@eltjs/embedding 0.1.0 → 0.1.2

package/README.md

@@ -31,6 +31,7 @@ new ComputedDataset({
 - The output column `vector` holds a unit vector of float32 values in the byte order of the platform. This is the form that `VectorSearch` of `@zukhruf/embeddings` reads, so the cosine similarity of two vectors is their dot product.
 - The identity of each row is the provider and the id of the model. When you change the model, the next refresh makes every vector again.
 - The rule sends 64 rows to `embedMany` at a time. When the model refuses a call, each row of that call is stored as `failed` with the error, and the next refresh asks again. The load does not fail.
+- A vector with no usable length, because it is all zeros or holds a value that is not finite, has no unit vector. Its row is stored as `failed` with the error, the other rows of the call stay `computed`, and the next refresh asks again.
 - To search, declare a view that gives `id` and `vector`, and give its name to `VectorSearch`. A search by publish date, such as `nearest`, also needs `published_at`. A query without a date range does not.
 
 ## Give the model its prompts

package/dist/embedding.js

@@ -57,9 +57,7 @@ export class Embedding extends Rule {
                 values: [...texts],
                 abortSignal: signal,
             });
-            return embeddings.map((embedding) => ({
-                value: { vector: unitVector(embedding) },
-            }));
+            return embeddings.map(unitVector);
         }
         catch (failure) {
             if (signal?.aborted)
@@ -71,7 +69,16 @@ export class Embedding extends Rule {
         }
     }
 }
+// A vector with no usable length has no unit vector: the division gives NaN,
+// which a search would rank as an answer. The row fails, and the next refresh
+// asks again.
 function unitVector(embedding) {
     const length = Math.hypot(...embedding);
-    return new Uint8Array(Float32Array.from(embedding, (value) => value / length).buffer);
+    if (length === 0 || !Number.isFinite(length))
+        return { error: `The model answered a vector of length ${length}.` };
+    return {
+        value: {
+            vector: new Uint8Array(Float32Array.from(embedding, (value) => value / length).buffer),
+        },
+    };
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@eltjs/embedding",
-  "version": "0.1.0",
+  "version": "0.1.2",
   "description": "A rule of an elt computed dataset: it gives each loaded row a vector of its text, made by an AI SDK embedding model, so a search can find rows by meaning.",
   "license": "Apache-2.0",
   "repository": {
@@ -40,16 +40,13 @@
     "node": ">=24"
   },
   "dependencies": {
-    "@eltjs/elt": "^0.1.0"
+    "@eltjs/elt": "^0.1.2",
+    "ai": "^7.0.165"
   },
   "devDependencies": {
     "@ai-sdk/provider": "^4.0.37",
-    "@eltjs/elt-sqlite": "^0.1.0",
+    "@eltjs/elt-sqlite": "^0.1.2",
     "@zukhruf/sqlite-open": "^0.8.10",
-    "@zukhruf/testing": "^0.8.10",
-    "ai": "^7.0.165"
-  },
-  "peerDependencies": {
-    "ai": "^7.0.148"
+    "@zukhruf/testing": "^0.8.10"
   }
 }