Search Over Groups
This route allows you to get groups as results instead of chunks. Each group returned will have the matching chunks sorted by similarity within the group. This is useful for when you want to get groups of chunks which are similar to the search query. If choosing hybrid search, the top chunk of each group will be re-ranked using scores from a cross encoder model. Compatible with semantic, fulltext, or hybrid search modes.
Authorizations
Headers
The dataset id or tracking_id to use for the request. We assume you intend to use an id if the value is a valid uuid.
The API version to use for this request. Defaults to V2 for orgs created after July 12, 2024 and V1 otherwise.
V1, V2 Body
Filters is a JSON object which can be used to filter chunks. This is useful for when you want to filter chunks by arbitrary metadata. Unlike with tag filtering, there is a performance hit for filtering on metadata.
Get total page count for the query accounting for the applied filters. Defaults to false, but can be set to true when the latency penalty is acceptable (typically 50-200ms).
Group_size is the number of chunks to fetch for each group. The default is 3. If a group has less than group_size chunks, all chunks will be returned. If this is set to a large number, we recommend setting slim_chunks to true to avoid returning the content and chunk_html of the chunks so as to lower the amount of time required for content download and serialization.
Highlight Options lets you specify different methods to highlight the chunks in the result set. If not specified, this defaults to the score of the chunks.
Page of group results to fetch. Page is 1-indexed.
Page size is the number of group results to fetch. The default is 10.
Query is the search query. This can be any string. The query will be used to create an embedding vector and/or SPLADE vector which will be used to find the result set. You can either provide one query, or multiple with weights. Multi-query only works with Semantic Search and is not compatible with cross encoder re-ranking or highlights.
If true, stop words (specified in server/src/stop-words.txt in the git repo) will be removed. Queries that are entirely stop words will be preserved.
Set score_threshold to a float to filter out chunks with a score below the threshold. This threshold applies before weight and bias modifications. If not specified, this defaults to 0.0.
fulltext, semantic, hybrid, bm25 Set slim_chunks to true to avoid returning the content and chunk_html of the chunks. This is useful for when you want to reduce amount of data over the wire for latency improvement (typicall 10-50ms). Default is false.
Sort Options lets you specify different methods to rerank the chunks in the result set. If not specified, this defaults to the score of the chunks.
Typo Options lets you specify different methods to correct typos in the query. If not specified, typos will not be corrected.
If true, quoted and - prefixed words will be parsed from the queries and used as required and negated words respectively. Default is false.
The user_id is the id of the user who is making the request. This is used to track user interactions with the search results.