@melaya/runner 1.0.38 → 1.0.40

package/dist/localRagIngest.py

@@ -113,6 +113,15 @@ async def _ingest(args) -> int:
                 dimensions=768,
             )
             embedder.client.base_url = "http://localhost:1234/v1"
+            # LM Studio's /v1/embeddings rejects the OpenAI-specific
+            # `dimensions` kwarg (only OpenAI text-embedding-3 supports
+            # output-dim reduction). Strip it before each call so the
+            # request is shaped exactly as LMS expects.
+            _orig_create = embedder.client.embeddings.create
+            async def _lms_create(**kw):
+                kw.pop("dimensions", None)
+                return await _orig_create(**kw)
+            embedder.client.embeddings.create = _lms_create  # type: ignore[assignment]
         else:
             _emit("error", reason=(
                 f"refused embedder provider '{args.embedder_provider}' for Mode B "
@@ -173,6 +182,10 @@ async def _ingest(args) -> int:
         ext = f.suffix.lower()
         if ext not in SUPPORTED_EXTS:
             continue
+        # macOS AppleDouble resource forks (._foo.docx, ._.DS_Store, …) —
+        # they're not real documents and every extractor will choke on them.
+        if f.name.startswith("._") or f.name == ".DS_Store":
+            continue
         try:
             raw = f.read_bytes()
         except Exception as exc:

package/dist/pythonEnv.js

@@ -108,6 +108,17 @@ const PIP_DEPS = [
     // ships Windows wheels, MilvusLite does not — using Qdrant for OS
     // parity across runners.)
     "ollama",
+    // nltk — used by agentscope.rag readers' default split_by="sentence"
+    // chunker. Without it, every reader raises ImportError at chunk time
+    // (the lazy import happens INSIDE the chunking call, not at module load,
+    // so the error surfaces per-file as "nltk is not installed"). The
+    // readers also lazily `nltk.download("punkt", quiet=True)` on first use.
+    "nltk",
+    // pandas — agentscope.rag.ExcelReader uses pandas.read_excel for sheet
+    // iteration. Without it, .xlsx files in a Mode B folder fail with
+    // "pandas is not installed". Numpy is already in PIP_DEPS (line 57) so
+    // the wheel install is fast.
+    "pandas",
 ];
 export function venvPython() {
     return platform() === "win32"

package/localRagIngest.py

@@ -113,6 +113,15 @@ async def _ingest(args) -> int:
                 dimensions=768,
             )
             embedder.client.base_url = "http://localhost:1234/v1"
+            # LM Studio's /v1/embeddings rejects the OpenAI-specific
+            # `dimensions` kwarg (only OpenAI text-embedding-3 supports
+            # output-dim reduction). Strip it before each call so the
+            # request is shaped exactly as LMS expects.
+            _orig_create = embedder.client.embeddings.create
+            async def _lms_create(**kw):
+                kw.pop("dimensions", None)
+                return await _orig_create(**kw)
+            embedder.client.embeddings.create = _lms_create  # type: ignore[assignment]
         else:
             _emit("error", reason=(
                 f"refused embedder provider '{args.embedder_provider}' for Mode B "
@@ -173,6 +182,10 @@ async def _ingest(args) -> int:
         ext = f.suffix.lower()
         if ext not in SUPPORTED_EXTS:
             continue
+        # macOS AppleDouble resource forks (._foo.docx, ._.DS_Store, …) —
+        # they're not real documents and every extractor will choke on them.
+        if f.name.startswith("._") or f.name == ".DS_Store":
+            continue
         try:
             raw = f.read_bytes()
         except Exception as exc:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@melaya/runner",
-  "version": "1.0.38",
+  "version": "1.0.40",
   "description": "Run Melaya AI pipelines locally with your own LM Studio or Ollama models",
   "license": "UNLICENSED",
   "private": false,