Recipes
Self-contained, runnable search scenarios over vector collections, grounded in real surfaces with troubleshooting tables for what actually goes wrong. These recipes are about reading — for creating collections (BYO or auto-embed), see the Pipelines recipes linked below.
What’s here
| Recipe | Key surfaces | When you need it |
|---|---|---|
| Multi-collection Search | POST /:bucket/search/vector (omit the collection list) · warnings[] · collectionStatuses[] | Search across multiple pipeline collections in one query. Covers the compatibility group key and observability. |
| Hybrid + Rerank | POST /:bucket/search/vector · the cross-group RRF merge at k=60 | What the route actually fuses, what the request fields do and don’t do, and where the verifiable evidence stops. |
| Multimodal Search | multipart POST /:bucket/search/vector · visual_embedding_index | Query by a file upload, and the content-type → modality routing rule. |
Creating collections — over in Pipelines
The two collection-creation recipes moved to Pipelines, because building a collection is a control-plane / ingest concern:
- Pipelines → Pipeline Collection — drop documents in a bucket → K3 chunks + embeds + indexes via a Scriptum template. The canonical RAG-ingest recipe.
- Pipelines → External Collection — you compute embeddings yourself and write vectors with a stock Qdrant / Pinecone client or SQL. The BYO path.
Conventions used in these recipes
- The search route has no wire client —
POST /:bucket/search/vectoris control-plane HTTP, so those examples usedodil data searchorcurl. Raw KNN by a vector you already hold is a data-plane op (SQL / Qdrant / Pinecone /dodil data vsearch), covered in the Pipelines → External Collection recipe. - Endpoints: control plane
https://api.data.dodil.io; data-plane doorsqdrant.uk-lon-1.dodil.io,pinecone.uk-lon-1.dodil.io,table.uk-lon-1.dodil.io,pg.uk-lon-1.dodil.io:5432— see Connect. - Consistent example collections:
docs(atext_embedding_indexpipeline collection),product-images(avisual_embedding_indexpipeline collection).
See also
- Vector — the vector engine overview
- Quickstart — a pure-SQL vector table + KNN in 5 minutes
- Core Concepts — the SQL data model, metrics, the search dispatch model
- Wire Compatibility — the Qdrant / Pinecone adapter surface
- API Reference → Search — the HTTP search route (mirror reading for the recipes here)
- CLI Guide —
dodil data vsearchand the collection commands - Connect & wire adapters — endpoints + credentials for the data plane