Assembling and Deploying Applications

A GenAI engineer built a RAG chain in a Databricks notebook using LangChain. The chain references a custom Python helper module (retrieval_utils.py) that lives alongside the notebook. When logging the chain with MLflow using mlflow.langchain.log_model() by passing the in-memory chain object directly, the model deploys successfully but fails at serving time with a ModuleNotFoundError for retrieval_utils. What is the recommended way to log the model so it serves correctly?

A Generative AI Engineer builds a RAG chain locally using LangChain 0.2.x and a specific version of the databricks-vectorsearch SDK. After logging the chain with MLflow and deploying it to a Model Serving endpoint, the endpoint fails to start with import errors that never appeared during local development. The engineer confirms the code is identical. What is the most likely cause and correct fix?

A GenAI engineer builds a RAG chain in a development Databricks workspace and logs it to MLflow, registering it in Unity Catalog. The organization uses a separate production workspace for serving. The team wants to promote the exact same logged model artifact from dev to prod without re-running the training/logging notebook in production, ensuring bit-for-bit parity of the deployed chain. What is the recommended approach on Databricks?

A Generative AI Engineer has built a RAG chain that queries a Databricks Vector Search index and calls a Foundation Model API endpoint. When they log the chain with MLflow using the models-from-code approach and deploy it to a Model Serving endpoint, the endpoint starts but every inference request fails with authentication errors when the chain tries to reach the Vector Search index and the LLM endpoint. What should the engineer do to fix this before re-logging the model?

A Generative AI Engineer logs a RAG chain to MLflow and deploys it to a Model Serving endpoint. During testing, some client requests fail with schema validation errors before the model even executes, while others with slightly different payload structures succeed. The team wants a reliable way to guarantee that incoming serving requests are validated against the exact expected input structure the model was logged with, and that they can inspect what that structure is. What should the engineer do when logging the model to achieve this?

A Generative AI Engineer is packaging a RAG chain for deployment to a Databricks Model Serving endpoint. The chain queries a Databricks Vector Search index and calls a Foundation Model API endpoint at inference time. During local testing the chain works, but after deployment the served endpoint fails with authentication errors when trying to reach the Vector Search index. What is the correct way to ensure the served model can securely access these Databricks resources without embedding personal credentials?

A Generative AI Engineer has built a RAG chain and is preparing to log it with MLflow before deploying it to a Model Serving endpoint. During a test deployment, the endpoint fails schema validation and rejects incoming requests because the expected input format is ambiguous. The engineer wants to ensure the served model reliably validates and documents the expected request/response structure at deployment time. What should the engineer do when logging the model to prevent this issue?

A GenAI engineer has built a RAG chain and wants to test it against a batch of 500 evaluation questions before deploying to a Model Serving endpoint. They log the chain with MLflow using the langchain flavor. After registering the model in Unity Catalog, they want to validate that the served model will accept the same request format they plan to send from a downstream client. Which action provides the strongest guarantee that the deployed endpoint schema matches what the client will send?

A GenAI engineer has registered a RAG chain in Unity Catalog Model Registry as 'catalog.schema.support_bot' with several versions. The team wants their Model Serving endpoint to always point to the version that has passed QA validation, without editing the endpoint configuration each time a new validated version is created. What is the recommended MLflow approach to accomplish this?

A Generative AI Engineer logs a custom RAG chain as an MLflow PyFunc model and deploys it to a Model Serving endpoint. During local testing the chain worked perfectly, but the serving endpoint fails to start with an import error for a third-party reranking library that was installed in the development notebook via a cluster library. What is the most likely root cause and the correct fix?

A GenAI engineer has registered version 3 of a RAG chain in Unity Catalog under the name main.genai.support_bot. They now need to serve this specific model version behind a Databricks Model Serving endpoint using the Python SDK, ensuring the endpoint routes 100% of traffic to that version. Which approach correctly deploys the model?

A Generative AI Engineer is logging a RAG chain with MLflow before deploying it to a Model Serving endpoint. They pass an input_example when calling mlflow.langchain.log_model() but do not explicitly define a ModelSignature. When they later query the deployed endpoint, some client requests are rejected due to schema mismatch, while the engineer expected MLflow to have captured a signature automatically. What is the most accurate explanation of MLflow's behavior here and how to make the endpoint's input contract reliable?

A production Model Serving endpoint currently serves version 7 of a RAG chain registered in Unity Catalog. After deploying version 8, monitoring shows a spike in latency and malformed responses. The team wants to immediately revert live traffic to the known-good version 7 with minimal downtime, while keeping version 8 registered for later debugging. What is the most appropriate action?

A GenAI engineer has built a custom retrieval-augmented chain in Python that orchestrates a vector search retriever, a custom re-ranking function, and a foundation model call. The chain is not implemented using a supported LangChain construct — it is plain Python with several helper modules. The engineer needs to log this chain to MLflow so it can be deployed to a Model Serving endpoint that accepts JSON requests and returns text responses. Which approach correctly packages this chain for serving?

A Generative AI Engineer has logged a RAG chain as an MLflow model in a Databricks notebook. The team requires that all production models be governed centrally, support cross-workspace sharing, and be referenced by a three-level namespace when deployed to a Model Serving endpoint. Which action correctly satisfies these governance and deployment requirements?

A GenAI engineer deploys a custom PyFunc RAG chain to a Databricks Model Serving endpoint. During a product launch, the endpoint receives sustained traffic of roughly 200 concurrent requests, but users report frequent HTTP 429 (Too Many Requests) errors even though CPU utilization on the compute is low. The endpoint is configured with a fixed compute size and a maximum concurrency well below the incoming load. What is the most appropriate action to eliminate the 429 errors while handling the sustained traffic?

A GenAI engineering team runs several Model Serving endpoints across multiple business units in a shared Databricks workspace. Finance asks the team to break down monthly serving compute costs by team so each unit can be charged back accurately. What is the recommended way to enable this cost attribution for the serving endpoints?

A GenAI engineer has fine-tuned a small custom classifier (a scikit-learn model wrapped in an MLflow pyfunc) that runs on a RAG pipeline's incoming queries to route them to different downstream handlers. The classifier is lightweight, receives sporadic traffic during business hours, and has near-zero traffic overnight. The team wants to minimize serving cost while accepting that occasional first-request latency after idle periods is acceptable. Which Model Serving endpoint configuration best fits these requirements?

A Generative AI Engineer is deploying two applications to Databricks Model Serving. The first is a lightweight sklearn classifier that returns fraud probabilities, expected to receive fewer than 5 requests per second. The second is a custom fine-tuned 7B-parameter transformer model that must generate text completions at interactive latency. The engineer wants to minimize cost while meeting the performance needs of each workload. Which endpoint configuration approach should the engineer take?

A GenAI engineer has two versions of a RAG chatbot registered in Unity Catalog: version 3 (currently in production) and version 4 (a newly tuned prompt template). Before fully promoting version 4, the team wants to route 90% of live traffic to version 3 and 10% to version 4 to compare real-world quality and latency. How should the engineer configure this on a single Databricks Model Serving endpoint?