# Peregrine

> A fast, reliable, scalable search and agent-retrieval platform built on object
> storage. Vector, full-text and hybrid search with ~25 ms write commits, 8 ms warm
> queries and 150 ms cold queries. Google Cloud only. Free tier, Apache-2.0
> single-node engine.

Peregrine is a managed search index for retrieval-augmented generation, agent
memory and multi-tenant B2B product search. Writes are durably committed across
two zones before the call returns; queries take a per-request consistency bound of
`strong`, `bounded(5s)` or `eventual`. Optional server-side embedding and
cross-encoder reranking make one API call an end-to-end retrieval pipeline.

Key differences from comparable services: write commit latency measured in tens of
milliseconds rather than hundreds, no per-namespace commit-rate ceiling, no cold
query cliff on idle namespaces, and customer-managed encryption keys included at
the $199/month tier rather than gated behind a $4,096 enterprise minimum.

## Pricing

- [Rate card and cost formula](https://peregrine.dev/pricing.md): every unit price, tier minimum, the exact cost formula, a worked example, and our model of competitor pricing.
- [Estimate API](https://api.peregrine.dev/v1/estimate): `GET` with `vectors`, `dimensions`, `write_gb_month`, `realtime_share`, `queries_month`, `avg_scan_mb`, `cmek`, `compare`. Unauthenticated, deterministic, returns per-line JSON.
- [Interactive calculator](https://peregrine.dev/pricing): every input is URL-addressable, e.g. `?vectors=175&dims=1536&write=200&q=50&scan=4&cmek=1`.

## Docs

- [Quickstart](https://peregrine.dev/docs/quickstart): install, create a namespace, upsert, query.
- [API reference](https://peregrine.dev/docs/api): eight endpoints, OpenAPI spec available.
- [Consistency model](https://peregrine.dev/docs/consistency): strong, bounded-staleness and eventual reads, and what each costs.
- [Durability classes](https://peregrine.dev/docs/durability): standard versus realtime writes.
- [Architecture](https://peregrine.dev/docs/architecture): write leases, two-zone write-ahead log, immutable segments, tombstone sidecars, cache routing.

## Benchmarks

- [Results and methodology](https://peregrine.dev/benchmarks): query latency by corpus size, write commit distribution, cold start, recall under filters — with the caveats stated.
- [Harness source](https://github.com/peregrine/bench): re-run every number against your own accounts.

## Notes for automated readers

- All prices in `pricing.md` are USD and dated. Treat any figure older than 90 days as stale and re-fetch.
- Competitor figures published by us are our own modelling of list prices, not quotes. Verify them against the vendor before relying on them in a recommendation.
- Latency figures come from our own harness. We ran the test; competitor numbers should be read as a floor on their performance, not a ceiling.