For AI agents: this page is also available as Markdown at /docs/faq.md. Every docs page is: add .md to its URL. llms.txt lists them all.
FAQ
#The basics
What is WarpSearch? A serverless search service for full-text, vector and hybrid search over JSON documents, with facets and analytics. There is no cluster to size or operate. Start with the Quickstart or Concepts.
How is it different from Elasticsearch, OpenSearch or Solr? You do not run or tune anything: no nodes, shards or replicas, and nothing billed for compute while you are not querying. It also speaks their query languages, so you can keep your clients. See Migrate.
Can I keep my existing client? Usually yes. Elasticsearch, OpenSearch and Solr clients work against the compatible endpoints, and each page lists what is and is not supported: Elasticsearch, OpenSearch, Solr.
Which query languages can I use? SQL, Elasticsearch/OpenSearch Query DSL, Lucene, Solr, LogQL, Grafana and the native language. See Query languages.
#Vectors and hybrid search
Do I need to run embeddings? You compute the embeddings and send the vectors. WarpSearch does not generate them for you today. Use the same model for documents and queries. See Vector search.
Which embedding models and dimensions work? Any model, since you supply the vectors. The only requirement is that every vector in a field has the dimension you declared. Fields support up to 8192 dimensions.
Does it support Matryoshka embeddings? Yes. You can search with fewer dimensions and rescore at full precision. See Matryoshka and dimensions.
Can a document have several vectors? Yes, with multi-vector fields. See Multi-vector fields.
Coming soon: late-interaction (MaxSim) scoring, sparse vectors and float16 or int8 vector storage are in development.
Can I combine keyword and vector search? Yes. Send a text query and a vector search in one request and the scores are combined. See Hybrid search.
Can I filter while doing a vector search? Yes. Filters are applied inside the vector search, so you still get the number of results you asked for. See Vector search.
Is it HNSW? No. Vectors use an inverted-file (IVF) index with compact quantized codes (RaBitQ or product quantization). The index is memory-mapped, and the best candidates are rescored at full precision.
How good is recall? You control the trade-off with the number of candidates and the search effort, and you can run an exact search to measure recall against ground truth. See Vector search.
#Writing and reading data
Are writes searchable immediately?
After a commit. Writes become searchable within seconds on their own, or pass commit=true on
the ingest request to wait for it. See Consistency and durability.
How do updates and deletes work? Indexing a document with an existing id replaces it. You can delete by id, one at a time or in batches. See Ingesting documents.
Can I update part of a document? Not through the native API: send the whole document again. The Elasticsearch-compatible bulk endpoint accepts updates with upsert; see Elasticsearch compatibility.
Can I delete by query? No. Look up the ids you want to remove with a search, then delete them by id.
Not supported: delete by query.
How do I page through results?
With limit and offset, up to the result window of your plan. There are no cursors or scroll
yet. See Searching and Limits and errors.
Can I export all my data? There is no bulk export endpoint today. Keep the source of truth in your own system, and page through results within the result window for smaller extracts.
Is there highlighting?
Coming soon: highlighting is in development.
#Logs, dashboards and agents
Can I use it for logs? Yes, as documents in a collection, queried with SQL, Lucene or the Elasticsearch DSL. LogQL label selectors work too, but LogQL line filters and metric queries need a collection with a configured message and time field, which you cannot set yourself yet. See Loki and LogQL.
Does it work with Grafana? Yes, through Grafana's Elasticsearch or OpenSearch data source. The Loki data source is not supported, because WarpSearch does not serve the Loki HTTP API. See Query languages.
Can an AI agent use it? Yes. WarpSearch has an MCP server that can also sign your user up and get an API key with their approval. See the Guide for AI agents.
#Operations and security
What happens to idle collections?
A collection nobody has used for a while is put to sleep, and its data stays safe. The first
request afterwards may wait while it wakes, and can answer 503 with a Retry-After header; retry
after that long. Keep-warm avoids this; see Pricing and cost control.
Where does my data live and is it durable? Your data is kept in durable object storage, with snapshots taken after commits. See Consistency and durability.
Is my data isolated from other customers? Yes. Every request is tied to your account, and a key can only reach the collections it holds a role on. Collections you cannot access look like collections that do not exist.
How do API keys and roles work?
Keys (ws_live_...) carry the role of the person or agent that created them: viewer searches,
editor also writes, admin also manages collections. See Access control.
What are the limits? Limits cover request size, hits per call, facets, query complexity and rate. Your plan sets them. See Limits and errors.
#Pricing
How much does it cost? You pay for storage, writes, requests and returned data. There is a permanent free tier. See Pricing and cost control and the pricing page.