revert: stop putting the endpoint in the embedding index identity

src/core/openai-embedding.ts

@@ -91,12 +91,6 @@ export class OpenAIEmbeddingClient {
     this.documentBody = options.documentBody ?? {};
     this.indexId = indexIdentity({
       model: this.model,
-      // The endpoint takes part in the identity. Two deployments of the same model name can
-      // serve different weights, and a vector from one is not comparable with a query
-      // embedded by the other - but nothing else in the identity would notice, so switching
-      // endpoints used to leave the stored vectors silently in place and mix two spaces.
-      // With it here, a switch makes the index stale and the existing backfill re-embeds.
-      providerEndpoint: endpointIdentity(options.baseUrl),
       ...(options.indexNamespace ? { indexNamespace: options.indexNamespace } : {}),
       dimensions: this.dimensions,
       documentTemplate: this.documentTemplate,
@@ -238,22 +232,8 @@ export class OpenAIEmbeddingClient {
  * per-call bodies means existing embeddings cannot be reused. Serialisation order
  * is part of the contract, so keep the spread order identical when editing.
  */
-/** The part of the endpoint that identifies the deployment: scheme, host and path, without
- *  a trailing slash or any query, so an equivalent spelling of the same endpoint keeps the
- *  same identity and a genuinely different one does not. */
-function endpointIdentity(baseUrl: string | undefined): string | null {
-  if (!baseUrl) return null;
-  try {
-    const url = new URL(baseUrl);
-    return `${url.protocol}//${url.host}${url.pathname.replace(/\/+$/, "")}`;
-  } catch {
-    return baseUrl.replace(/\/+$/, "");
-  }
-}
-
 function indexIdentity(input: {
   model: string;
-  providerEndpoint?: string | null;
   indexNamespace?: string;
   dimensions?: number;
   documentTemplate: string;
@@ -266,7 +246,6 @@ function indexIdentity(input: {
   const fingerprint = createHash("sha256")
     .update(
       JSON.stringify({
-        ...(input.providerEndpoint ? { providerEndpoint: input.providerEndpoint } : {}),
         ...(input.indexNamespace ? { indexNamespace: input.indexNamespace } : {}),
         dimensions: input.dimensions ?? null,
         documentTemplate: input.documentTemplate,

tests/core/openai-embedding.test.ts

@@ -286,28 +286,3 @@ test("a dropped connection is still retried", async () => {
     );
   }
 });
-
-test("the embedding index identity separates deployments of the same model", () => {
-  // Two regions of the same provider serve the same model name. A vector from one is not
-  // comparable with a query embedded by the other, and nothing else in the identity - model,
-  // dimensions, templates - differs, so the endpoint has to be part of it. Without this, a
-  // switch left every stored vector in place and silently mixed the two spaces.
-  const mainland = new OpenAIEmbeddingClient({
-    baseUrl: "https://embed.example.cn/compatible-mode/v1",
-    model: "text-embedding-v4",
-    dimensions: 1024,
-  });
-  const singapore = new OpenAIEmbeddingClient({
-    baseUrl: "https://embed.example.sg/compatible-mode/v1",
-    model: "text-embedding-v4",
-    dimensions: 1024,
-  });
-  const spelling = new OpenAIEmbeddingClient({
-    baseUrl: "https://embed.example.cn/compatible-mode/v1/",
-    model: "text-embedding-v4",
-    dimensions: 1024,
-  });
-  assert.notEqual(mainland.indexId, singapore.indexId, "a different endpoint is a different index");
-  assert.equal(spelling.indexId, mainland.indexId, "a trailing slash is the same endpoint");
-  assert.match(mainland.indexId, /^text-embedding-v4@[0-9a-f]{12}$/);
-});