Repository navigation
Proposal: Vector Store & Search Index primitives #2406
msdbcardoso
started this conversation in
Suggestions
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Areas: Encore.ts (TypeScript), Encore.go (Go), Encore Cloud
Is your feature request related to a problem? Please describe it
AI-powered search and vector storage are becoming standard backend infrastructure. Every major cloud offers managed services for both — AWS (OpenSearch k-NN, Bedrock Knowledge Bases), GCP (Vertex AI Vector Search, Vertex AI Search), Azure (Cosmos DB vector, AI Search), and Cloudflare (Vectorize, AI Search).
I noticed there's already significant scaffolding for a
SearchIndexprimitive in the codebase — protobuf definitions, both parsers (Go + TS/Rust), infra config structs, and SDK stubs inencore.dev/ai/search. The runtime methods aren't wired up yet and the Rust infra config mapping is commented out pending protobuf codegen.There's no equivalent
VectorStoreprimitive for low-level vector operations (BYO embeddings), which would complementSearchIndexat a lower abstraction level.Describe the solution you'd like
I work on Cloudflare's AI Search team and would like to contribute the Cloudflare provider implementations to help finish what's been started:
SearchIndex (existing scaffolding) — Wire up the runtime for Cloudflare AI Search: the API client, NAPI bridge, and uncomment the Rust infra config mapping. The SDK surface already looks right.
VectorStore (new) — A lower-level primitive at
encore.dev/storage/vectorfor BYO-embeddings vector operations (upsert, query, delete with metadata filtering). Cloudflare Vectorize as the initial provider, pgvector in local Postgres for local dev.Before I start, I'd appreciate any guidance:
SearchIndexscaffolding still the intended direction, or has thinking evolved?VectorStoreshould relate to the existing pgvector support inSQLDatabase?What is the motivation / use case for changing the behavior?
Common use cases that currently require leaving Encore's infrastructure model:
These are the same kinds of backend concerns that Encore already handles declaratively for SQL, caching, object storage, and pub/sub. Finishing the search primitive and adding a vector primitive would complete the picture for AI-native applications.
All reactions