Synthetic data is artificially generated training or evaluation data. In search, foundation models can generate domain-specific queries, documents, and relevance signals when user-behavior data is scarce. section 15.2.3
A practical pipeline supplies an LLM with several related search results, asks for queries that should retrieve them, and requests a result index for each query. The generated judgments can then become query-document relevance pairs. Related documents provide niche context and improve query specificity compared with a single document. section 15.2.3
Generating judgments for minimalist-shoe searches
For results retrieved by “What are minimalist shoes?”, the prompt produces queries such as “What is the definition of a minimalist shoe?” and “Are minimalist shoes suitable for all terrains?”, paired with result indexes. A parser converts those indexes into records containing each query and its relevant document ID. section 15.2.3
Listing 15.6 A prompt that generates queries describing documents
What are some queries that should find the following documents? List at
least 5 unique queries, where these documents are better than others
in an outdoors question and answer dataset. Be concise and only
output the list of queries, and a result number in the format [n] for
the best result in the resultset. Don't print a relevance summary at
the end.
### Results:
{resultset}
Use the prompt template to request multiple concise queries and a best-result index from a supplied result set. section 15.2.3
