for creating, managing, and querying vector search indexes for RAG and semantic search applications.

When to Use

Use this skill when:

• Building RAG (Retrieval-Augmented Generation) applications
• Implementing semantic search or similarity matching
• Creating vector indexes from Delta tables
• Choosing between storage-optimized and standard endpoints
• Querying vector indexes with filters

Overview

Databricks Vector Search provides managed vector similarity search with automatic embedding generation and Delta Lake integration.

Component	Description
Endpoint	Compute resource hosting indexes (Standard or Storage-Optimized)
Index	Vector data structure for similarity search
Delta Sync	Auto-syncs with source Delta table
Direct Access	Manual CRUD operations on vectors

Endpoint Types

Type	Latency	Capacity	Cost	Best For
Standard	20-50ms	320M vectors (768 dim)	Higher	Real-time, low-latency
Storage-Optimized	300-500ms	1B+ vectors (768 dim)	7x lower	Large-scale, cost-sensitive

Index Types

Type	Embeddings	Sync	Use Case
Delta Sync (managed)	Databricks computes	Auto from Delta	Easiest setup
Delta Sync (self-managed)	You provide	Auto from Delta	Custom embeddings
Direct Access	You provide	Manual CRUD	Real-time updates

Quick Start

Create Endpoint

from databricks.sdk import WorkspaceClient

w = WorkspaceClient()

# Create a standard endpoint
endpoint = w.vector_search_endpoints.create_endpoint(
    name="my-vs-endpoint",
    endpoint_type="STANDARD"  # or "STORAGE_OPTIMIZED"
)
# Note: Endpoint creation is asynchronous; check status with get_endpoint()

Create Delta Sync Index (Managed Embeddings)

# Source table must have: primary key column + text column
index = w.vector_search_indexes.create_index(
    name="catalog.schema.my_index",
    endpoint_name="my-vs-endpoint",
    primary_key="id",
    index_type="DELTA_SYNC",
    delta_sync_index_spec={
        "source_table": "catalog.schema.documents",
        "embedding_source_columns": [
            {
                "name": "content",  # Text column to embed
                "embedding_model_endpoint_name": "databricks-gte-large-en"
            }
        ],
        "pipeline_type": "TRIGGERED"  # or "CONTINUOUS"
    }
)

Query Index

results = w.vector_search_indexes.query_index(
    index_name="catalog.schema.my_index",
    columns=["id", "content", "metadata"],
    query_text="What is machine learning?",
    num_results=5
)

for doc in results.result.data_array:
    score = doc[-1]  # Similarity score is last column
    print(f"Score: {score}, Content: {doc[1][:100]}...")

Common Patterns

Create Storage-Optimized Endpoint

# For large-scale, cost-effective deployments
endpoint = w.vector_search_endpoints.create_endpoint(
    name="my-storage-endpoint",
    endpoint_type="STORAGE_OPTIMIZED"
)

Delta Sync with Self-Managed Embeddings

# Source table must have: primary key + embedding vector column
index = w.vector_search_indexes.create_index(
    name="catalog.schema.my_index",
    endpoint_name="my-vs-endpoint",
    primary_key="id",
    index_type="DELTA_SYNC",
    delta_sync_index_spec={
        "source_table": "catalog.schema.documents",
        "embedding_vector_columns": [
            {
                "name": "embedding",  # Pre-computed embedding column
                "embedding_dimension": 768
            }
        ],
        "pipeline_type": "TRIGGERED"
    }
)

Direct Access Index

import json

# Create index for manual CRUD
index = w.vector_search_indexes.create_index(
    name="catalog.schema.direct_index",
    endpoint_name="my-vs-endpoint",
    primary_key="id",
    index_type="DIRECT_ACCESS",
    direct_access_index_spec={
        "embedding_vector_columns": [
            {"name": "embedding", "embedding_dimension": 768}
        ],
        "schema_json": json.dumps({
            "id": "string",
            "text": "string",
            "embedding": "array<float>",
            "metadata": "string"
        })
    }
)

# Upsert data
w.vector_search_indexes.upsert_data_vector_index(
    index_name="catalog.schema.direct_index",
    inputs_json=json.dumps([
        {"id": "1", "text": "Hello", "embedding": [0.1, 0.2, ...], "metadata": "doc1"},
        {"id": "2", "text": "World", "embedding": [0.3, 0.4, ...], "metadata": "doc2"},
    ])
)

# Delete data
w.vector_search_indexes.delete_data_vector_index(
    index_name="catalog.schema.direct_index",
    primary_keys=["1", "2"]
)

Query with Embedding Vector

# When you have pre-computed query embedding
results = w.vector_search_indexes.query_index(
    index_name="catalog.schema.my_index",
    columns=["id", "text"],
    query_vector=[0.1, 0.2, 0.3, ...],  # Your 768-dim vector
    num_results=10
)

Hybrid Search (Semantic + Keyword)

Hybrid search combines vector similarity (ANN) with BM25 keyword scoring. Use it when queries contain exact terms that must match — SKUs, error codes, proper nouns, or technical terminology — where pure semantic search might miss keyword-specific results. See references/search-modes.md for detailed guidance on choosing between ANN and hybrid search.

