Task 3.2: Augment a prompt with additional context from a user's input based on key fields, terms, and intents; Create a prompt that adjusts an LLM's response from a baseline to a desired output

• ADefine a join relationship with that condition in the knowledge store

  Correct

  Join relationships tell Genie how tables relate so that it writes accurate JOIN statements; for complicated conditions a SQL expression can record the join.

• BAdd a column synonym "account" to the orders.account_id column

  Incorrect

  Synonyms help match user wording to a column; they do not tell Genie which columns link two tables.

• CHide the orders table so that Genie answers from the accounts table alone

  Incorrect

  Revenue lives in orders, so hiding it removes the data the question needs instead of fixing the join.

• DRaise the SQL warehouse size so Genie can try more join paths

  Incorrect

  Warehouse size changes query speed and concurrency, not which join condition Genie writes.

The Genie knowledge store holds agent-scoped metadata, synonyms, SQL expressions and join relationships. Primary and foreign keys declared in Unity Catalog are saved as join relationships automatically; without them, the author defines the relationship so that Genie joins the tables the approved way.

• APin ai_extract to version 1.0 in the options map for stable output

  Incorrect

  Version 1.0 is a legacy interface that does not support field descriptions or rich schemas.

• BRun ai_fix_grammar on the extracted values afterwards

  Incorrect

  ai_fix_grammar corrects grammatical errors in text; it does not reformat dates or look up legal names.

• CRun ai_summarize on each invoice before extracting the fields

  Incorrect

  A summary can drop the fields the job needs, and it does not tell the model which format to use.

• DAdd property descriptions, e.g. "Date in YYYY-MM-DD format"

  Correct

  Each property in an ai_extract JSON schema can carry an optional description that guides extraction, which is where a format or a meaning such as legal name belongs.

ai_extract version 2.0 and later accepts a JSON schema with types and an optional description per property. Descriptions are the lever for steering what a field means and how its value is written.

• AA text instruction that contains the approved query

  Incorrect

  Instructions guide how Genie writes SQL, but they are not trusted assets, so answers are not marked as verified.

• BA parameterized example SQL query with :store_id

  Correct

  Parameterized example queries take a value from the user's question at runtime and reuse the approved structure; in chat mode, using one gives a verified answer.

• CA column synonym that maps "store" to store_id

  Incorrect

  Synonyms help Genie pick the right column; they do not reuse an approved query or produce verified answers.

• DA benchmark question for each store number

  Incorrect

  Benchmarks are for evaluating the agent; Genie does not use them to answer users.

A parameterized example SQL query gives a verified answer when Genie uses it. A parameter turns the values users type, such as a store number, into inputs of the approved query.

• AThe name of the table that the query reads

  Incorrect

  A table name does not help Genie match how users ask questions.

• BA summary of the SQL logic, written for reviewers

  Incorrect

  Logic belongs in SQL comments; the title should mirror the user's question.

• CThe most typical phrasing of the user's question

  Correct

  Using the way users usually ask the question as the title improves Genie's ability to match prompts to the example.

• DA unique ID such as EX-001

  Incorrect

  An ID carries no wording for Genie to match against.

Give each example query its SQL and a title in the users' own words. Genie can use a matching example directly, or learn from it to structure SQL for similar questions.

• APrompts often do not transfer across models

  Correct

  Each model responds differently, so prompts should be re-tested and adapted when the model changes.

• BCheaper models ignore system prompts

  Incorrect

  Price does not decide whether a model follows system prompts.

• CPrompt templates must be retrained together with the new model

  Incorrect

  Prompt templates are text, not trained artifacts.

• DThe index must be rebuilt for each model

  Incorrect

  The chat model change does not affect the retrieval index.

Prompts that work well for one model may not for another; experiment with prompt formats and lengths and be ready to refine prompts when switching models.

• AShorten the query to its first keyword before searching

  Incorrect

  Dropping words loses meaning and does not add the missing terms.

• BExpand the query into synonym variations, searching each

  Correct

  Query expansion generates variations of the query to find documents that use different terminology, improving recall.

• CLower num_results to 1 so that only the top hit is kept

  Incorrect

  Fewer results lower recall further.

• DRewrite the manuals so that they use only one wording

  Incorrect

  Changing the source text to match one wording is impractical and loses content.

Query optimization techniques include expansion (multiple variations, deduplicated results) and reformulation (breaking down or rephrasing complex queries).

• AAdd the question as a benchmark so that Genie learns from it

  Incorrect

  Benchmarks score the agent's accuracy; they do not teach Genie how to answer.

• BPaste the corrected SQL into the general text instructions

  Incorrect

  Text instructions are for global guidance; SQL for one kind of question belongs in an example query.

• CRename the columns in the table to match the user's words

  Incorrect

  Renaming source columns is not needed; synonyms and descriptions handle wording, and the SQL logic still has to be taught.

• DClick Add as instruction to save it as an example SQL query

  Correct

  Add as instruction opens the example SQL editor filled with the question and the corrected SQL, so Genie can reuse it for related questions.

Authors with CAN EDIT can correct generated SQL, run it and save it with Add as instruction. Example SQL queries let Genie match user prompts to verified SQL and learn from them for related questions, while benchmarks only measure accuracy.