Task 1.2: Select chain components for a desired model input and output; Translate business use case goals into a description of the desired inputs and outputs for the AI pipeline

• AA filter-only query on chunk_id, with no query text, that returns the chunk's text column

  Correct

  A filter-only query omits query_text, query_vector and query_type and returns the rows that match the filter; filtering on the primary key is a point lookup.

• BA similarity search that sends the chunk_id value as query_text and keeps the top result

  Incorrect

  Similarity search ranks chunks by meaning, and an ID string carries no meaning to match, so the top result need not be that chunk.

• CA SQL query on the index's source Delta table, filtered on the chunk_id value

  Incorrect

  The service principal has no access to the source table, so this path is closed; it can only query the index.

• DThe original question searched again, with num_results raised until that same chunk comes back

  Incorrect

  A repeated search may not return the same chunk at all and wastes calls, although the chain already knows which row it needs.

For a known primary key, AI Search supports filter-only queries: omit query_text, query_vector and query_type and pass only filters, for example {"chunk_id": "c-1042"} on a standard endpoint. The caller needs SELECT on the index, and the num_results and pagination limits still apply.

• AStore doc_uri and page as chunk columns, return them from the retriever and cite them in post-processing

  Correct

  Document-level and structural metadata such as URLs and page numbers belong next to the chunks, and post-processing can add citations from what was retrieved.

• BInstruct the model to cite the file name and page that each passage most likely came from

  Incorrect

  The retrieved text carries no source fields, so the model can only guess, which is how invented file names appear.

• CList every policy URL with its title in the system prompt and ask the model to pick the one that matches each passage

  Incorrect

  A list in the prompt does not tie a chunk to its source or page, and the list grows with the corpus.

• DReturn the chunk_id values to the user so they can look up the source themselves

  Incorrect

  A chunk_id is not a link to a PDF page, so the compliance requirement is not met.

Metadata such as file names, URLs and page numbers should be stored alongside chunks. The retriever returns it with each document (for example as doc_uri in the MLflow retriever schema), and a post-processing step adds citations.

• Arequest => a CONCAT of the three values into one prompt string, with returnType => 'DOUBLE'

  Incorrect

  The endpoint expects the named input features of its signature, not a text prompt that the model would have to parse.

• Brequest => a named_struct of the three columns, plus a returnType that is required because ai_query cannot read the model schema

  Incorrect

  On Databricks Runtime 15.2 and above, returnType is optional for this endpoint: ai_query infers the return type from its model schema.

• Crequest => named_struct('tenure_months', tenure_months, 'plan', plan, 'region', region), with returnType left out

  Correct

  For a custom endpoint the request can be a STRUCT whose field names match the model's input features, and the return type is inferred from the model schema on Databricks Runtime 15.2 and above.

• Drequest => a messages array that holds the row as one user message, with returnType => 'STRING'

  Incorrect

  A messages array suits a chat model; this endpoint expects the three features of its signature and returns a number.

ai_query sends a custom model endpoint a single column or a STRUCT whose field names match the input features the endpoint expects. When the endpoint has a model schema, Databricks Runtime 15.2 and above infer the return type, so returnType can be omitted.

• AQuery preprocessing

  Incorrect

  Preprocessing prepares the retrieval query before retrieval runs.

• BRetrieval

  Incorrect

  Retrieval finds and ranks chunks; it does not build the prompt.

• CPrompt augmentation

  Correct

  Prompt augmentation forms the prompt from the query, retrieved context and instructions in a template.

• DPost-processing

  Incorrect

  Post-processing works on the LLM's response after generation.

The online RAG chain runs optional query preprocessing, retrieval, prompt augmentation, LLM generation and optional post-processing.

• AOnly the prompt template, so that the generation model sees the product name

  Incorrect

  The prompt template shapes generation; it does not narrow what retrieval returns.

• BThe generation model, which must be fine-tuned on every product name

  Incorrect

  Filtering happens in the retrieval step, so the generation model is not involved.

• CThe post-processing step, which must drop answers that mention other products

  Incorrect

  Post-processing acts after generation and does not change which chunks are retrieved.

• DMetadata extraction in the pipeline, and a retriever that applies the filter

  Correct

  Each chunk needs the product field from the data pipeline, and the retrieval step must accept and apply the extracted filter.

Filter extraction must be done together with changes to two other components: the metadata extraction step of the data pipeline, so that each chunk carries the field, and the retriever, so that it accepts and applies the extracted filters.

• AIntent, then entities, then rewrite

  Correct

  Entity extraction can depend on the intent, and the rewrite uses both, which is the order Databricks describes for a support bot.

• BRewrite, then intent, then entities

  Incorrect

  Rewriting first means the rewrite cannot use the intent and entities it needs.

• CEntities, then rewrite, then intent

  Incorrect

  Classifying intent last leaves the entity step without its guidance and the rewrite without the intent.

• DAll three in one parallel call

  Incorrect

  Parallel calls cannot feed one step's output into the next.

In the support-bot example, the chain classifies the intent, extracts entities based on that intent, then rewrites the query with both before retrieval.

• AFine-tune the generation model each week on the promotions that are current

  Incorrect

  Weekly fine-tuning is costly and still leaves the model without the date at answer time.

• BParameterize the prompt template and pass in the current date at inference time

  Correct

  A prompt template can take variables beyond the query and the retrieved data, such as the current date or user context, filled in at inference time.

• CRaise the temperature so that the model reasons more freely about the dates it sees

  Incorrect

  Temperature controls randomness; it gives the model no knowledge of the date.

• DAdd few-shot examples of last month's promotions to the prompt template

  Incorrect

  Examples show format and style, but they do not tell the model which day it is.

Prompt augmentation combines the query, the retrieved information and instructions in a template. Parameterizing that template with values such as the current date or user context makes responses context-aware.