What is samesake

samesake is a search engine compiler for visual commerce, starting with fashion. You declare your catalog and retrieval spaces in TypeScript; samesake compiles them into a Postgres-backed search layer you run inside your own app — no separate search cluster, no hosted vector database.

It is built for shoppers who do not know the product name: screenshots, similar-look search, vague intent, budget constraints, occasion, size, availability, and merchant ranking policy.

What makes it different

It is not a hosted vector DB, a generic RAG framework, or keyword search alone. It is a typed retrieval layer for commerce catalogs where:

• image similarity, text intent, structured attributes, price, freshness, and availability are separate signals;
• hard filters stay hard — “under 20000” and “available now” are typed constraint predicates pushed into SQL, not soft semantic vibes;
• query-time weights tune visual, intent, price, and freshness influence without reindexing;
• every search emits an auditable constraint trace (/search/explain shows per-leg ranks and space cosines);
• the same factory also does entity resolution and deduplication for catalog/customer records.

How a search is composed

catalog.ts                      collection() + entity() declarations
      │
      ▼
createMatcher({ embed, generate?, ... })
      ├── schema generation     →  per-project search tables (fts, vector, filter columns)
      ├── ingest / enrich / index →  connectors, enrichment pipeline, embeddings
      ├── search / facets / nlq →  hybrid RRF retrieval (keyword + cosine + optional spaces)
      └── match / dedup / explain →  entity resolution
      │
      ▼
Postgres (pgvector + pg_trgm + unaccent + fuzzystrmatch)

Three retrieval signals combine through reciprocal-rank fusion (RRF):

• FTS — Postgres full-text keyword matching.
• Cosine ANN — vector similarity over embeddings you supply (bring your own model).
• Spaces (optional) — typed segmented vectors (style / price / freshness / category / image) with query-time weights. Off by default; enable behind your own eval gate.

Hard filters (price ≤ X, available = true, colors ∋ red) compile to SQL predicates that gate the result set before ranking.

Three ways to call it

createMatcher(config) returns one object you can call three ways:

Surface	Use when
In-process — matcher.search(...)	hot paths inside your app; no HTTP overhead
Web-standard — matcher.fetch(request)	Bun.serve, Cloudflare Workers, Vercel, Deno
Composable — matcher.app (Hono)	mount at /v1 inside an existing Hono service

The stack

Layer	Choice
Runtime	Bun 1.3+ (Node supported)
HTTP	Hono — universal fetch handler
Database	Postgres 15+ with pgvector, pg_trgm, unaccent, fuzzystrmatch
Validation	Zod
AI	Bring your own — you supply embed and optional generate / parse

Note

Two containers in production: Postgres and your app process. No Redis, no Elasticsearch, no LanceDB.

Packages: @samesake/core (SDK), @samesake/server, @samesake/cli.

Next: build a search experience in plain language, or jump to the Quickstart.