Problem description
We created three data stores and an AI application (of type Custom search (general)) that uses them. The data stores contain many Python code snippets demonstrating how our SDK is used.
We interact with the AI application using the projects.locations.dataStores.servingConfigs.answer REST API.
Our intended use case is to allow customers to submit their existing Python code and ask questions about it. For example:
- How can I add authentication to
CompilerServicefrom your SDK? Here is my current code:<LONG_CODE_SNIPPET>. - What’s wrong with my current implementation? I use the
invoke_ctrl_asyncmethod from your SDK:<LONG_CODE_SNIPPET>.
In practice, these code snippets can easily make the query longer than 2000 characters.
However, we discovered that if query.text exceeds 2000 characters, the answer API returns 400 INVALID_ARGUMENT.
Request:
curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
"https://eu-discoveryengine.googleapis.com/v1/projects/test-project-123456/locations/eu/collections/default_collection/engines/test-application_1234567898765/servingConfigs/default_serving_config:answer" \
-d '
{
"query": {
"text": "Lorem ipsum dolor sit amet, consectetur adipiscing elit. Integer efficitur urna id pulvinar condimentum. Morbi non turpis purus. Vestibulum lorem risus, vulputate et erat id, sagittis tempus erat. Phasellus rhoncus turpis sit amet diam tincidunt imperdiet. Nulla cursus tellus non ultrices maximus. Aliquam vulputate gravida est, id vestibulum est volutpat id. Vestibulum commodo est nec ligula iaculis, at rhoncus dolor tempor. Nullam congue pulvinar vestibulum. Mauris a purus nec risus tempor porta. Sed ipsum mauris, tempor vel orci non, euismod euismod risus. Duis sit amet nunc vel velit volutpat feugiat. Nulla tincidunt sed lacus id imperdiet. Aliquam convallis justo ac eros congue, ac sodales enim hendrerit. Fusce efficitur dui sed turpis aliquam, id finibus elit fermentum. Morbi quis accumsan metus. Aliquam elementum suscipit finibus. Nam blandit nibh et nunc maximus, eu euismod mauris viverra. Maecenas nec urna massa. Phasellus congue ante maximus est sodales fermentum. Maecenas dui sapien, condimentum eu venenatis id, vestibulum dapibLorem ipsum dolor sit amet, consectetur adipiscing elit. Vestibulum dapibus diam vitae congue lobortis. Mauris a tincidunt urna. Sed justo ipsum, eleifend ac dui vehicula, mollis bibendum dui. Donec eget enim eu ipsum tempor venenatis in non enim. Nam tincidunt dui blandit augue dictum, vitae tempus libero dignissim. Mauris gravida euismod magna ut dictum. Morbi vel risus in lacus blandit pretium. Sed non ex vel lectus mollis lobortis. Ut nec massa diam. Ut sit amet urna vel urna fermentum sodales quis nec erat. Etiam metus dui, pharetra in consectetur ornare, interdum vel tellus. Nulla vestibulum, erat ut aliquet pulvinar, libero nunc scelerisque risus, in lobortis nibh lectus vitae elit. Pellentesque habitant morbi tristique senectus et netus et malesuada fames ac turpis egestas. Nunc dignissim a metus id rutrum. Curabitur tristique urna quis faucibus tincidunt. Proin dui arcu, gravida convallis orci ac, imperdiet molestie nisl abc"
}
}
'
Response:
{
"error": {
"code": 400,
"message": "Request contains an invalid argument: the max query length should be \u003c= 2000 characters.",
"status": "INVALID_ARGUMENT"
}
}
Questions
1. Where is the 2000 character limit documented?
The Query API documentation does not appear to specify a maximum length for the query.text field.
Could you point us to the documentation that defines this 2000-character limit?
2. What is the recommended approach for this use case?
Our use case is to answer questions about customer-provided source code, including diagnosing an existing implementation and explaining how to modify it. Therefore, it is expected that the query will sometimes contain several thousand characters of code.
A 2000-character limit makes this use case difficult because truncating or arbitrarily splitting the customer’s code may remove the context required to answer the question correctly.
Is there a recommended way to provide longer code snippets to the answer API while preserving them as part of the user’s query/context?
More generally, what is the recommended architecture for using the answer API when the user’s query includes a potentially long question, exceeding 2000 characters?