@equinor/fusion-framework-cli-plugin-ai-index 3.0.0 → 3.0.1

package/dist/types/version.d.ts

@@ -1 +1 @@
-export declare const version = "3.0.0";
+export declare const version = "3.0.1";

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@equinor/fusion-framework-cli-plugin-ai-index",
-  "version": "3.0.0",
+  "version": "3.0.1",
   "description": "AI indexing plugin for Fusion Framework CLI providing document embedding and chunking utilities",
   "main": "dist/esm/index.js",
   "type": "module",
@@ -54,17 +54,17 @@
     "ts-morph": "^28.0.0",
     "zod": "^4.3.6",
     "@equinor/fusion-framework-cli-plugin-ai-base": "4.0.0",
-    "@equinor/fusion-framework-module-ai": "4.0.0",
     "@equinor/fusion-framework-module": "6.0.0",
+    "@equinor/fusion-framework-module-ai": "4.0.0",
     "@equinor/fusion-imports": "2.0.0"
   },
   "peerDependencies": {
-    "@equinor/fusion-framework-cli": "^15.0.0"
+    "@equinor/fusion-framework-cli": "^15.0.1"
   },
   "devDependencies": {
     "typescript": "^5.9.3",
     "vitest": "^4.1.0",
-    "@equinor/fusion-framework-cli": "^15.0.0"
+    "@equinor/fusion-framework-cli": "^15.0.1"
   },
   "scripts": {
     "build": "tsc -b",

package/src/bin/embed.ts

@@ -175,17 +175,24 @@ const EMBED_BATCH_SIZE = 500;
  * Number of concurrent embedding API requests in flight.
  *
  * Each request now carries EMBED_BATCH_SIZE texts in a single HTTP call
- * (LangChain batchSize is aligned), so 2 concurrent requests already
- * saturate most Azure OpenAI TPM quotas.
+ * (LangChain batchSize is aligned), so 3 concurrent requests should
+ * saturate most Azure OpenAI TPM quotas without triggering rate limits.
  */
-const EMBED_BATCH_CONCURRENCY = 2;
+const EMBED_BATCH_CONCURRENCY = 3;
 
 /**
  * Maximum time (ms) to wait before flushing a partial embedding batch.
- * Without this, `bufferCount` waits indefinitely for a full batch, which
- * starves `mergeMap` concurrency when upstream document throughput is slow.
+ *
+ * A longer window lets more documents accumulate before triggering an
+ * HTTP call, which drastically cuts the number of round-trips when
+ * upstream (metadata enrichment) feeds documents slowly.
+ *
+ * Full batches (EMBED_BATCH_SIZE) are emitted immediately regardless of
+ * the timer, and `bufferTime` flushes any remainder the instant the
+ * source stream completes — so a large value only affects mid-stream
+ * partial batches, not tail latency.
  */
-const EMBED_BUFFER_FLUSH_MS = 500;
+const EMBED_BUFFER_FLUSH_MS = 10_000;
 
 /** Maximum retry attempts for transient / rate-limit errors per chunk. */
 const MAX_RETRIES = 4;

package/src/version.ts

@@ -1,2 +1,2 @@
 // Generated by genversion.
-export const version = '3.0.0';
+export const version = '3.0.1';