Vector collections
A collection is one split of one of your Hugging Face datasets, embedded into Qdrant so it can be searched by meaning. The Collections page builds them, searches them and tests how well their questions find their answers. The dataset stays on Hugging Face: a collection holds vectors computed from it at one commit, and its definition.
Build one
Open Collections → New collection, or Build a collection on a dataset's page, then pick:
| Field | What it decides |
|---|---|
| Dataset and split | The rows embedded, read at the dataset's latest commit. A split over 50 MB is refused, as in the viewer. |
| Text to embed | One to four text columns, joined by a blank line. A row empty in all of them is skipped. |
| Questions for the retrieval test | The column the test asks with. For a Q/A split, the question column. |
| Embedding model | One of your OpenAI-compatible endpoint's embedding models: OPENAI_EMBEDDING_MODEL (the default), or another it serves. Without an endpoint that embeds, no collection can be built. |
The worker then reads the split with your own token and embeds every row on your OpenAI-compatible endpoint (the model is named openai:<id>), with your key: every row is sent there. The page shows the progress; a build can be cancelled. Each row becomes a point whose id is the row's position in the split, and whose payload holds the embedded text and the row's text fields.
A collection keeps the commit it was built from. Once the dataset has a newer commit, its page says so and Rebuild embeds the new version, replacing the vectors.
Search it
Search embeds a query with the collection's model and lists the closest rows, each with its cosine similarity and a link to the row in the dataset viewer, at the commit the collection was built from.
Test it
Retrieval test samples rows of the split (50 to 500). Each one asks with its question column, and a hit is its own row coming back:
| Result | Meaning |
|---|---|
| Found first | Share of the rows whose own row is the top hit |
| In the top k | Share of the rows whose own row comes back at all |
| Mean reciprocal rank | 1 when every row comes first, 0 when none is in the top k |
The rows not found first are listed with what came first instead. A top hit with a similarity of 0.95 or more is flagged as a possible duplicate: the same pair twice in the dataset.
Compare models and columns
The sample depends only on the split and its commit. Build two collections of the same split, with another model or other columns, and they are tested on the same rows.
Embed the answers, ask with the questions
When the question column is embedded too, every question finds its own text and the test is meaningless; the page says so. Embed the answer (and its context) alone to measure retrieval.
Use it from your own code
Use it gives the collection's Qdrant name and a Python snippet that queries it with qdrant-client, embedding the query with the openai client on the same endpoint and model id as the build.
An endpoint model is named by its id
A collection built with openai:embed-default is searched with whatever your endpoint serves under embed-default at the time. If the alias moves to another model, or you switch to another endpoint, rebuild the collection.
Limits
- Qdrant must be configured: see Configuration. Without it the page says so and offers nothing to build.
- A collection built with one of the local models the server used to run (multilingual MiniLM, BGE, MiniLM L6) can no longer be searched or rebuilt: delete it and build it again with an endpoint model.
- 20 collections per user, and one build at a time.
- Builds count towards the job quotas (
QUOTA_ACTIVE_RUNS,QUOTA_RUNS_PER_DAY, see Jobs). - Deleting a collection drops its vectors; deleting your account drops all of them, those of a build still running included. The data export lists your collections, without their vectors.