# Combines vector similarity with keyword matching
results = w.vector_search_indexes.query_index(
    index_name="catalog.schema.my_index",
    columns=["id", "content"],
    query_text="SPARK-12345 executor memory error",
    query_type="HYBRID",
    num_results=10
)

Filtering

Standard Endpoint Filters (Dictionary)

# filters_json uses dictionary format
results = w.vector_search_indexes.query_index(
    index_name="catalog.schema.my_index",
    columns=["id", "content"],
    query_text="machine learning",
    num_results=10,
    filters_json='{"category": "ai", "status": ["active", "pending"]}'
)

Storage-Optimized Filters (SQL-like)

Storage-Optimized endpoints use SQL-like filter syntax via the databricks-vectorsearch package's filters parameter (accepts a string):

from databricks.vector_search.client import VectorSearchClient

vsc = VectorSearchClient()
index = vsc.get_index(endpoint_name="my-storage-endpoint", index_name="catalog.schema.my_index")

# SQL-like filter syntax for storage-optimized endpoints
results = index.similarity_search(
    query_text="machine learning",
    columns=["id", "content"],
    num_results=10,
    filters="category = 'ai' AND status IN ('active', 'pending')"
)

# More filter examples
# filters="price > 100 AND price < 500"
# filters="department LIKE 'eng%'"
# filters="created_at >= '2024-01-01'"

Trigger Index Sync

# For TRIGGERED pipeline type, manually sync
w.vector_search_indexes.sync_index(
    index_name="catalog.schema.my_index"
)

Scan All Index Entries

# Retrieve all vectors (for debugging/export)
scan_result = w.vector_search_indexes.scan_index(
    index_name="catalog.schema.my_index",
    num_results=100
)

Reference Files

Topic	File	Description
Index Types	references/index-types.md	Detailed comparison of Delta Sync (managed/self-managed) vs Direct Access
End-to-End RAG	references/end-to-end-rag.md	Complete walkthrough: source table → endpoint → index → query → agent integration
Search Modes	references/search-modes.md	When to use semantic (ANN) vs hybrid search, decision guide
Operations	references/troubleshooting-and-operations.md	Monitoring, cost optimization, capacity planning, migration

CLI Quick Reference

# List endpoints
databricks vector-search-endpoints list-endpoints

# Create endpoint (positional args: NAME ENDPOINT_TYPE)
databricks vector-search-endpoints create-endpoint my-endpoint STANDARD

# List indexes on endpoint (positional arg: ENDPOINT_NAME)
databricks vector-search-indexes list-indexes my-endpoint

# Get index status (positional arg: INDEX_NAME)
databricks vector-search-indexes get-index catalog.schema.my_index

# Sync index (positional arg: INDEX_NAME)
databricks vector-search-indexes sync-index catalog.schema.my_index

# Delete index (positional arg: INDEX_NAME)
databricks vector-search-indexes delete-index catalog.schema.my_index

Common Issues

Issue	Solution
Index sync slow	Use Storage-Optimized endpoints (20x faster indexing)
Query latency high	Use Standard endpoint for <100ms latency
filters_json not working	Storage-Optimized uses SQL-like string filters via databricks-vectorsearch package's filters parameter
Embedding dimension mismatch	Ensure query and index dimensions match
Index not updating	Check pipeline_type; use sync_index() for TRIGGERED
Out of capacity	Upgrade to Storage-Optimized (1B+ vectors)
query_vector truncated	Large vectors (e.g. 1024-dim) can be truncated when serialized as JSON. Use query_text instead (for managed embedding indexes), or use the Databricks SDK to pass raw vectors

Embedding Models

Databricks provides built-in embedding models:

Model	Dimensions	Context Window	Use Case
databricks-gte-large-en	1024	8192 tokens	English text, high quality
databricks-bge-large-en	1024	512 tokens	English text, general purpose

# Use with managed embeddings
embedding_source_columns=[
    {
        "name": "content",
        "embedding_model_endpoint_name": "databricks-gte-large-en"
    }
]

Notes

• Storage-Optimized is newer — better for most use cases unless you need <100ms latency
• Delta Sync recommended — easier than Direct Access for most scenarios
• Hybrid search — available for both Delta Sync and Direct Access indexes
• columns_to_sync matters — only synced columns are available in query results; include all columns you need
• Filter syntax differs by endpoint — Standard uses dict-format filters, Storage-Optimized uses SQL-like string filters. Use the databricks-vectorsearch package's filters parameter which accepts both formats
• Management vs runtime — CLI and SDK handle lifecycle management; for agent tool-calling at runtime, use VectorSearchRetrieverTool

Related Skills

• databricks-model-serving - Deploy agents that use VectorSearchRetrieverTool
• databricks-agent-bricks - Knowledge Assistants use RAG over indexed documents