Vertex AI Search ranking API
Website: https://cloud.google.com/generative-ai-app-builder/docs/ranking
Provider: Google Cloud
Category: Curated Resource Lists / AI & Machine Learning
Type: API for document reranking
Overview
Vertex AI Search ranking API is a Google Cloud service that reranks a list of documents (records) based on how well they answer a given query. It uses pretrained language models to go beyond basic semantic similarity and assign precise relevance scores, making it well-suited for improving search and Retrieval-Augmented Generation (RAG) pipelines.
The API is stateless: you don’t need to index documents beforehand. You supply a query and a set of records, and it returns the same records in a new order with relevance scores.
Use Cases
- Improve search result quality after retrieving an initial candidate set from Vector Search or another search system.
- RAG grounding: find the most relevant pieces of content to pass to an LLM as grounding context.
- Improve existing search experiences by reranking results for better relevance to user intent.
- Identify relevant sections within documents, especially for long or chunked content.
- Chunked-document workflows, e.g.:
- Split documents into chunks with Document AI Layout Parser.
- Create embeddings for each chunk.
- Load embeddings into Vector Search or another search solution.
- Query and retrieve the most relevant chunks.
- Rerank those chunks with the ranking API.
Features
Core Ranking
- Query-based reranking: Takes a natural-language query plus a list of records and returns the records ordered by relevance.
- Precise relevance scoring: Provides scores that indicate how well each record answers the query, beyond simple embedding similarity.
- Complements vector search: Intended to be used after an initial retrieval step from Vector Search or any other search backend.
Stateless Operation
- No indexing required: The API does not maintain state or indexes. You send documents with each request.
- Flexible integration: Can be applied to results from different search systems or databases without changing underlying infrastructure.
- Query input: A text query string, e.g.
"Why is the sky blue?".
- Records array:
- Each record can include:
id (unique identifier)title (optional but recommended)content (optional but recommended)
- You must provide at least one of
title or content per record.
- Record limits:
- Up to 200 records per request.
- Token limits per record depend on model version:
- Versions up to
003: up to 512 tokens per record.
- Version
004: up to 1024 tokens per record.
- If combined
title + content exceeds the model’s token limit, the extra content is truncated.
Output Control
- Configurable number of results: Optionally specify the maximum number of records to return. By default, all records are returned, reranked.
Model Versions
- Multiple model versions supported, including at least:
- Different token capacities: newer versions (e.g.,
004) support longer records.
Typical Workflow Integration
- Prepare documents
- Optionally split large documents into smaller chunks (for example, using Document AI Layout Parser).
- Initial retrieval
- Use an embeddings API to embed the chunks.
- Load embeddings into Vertex AI Vector Search or another search solution.
- Run a query to retrieve an initial set of candidate chunks or documents.
- Rerank with ranking API
- Send the query and the retrieved records to the ranking API.
- Use the returned order and scores to select the most relevant content for search display or for RAG grounding.
Pricing
The provided content does not include pricing details or plan information for the Vertex AI Search ranking API. Refer to Google Cloud’s official pricing pages for up-to-date pricing information.
