REST Resource: projects.locations.evaluations

string

Optional. The ranking expression controls the customized ranking on retrieval documents. This overrides ServingConfig.ranking_expression . The syntax and supported features depend on the rankingExpressionBackend value. If rankingExpressionBackend is not provided, it defaults to RANK_BY_EMBEDDING .

If rankingExpressionBackend is not provided or set to RANK_BY_EMBEDDING , it should be a single function or multiple functions that are joined by "+".

• rankingExpression = function, { " + ", function };

Supported functions:

• double * relevanceScore
• double * dotProduct(embedding_field_path)

Function variables:

• relevanceScore : pre-defined keywords, used for measure relevance between query and document.
• embedding_field_path : the document embedding field used with query embedding vector.
• dotProduct : embedding function between embedding_field_path and query embedding vector.

Example ranking expression:

If document has an embedding field doc_embedding, the ranking expression could be 0.5 * relevanceScore + 0.3 * dotProduct(doc_embedding) .

If rankingExpressionBackend is set to RANK_BY_FORMULA , the following expression types (and combinations of those chained using + or * operators) are supported:

• double
• signal
• log(signal)
• exp(signal)
• rr(signal, double > 0) -- reciprocal rank transformation with second argument being a denominator constant.
• isNan(signal) -- returns 0 if signal is NaN, 1 otherwise.
• fillNan(signal1, signal2 | double) -- if signal1 is NaN, returns signal2 | double, else returns signal1.

Here are a few examples of ranking formulas that use the supported ranking expression types:

• 0.2 * semanticSimilarityScore + 0.8 * log(keywordSimilarityScore) -- mostly rank by the logarithm of keywordSimilarityScore with slight semantic_smilarity_score adjustment.
• 0.2 * exp(fillNan(semanticSimilarityScore, 0)) + 0.3 * isNan(keywordSimilarityScore) -- rank by the exponent of semanticSimilarityScore filling the value with 0 if it's NaN, also add constant 0.3 adjustment to the final score if semanticSimilarityScore is NaN.
• 0.2 * rr(semanticSimilarityScore, 16) + 0.8 * rr(keywordSimilarityScore, 16) -- mostly rank by the reciprocal rank of keywordSimilarityScore with slight adjustment of reciprocal rank of semantic_smilarity_score .

The following signals are supported:

• semanticSimilarityScore : semantic similarity adjustment that is calculated using the embeddings generated by a proprietary Google model. This score determines how semantically similar a search query is to a document.
• keywordSimilarityScore : keyword match adjustment uses the Best Match 25 (BM25) ranking function. This score is calculated using a probabilistic model to estimate the probability that a document is relevant to a given query.
• relevanceScore : semantic relevance adjustment that uses a proprietary Google model to determine the meaning and intent behind a user's query in context with the content in the documents.
• pctrRank : predicted conversion rate adjustment as a rank use predicted Click-through rate (pCTR) to gauge the relevance and attractiveness of a search result from a user's perspective. A higher pCTR suggests that the result is more likely to satisfy the user's query and intent, making it a valuable signal for ranking.
• freshnessRank : freshness adjustment as a rank
• documentAge : The time in hours elapsed since the document was last updated, a floating-point number (e.g., 0.25 means 15 minutes).
• topicalityRank : topicality adjustment as a rank. uses proprietary Google model to determine the keyword-based overlap between the query and the document.
• baseRank : the default rank of the result