@kelceyp/caw-server 1.0.189 → 1.0.191

package/dist/embedding-worker.js

@@ -0,0 +1,98 @@
+#!/usr/bin/env node
+/**
+ * embedding-worker.js — Standalone Node.js child process for embedding generation.
+ *
+ * Used as a fallback when Bun's ONNX runtime cannot load @huggingface/transformers
+ * natively. Spawned by EmbeddingProvider when the native path fails.
+ *
+ * IPC protocol (JSON-lines over stdin/stdout):
+ *   Startup:  sends  { ready: true }  after pipeline is loaded
+ *             sends  { error: "..." } if startup fails
+ *   Request:  parent → worker: { texts: ["...", "..."] }\n
+ *   Response: worker → parent: { vectors: [[...], ...] }\n
+ *                           or { error: "..." }\n
+ *
+ * Vectors are raw (un-normalised) Float32Arrays serialised as plain number arrays.
+ * Normalisation is performed by EmbeddingProvider after receiving the response.
+ *
+ * This script is intentionally standalone — no imports from the CAW codebase.
+ * It runs in Node.js, not Bun, to avoid the ONNX runtime incompatibility.
+ */
+
+import { createInterface } from 'readline';
+
+const MODEL_NAME = process.env.EMBEDDING_MODEL || 'nomic-ai/nomic-embed-text-v1.5';
+const CACHE_DIR = process.env.EMBEDDING_CACHE_DIR || null;
+
+const sendLine = (obj) => {
+    process.stdout.write(JSON.stringify(obj) + '\n');
+};
+
+const run = async () => {
+    let pipe;
+
+    try {
+        const { pipeline, env } = await import('@huggingface/transformers');
+
+        if (CACHE_DIR) {
+            env.cacheDir = CACHE_DIR;
+        }
+
+        // Log download progress to stderr (visible in parent process output)
+        pipe = await pipeline('feature-extraction', MODEL_NAME, {
+            quantized: true,
+            progress_callback: (progress) => {
+                if (progress.status === 'download') {
+                    process.stderr.write(`[EmbeddingWorker] Downloading: ${progress.file} (${Math.round((progress.loaded / progress.total) * 100)}%)\n`);
+                } else if (progress.status === 'ready') {
+                    process.stderr.write(`[EmbeddingWorker] Ready: ${progress.file}\n`);
+                }
+            }
+        });
+
+        sendLine({ ready: true });
+    } catch (err) {
+        sendLine({ error: `Pipeline load failed: ${err.message}` });
+        process.exit(1);
+    }
+
+    // Process requests from parent line by line
+    const rl = createInterface({ input: process.stdin });
+
+    rl.on('line', async (line) => {
+        let req;
+        try {
+            req = JSON.parse(line);
+        } catch {
+            sendLine({ error: `Invalid JSON request: ${line}` });
+            return;
+        }
+
+        const { texts } = req;
+        if (!Array.isArray(texts)) {
+            sendLine({ error: 'Request must have a texts array' });
+            return;
+        }
+
+        try {
+            const output = await pipe(texts, { pooling: 'mean', normalize: false });
+            // Serialise: output.data is a flat Float32Array [batch * dims]
+            const dims = output.dims[output.dims.length - 1];
+            const vectors = [];
+            for (let i = 0; i < texts.length; i++) {
+                const start = i * dims;
+                const end = start + dims;
+                vectors.push(Array.from(output.data.slice(start, end)));
+            }
+            sendLine({ vectors });
+        } catch (err) {
+            sendLine({ error: `Embedding failed: ${err.message}` });
+        }
+    });
+
+    rl.on('close', () => {
+        process.exit(0);
+    });
+};
+
+run();