@kuralle-agents/vectorize-store 0.9.0 → 0.11.0

package/README.md

@@ -20,27 +20,60 @@ Peer: `@kuralle-agents/rag`.
 
 ## Usage
 
+Pair Vectorize with **Workers AI embeddings** — the embedding model runs inside
+Cloudflare's network (`env.AI` binding), so every query embedding skips the
+public-internet round trip a cloud embedding API costs, and no provider API key
+is needed. `AiSdkEmbedder` accepts the model from
+[`workers-ai-provider`](https://www.npmjs.com/package/workers-ai-provider)
+directly:
+
 ```ts
 import { CloudflareVectorizeStore } from '@kuralle-agents/vectorize-store';
 import { AiSdkEmbedder, VectorRetriever } from '@kuralle-agents/rag';
-import { createOpenAI } from '@ai-sdk/openai';
+import { createWorkersAI } from 'workers-ai-provider';
 
 interface Env {
   VECTORIZE: VectorizeIndex;
-  OPENAI_API_KEY: string;
+  AI: Ai;
 }
 
 export default {
   async fetch(request: Request, env: Env) {
-    const openai = createOpenAI({ apiKey: env.OPENAI_API_KEY });
+    const workersai = createWorkersAI({ binding: env.AI });
     const vectorStore = new CloudflareVectorizeStore({ index: env.VECTORIZE });
-    const embedder = new AiSdkEmbedder({ model: openai.embedding('text-embedding-3-small') });
+    const embedder = new AiSdkEmbedder({
+      model: workersai.textEmbeddingModel('@cf/baai/bge-m3'),
+    });
     const retriever = new VectorRetriever({ store: vectorStore, embedder, indexName: 'docs', topK: 5 });
     // ...
   },
 };
 ```
 
+```toml
+# wrangler.toml
+[ai]
+binding = "AI"
+
+[[vectorize]]
+binding = "VECTORIZE"
+index_name = "docs"
+```
+
+The Vectorize index dimension must match the embedding model
+(`@cf/baai/bge-m3` → 1024):
+
+```bash
+npx wrangler vectorize create docs --dimensions=1024 --metric=cosine
+```
+
+Any other AI SDK embedding provider (OpenAI, Google, Cohere) plugs into
+`AiSdkEmbedder` the same way when you need a specific model — at the cost of a
+cross-internet API call per embedding and a provider key. Whichever model you
+choose, configure `RagPipeline`'s `manifest` (see `@kuralle-agents/rag`) so the
+index is locked to the model that built it — mixing embedding models in one
+index silently corrupts relevance.
+
 Use `createVectorRetrievalTool` from `@kuralle-agents/tools` to attach this retriever as an agent tool.
 
 ## Related

package/package.json

@@ -6,7 +6,7 @@
     "url": "git+https://github.com/kuralle/kuralle-agents.git",
     "directory": "packages/kuralle-vectorize-store"
   },
-  "version": "0.9.0",
+  "version": "0.11.0",
   "description": "Cloudflare Vectorize-backed VectorStore for Kuralle (edge/Workers runtime)",
   "type": "module",
   "main": "dist/index.js",
@@ -18,12 +18,12 @@
     }
   },
   "peerDependencies": {
-    "@kuralle-agents/rag": "0.9.0"
+    "@kuralle-agents/rag": "0.11.0"
   },
   "devDependencies": {
     "@types/node": "^20.11.0",
     "typescript": "^5.3.0",
-    "@kuralle-agents/rag": "0.9.0"
+    "@kuralle-agents/rag": "0.11.0"
   },
   "publishConfig": {
     "access": "public"