AI Search & Agents
Databricks · Machine Learning & AI
Theory
AI Search (renamed from Vector Search in 2026) stores embeddings of your text so you can find content by meaning, not exact keywords. You create an endpoint, then an index on a Delta table: a Delta Sync index keeps itself up to date as the source table changes and can compute embeddings for you with an embedding model endpoint. Queries return the most similar rows from Python, REST or SQL (vector_search()). This is the retrieval part of RAG (retrieval-augmented generation): an LLM answers using the retrieved passages. Agent Bricks is Databricks' platform for building, deploying and governing such agents in code, with MLflow Tracing and evaluation; for low-code agents Databricks now points to Genie Agents (Knowledge Assistant is legacy).
Example: a support agent searches an index of product FAQs and answers "How long is delivery to Jaipur?" citing the matching FAQ rows.
Why we use it
LLMs do not know your company's documents and can make things up. Retrieval from a governed index grounds answers in your own data, and an agent platform adds evaluation, tracing and governed deployment so the result is reliable enough for users.
When to use it
Use AI Search when users search or ask questions over large text collections (docs, tickets, product descriptions). Use Genie Agents for low-code Q&A, especially over structured tables, and code-first Agent Bricks agents for complex, multi-step workflows.
How it works
- Documents or FAQ rows in a Delta table
- Chunk text and create a Delta Sync index on an AI Search endpoint
- User asks a question; embed it and search the index
- Pass the top matches plus the question to an LLM
- Is the answer grounded and correct in evaluation? Yes: Deploy the agent (Model Serving / Agent Bricks). No: Improve chunks, prompts or retrieval and re-evaluate.
- Trace and monitor real conversations
Syntax
Python: from databricks.vector_search.client import VectorSearchClient
vsc.create_delta_sync_index(endpoint_name, index_name, source_table_name, primary_key,
pipeline_type="TRIGGERED", embedding_source_column, embedding_model_endpoint_name)
index.similarity_search(query_text="...", columns=[...], num_results=5)
SQL: SELECT * FROM vector_search(index => '<cat.sch.index>', query_text => '...', num_results => 5)
UI: Catalog → source table → Create → AI Search index; Agents (sidebar) for Agent BricksExample
from databricks.vector_search.client import VectorSearchClient
vsc = VectorSearchClient()
index = vsc.create_delta_sync_index(
endpoint_name="techtown-search",
index_name="main.sales.faq_index",
source_table_name="main.sales.faq", # faq_id, question, answer (CDF enabled)
primary_key="faq_id",
pipeline_type="TRIGGERED",
embedding_source_column="answer",
embedding_model_endpoint_name="databricks-gte-large-en")
hits = index.similarity_search(query_text="delivery time to Jaipur",
columns=["faq_id", "question", "answer"], num_results=3)
-- Same search from SQL
SELECT faq_id, question
FROM vector_search(index => 'main.sales.faq_index',
query_text => 'delivery time to Jaipur', num_results => 3);Key points
- Former name: Vector Search; the Python client and vector_search() SQL function keep the old name
- Delta Sync indexes stay in sync with the source table (enable change data feed)
- Indexes are Unity Catalog objects; grant SELECT to readers
- Hybrid search mixes keyword and semantic matching for better recall
- RAG = retrieve relevant chunks, then pass them to the LLM in the prompt
- Agent Bricks and MLflow Tracing help evaluate and monitor agents before users see them