Overview

Queries are the primary way to access information in a capsule. They are designed to be easy for LLMs to write. Every query has three parts:

• query: natural language query, like “What is the capital of France?”
• wishlist: metadata filters, find out more about wishlists
• limit: the maximum number of results to return

A simple query request:

{
    "query": "What is the capital of France?",
    "wishlist": {}, // see article on wishlists
    "limit": 10
}

​

How SID runs queries

The easiest type of query is a semantic query, but for most production systems, many queries will be more complex. For example, “give me the ten newest documents on token selection for pre-training” cannot be answered by a semantic query. You might think this is an edge case, but over our production traffic, they can easily make up 25% of all queries.

To solve this, SID uses multiple query engines under the hood, and we dynamically switch between them. A single request may be parsed into multiple subqueries.

​

What a query returns

A query will return a list of objects that are subsections of the original data (aka. chunks). The response will have at most as many items as specified in the limit parameter.

[
    {
        "item_id": "401ebb14-d0a1-4244-b720-7f4fcdb02230", // can be used to delete the item
        "idx": 8, // index of the item in the original document/data
        "content": "The treaty defines...",
        "uri": "https://example.com/document/url/for/displaying", // link to the original document
        "kind": "file",
        "file_name": "tax-treaty-table-2.pdf",
        "file_type": "application/pdf",
        "time_added": "2024-05-22T23:27:56.609683Z",
        "time_authored": "2024-01-13T13:24:11.987324Z",
        "score": 0.5021019,
        "metadata": null
    },
    {...}
]

For more details, please see the API reference.