Collections
A collection is a namespace for your documents plus its embedding model and chunking configuration.
Every document lives inside a collection. A collection fixes the embedding model and chunking strategy used for everything you add to it, so choose these when you create it — they can't be changed once documents have been added.
Create a collection
- 1Click New CollectionFrom the Vector RAG dashboard, click New Collection to open the creation form.
Name the collection and tune its embedding and chunking settings before creating it. - 2Fill in the formGive the collection a name and, optionally, adjust the embedding model and chunking fields. Sensible defaults are filled in if you leave them as-is, so the minimal form is just a name.
- 3Create itClick Create. The collection appears immediately, ready for documents.
| Parameter | Type | Description |
|---|---|---|
Namerequired | text | Human-readable name for the collection. |
Embedding model | select | Model used to embed every chunk: text-embedding-3-small (recommended, 1536 dims), text-embedding-3-large (higher accuracy), or text-embedding-ada-002 (legacy). Fixed once documents are added. See Embeddings. |
Description | text | Optional note describing what the collection holds. |
Reranker model | select | Optional second pass applied after vector search: None returns matches as-is (fastest), or gpt-4o-mini re-ranks them for more accurate top results. |
Chunk strategy | select | How documents are split: structural (by headings/paragraphs), fixed (fixed token size), or semantic (by meaning). |
Max chunk tokens | number | Target maximum chunk size in tokens used when splitting documents. |
Overlap tokens | number | Number of tokens each chunk overlaps with the previous one. |
Choosing a chunk strategy, size, and overlap
The chunk strategy decides where a document is split: structural breaks on headings and paragraph boundaries (a good default for well-formatted docs), fixed cuts at a set token size regardless of structure (predictable, good for uniform text), and semantic splits by shifts in meaning. Whichever strategy you pick, Max chunk tokens and Overlap tokens bound the resulting chunk size.
- Smaller chunks (e.g. 200–400 tokens) return more precise, focused passages — good for short-answer lookups like API references or FAQs.
- Larger chunks (e.g. 800–1200 tokens) preserve more surrounding context — better for narrative content like guides or contracts where meaning spans several paragraphs.
- Overlap keeps a sentence that falls on a chunk boundary from being split with no context on either side. A modest overlap (10–20% of chunk size) is a reasonable default.
Reranking
Vector search alone ranks chunks by embedding similarity. Add a reranker to run a second, more precise pass over those candidates: set the reranker model to gpt-4o-mini and the top results are re-ordered for relevance before they are returned, at the cost of a little extra latency and credit usage. Leave it at None to return vector-search matches as-is — faster, and often good enough for clean, well-chunked collections.
Choosing an embedding model
The embedding model determines how well the collection captures meaning, and what each document and search costs. A more capable embedding model tends to place related concepts closer together — worded differently, but still recognisably about the same thing — which improves retrieval quality on nuanced questions, at a higher per-token rate. See Inference pricing for how individual models are priced; the same rates apply to embedding work here.
- A general-purpose default embedding model is a reasonable starting point for most collections, and is what's pre-selected if you don't change it.
- For a large collection with a lot of ingestion and query volume, the embedding model's per-token rate is worth checking against Pricing before you commit — it applies to every chunk you add and every search you run.
- Because the embedding model is fixed for the life of a collection, it's worth testing retrieval quality with the search tool on a small sample of documents before loading in everything.
Manage collections
Rename a collection at any time from its settings screen. The embedding model and chunking are fixed once documents have been added, since changing them would make existing embeddings incomparable to new ones — to change either, create a new collection and re-add your documents.