// the find
devflowinc/trieve
All-in-one platform for search, recommendations, RAG, and analytics offered via API
Trieve is a self-hostable search and RAG API platform built in Rust on actix-web, combining dense vector search (via Qdrant), neural sparse search (SPLADE), and cross-encoder reranking into a single service. It targets teams who want Algolia-tier search or a managed RAG pipeline without handing their data to a SaaS. The managed cloud tier exists, but the self-hosting story is the real pitch.
The hybrid search implementation is the genuine differentiator — dense + SPLADE sparse + optional cross-encoder reranking is a thoughtful stack, not just 'we bolted on vectors.' Sub-sentence highlighting via simsearch is a small but useful touch that most search APIs skip entirely. The worker architecture is well-separated: ingestion, file processing, deletion, and group operations each run as independent binaries, so a slow crawl job doesn't block search throughput. Helm charts with separate values files for AWS, GCP, and Azure means the Kubernetes self-hosting path is actually documented and differentiated, not an afterthought.
Local dev setup requires running at minimum 5 separate terminal processes — cargo watch, three worker binaries, and the frontend yarn dev — plus a Docker Compose stack with embedding servers. The convenience.sh abstraction leaks badly and there's no single command to get a working local instance. The dependency graph is operationally heavy for self-hosters: Qdrant, PostgreSQL, ClickHouse, Keycloak, and Redis are all required; Keycloak alone is a significant burden for teams who just want API-key auth on a small deployment. The Python SDK is OpenAPI-generator output, not hand-written — the abstractions are mechanical and the usage patterns don't feel idiomatic. Analytics requiring ClickHouse means you're maintaining a column-store just to see query latencies, which is a lot of infrastructure for what most teams need from search analytics.