Task 3.1: Select Langchain/similar tools for use in a Generative AI application; Utilize MLflow and Agent Framework for developing agentic systems; Enable multi-agent systems to leverage Genie Spaces or conversational API to retrieve data

• AWrap the index in a Unity Catalog function that calls vector_search and register that function as an agent tool

  Incorrect

  Querying AI Search through Unity Catalog functions is marked deprecated on the retrieval tools page, and the vector_search SQL function is in Public Preview.

• BLoad DatabricksEmbeddings in the agent and send precomputed query vectors to the REST API

  Incorrect

  With Databricks-managed embeddings the index embeds the query text itself, so the agent has no need to compute query vectors.

• CCopy the chunk table into the agent process at startup and rank chunks with cosine similarity

  Incorrect

  This bypasses the index, its governance and its scaling, and the copy goes stale as the table changes.

• DCreate a VectorSearchRetrieverTool from databricks-langchain with the index name and a tool description

  Correct

  AI Bridge exposes an AI Search index as an agent tool through VectorSearchRetrieverTool in databricks-langchain or databricks-openai.

For new Python agents, Databricks recommends AI Bridge: VectorSearchRetrieverTool turns an index into a tool with a name and description that tell the model when to search.

• AA Delta Sync index with Databricks-managed embeddings

  Incorrect

  Managed embeddings let the index embed the query text itself, so only the index name and description are needed.

• BAny index served from a storage-optimized endpoint

  Incorrect

  The endpoint type does not decide this; the embedding option of the index does.

• CAny index that the agent queries with HYBRID search

  Incorrect

  The query type does not change who computes embeddings for the query.

• DA Direct Vector Access index

  Correct

  With a direct-access index the tool cannot embed query text on its own, so text_column and an embedding model are required.

• EA Delta Sync index with self-managed embeddings

  Correct

  Self-managed embeddings mean Databricks does not hold the embedding model for the index, so the tool must be told which one to use.

For direct-access indexes or Delta Sync indexes with self-managed embeddings, configure VectorSearchRetrieverTool with a text column and the same embedding model that created the index vectors.

• AKeep one long-lived conversation under a service principal and send every user's question into it

  Incorrect

  Reusing threads across sessions lowers accuracy through unintended context, and a shared service principal does not apply each user's permissions.

• BStart a Chat mode conversation per user session with the user's OAuth credentials and send follow-ups as messages in it

  Correct

  Chat mode conversations keep history for follow-ups, and Genie evaluates data access with the caller's own Unity Catalog permissions.

• CStart a new conversation for every message so that no earlier context can leak between turns

  Incorrect

  A new conversation per message drops the history that follow-up questions depend on.

• DLet the agent write SQL against the sales tables directly and skip the Genie Agent

  Incorrect

  This throws away the curated instructions, example SQL and semantics that make the Genie Agent reliable for these questions.

The Genie Agents API provides Chat mode APIs for stateful conversations with follow-ups. Databricks advises a new conversation for each session, OAuth for users when a user is present, and evaluates data access with the caller's permissions.

• AKeep Model Serving and turn on express deployments

  Incorrect

  Express deployments speed up endpoint deployment, but the agent stays on Model Serving, which Databricks now lists as the legacy runtime for agents.

• BKeep Model Serving and enable route optimization

  Incorrect

  Route optimization lowers overhead latency and raises throughput for an endpoint; it does not change where or how the agent runs.

• CMove the agent's server code to run on Databricks Apps

  Correct

  Databricks lists Model Serving as the legacy runtime for agents and runs new agents on Databricks Apps, where the agent server can also be your own server with custom endpoints.

• DMove the agent behind a provisioned throughput endpoint

  Incorrect

  Provisioned throughput serves foundation models with dedicated capacity; it is not a host for custom agent code.

A deployed agent has three layers: the framework, an agent server and an agent runtime. Databricks runs new agents on Databricks Apps and marks agents on Model Serving endpoints, including Custom Agent on Model Serving, as legacy offerings to maintain and migrate rather than build on.

• AA model signature on the chain

  Incorrect

  Signatures describe model inputs and outputs for registration; they do not enable tracing.

• BA provisioned throughput endpoint

  Incorrect

  The serving mode of the LLM does not affect whether MLflow captures traces.

• CAn explicit mlflow.langchain.autolog() call

  Correct

  On serverless compute clusters autologging is not enabled automatically, so tracing for LangChain must be turned on explicitly.

• DA call to mlflow.autolog(disable=True) at the top of the notebook

  Incorrect

  That disables autologging, the opposite of what is needed.

mlflow.langchain.autolog() enables automatic tracing for LangChain, logging nested traces to the active experiment. It must be called explicitly on serverless compute, and autolog(disable=True) turns it off.

• APoll in a tight loop with no delay until a result arrives

  Incorrect

  Tight loops waste requests; backoff keeps fast queries fast without hammering the API.

• BPoll once after a fixed 60 seconds and return what is there

  Incorrect

  A fixed wait slows quick answers and can still be too early for long ones.

• CNever poll; register a webhook and wait for its callback

  Incorrect

  The documented pattern is polling until a conclusive status such as COMPLETED or FAILED.

• DPoll every 1-5 s, back off to a minute, stop near 10 minutes

  Correct

  Databricks recommends polling every 1 to 5 seconds with exponential backoff up to one minute and limiting polling to 10 minutes for most queries.

Poll until a conclusive message status (COMPLETED, FAILED or CANCELLED), using exponential backoff, and return a timeout or ask the user to check later if nothing arrives within about 10 minutes.

• AA 1,000 message limit per workspace

  Incorrect

  No such workspace-wide message limit is described.

• BThe 200,000 conversation limit

  Correct

  Databricks advises managing old conversations to stay under the 200,000 conversation limit.

• CA 24-hour lifetime per conversation

  Incorrect

  Conversations are not documented as expiring after a day.

• DA 10-table limit per conversation

  Incorrect

  Table limits concern the agent's configuration, not conversations.

Use the endpoint that lists a Genie Agent's conversations to find old ones, and delete them to stay under the 200,000 conversation limit.

• AChatDatabricks

  Correct

  Databricks examples build LangGraph agents with ChatDatabricks pointing at a Databricks model endpoint.

• BDatabricksEmbeddings

  Incorrect

  This class produces embeddings, not chat responses.

• CDatabricksOpenAI

  Incorrect

  This is the OpenAI-compatible client from databricks-openai, not a LangChain chat model class.

• DGenieAgent

  Incorrect

  This helper wraps a Genie Agent as a tool, not the agent's own model.

databricks-langchain provides ChatDatabricks for chat models and DatabricksEmbeddings for embeddings, among other integrations; databricks-openai offers an OpenAI-compatible client.