Search

For pre-filtering documents, using tags_v2 is preferred to using tags (which is now deprecated). If both tags_v2 and tags are specified, tags is ignored. tags_v2 enables building complex filters through the use of “AND”, “OR”, and negation logic. Take the below input as an example:

{
    "OR": [
        {
            "key": "subject",
            "value": "holy-bible",
            "negate": false
        },
        {
            "key": "person-of-interest",
            "value": "jesus christ",
            "negate": false
        },
        {
            "key": "genre",
            "value": "religion",
            "negate": true
        }
        {
            "AND": [
                {
                    "key": "subject",
                    "value": "tao-te-ching",
                    "negate": false
                },
                {
                    "key": "author",
                    "value": "lao-tzu",
                    "negate": false
                }
            ]
        }
    ]
}

In this case, files will be filtered such that:

1. “subject” = “holy-bible” OR
2. “person-of-interest” = “jesus christ” OR
3. “genre” != “religion” OR
4. “subject” = “tao-te-ching” AND “author” = “lao-tzu”

Note that the top level of the query must be either an “OR” or “AND” array. Currently, nesting is limited to 3. For tag blocks (those with “key”, “value”, and “negate” keys), the following typing rules apply:

1. “key” isn’t optional and must be a string
2. “value” isn’t optional and can be any or list[any]
3. “negate” is optional and must be true or false. If present and true, then the filter block is negated in the resulting query. It is false by default.

When querying embeddings, you can optionally specify the media_type parameter in your request. By default (if not set), it is equal to “TEXT”. This means that the query will be performed over files that have been parsed as text (for now, this covers all files except image files). If it is equal to “IMAGE”, the query will be performed over image files (for now, .jpg and .png files). You can think of this field as an additional filter on top of any filters set in file_ids and

When hybrid_search is set to true, a combination of keyword search and semantic search are used to rank and select candidate embeddings during information retrieval. By default, these search methods are weighted equally during the ranking process. To adjust the weight (or “importance”) of each search method, you can use the hybrid_search_tuning_parameters property. The description for the different tuning parameters are:

• weight_a: weight to assign to semantic search
• weight_b: weight to assign to keyword search

You must ensure that sum(weight_a, weight_b,..., weight_n) for all n weights is equal to 1. The equality has an error tolerance of 0.001 to account for possible floating point issues.

In order to use hybrid search for a customer across a set of documents, two flags need to be enabled:

1. Use the /modify_user_configuration endpoint to to enable sparse_vectors for the customer. The payload body for this request is below:

{
  "configuration_key_name": "sparse_vectors",
  "value": {
    "enabled": true
  }
}

2. Make sure hybrid search is enabled for the documents across which you want to perform the search. For the /uploadfile endpoint, this can be done by setting the following query parameter: generate_sparse_vectors=true

Carbon supports multiple models for use in generating embeddings for files. For images, we support Vertex AI’s multimodal model; for text, we support OpenAI’s text-embedding-ada-002 and Cohere’s embed-multilingual-v3.0. The model can be specified via the embedding_model parameter (in the POST body for /embeddings, and a query parameter in /uploadfile). If no model is supplied, the text-embedding-ada-002 is used by default. When performing embedding queries, embeddings from files that used the specified model will be considered in the query. For example, if files A and B have embeddings generated with OPENAI, and files C and D have embeddings generated with COHERE_MULTILINGUAL_V3, then by default, queries will only consider files A and B. If COHERE_MULTILINGUAL_V3 is specified as the embedding_model in /embeddings, then only files C and D will be considered. Make sure that the set of all files you want considered for a query have embeddings generated via the same model. For now, do not set VERTEX_MULTIMODAL as an embedding_model. This model is used automatically by Carbon when it detects an image file.