@melaya/runner 1.0.46 → 1.0.48

package/dist/localRagIngest.py

@@ -206,6 +206,25 @@ async def _ingest(args) -> int:
         from agentscope.rag import QdrantStore, SimpleKnowledge
         store_path = _resolve_store_path(args.pipeline_name)
         store_path.mkdir(parents=True, exist_ok=True)
+        # ── Fix qdrant-client's location-vs-path ambiguity ───────────────
+        # agentscope's QdrantStore only exposes `location=` and forwards
+        # it verbatim to `AsyncQdrantClient(location=...)`. When `location`
+        # is a filesystem path (not `:memory:` and not a URL), qdrant-client
+        # tries to parse it as a host:port → getaddrinfo → fails with
+        # `[Errno 8] nodename nor servname provided` on Mac / `[Errno
+        # 11001] getaddrinfo failed` on Windows. Per-file embedding then
+        # crashes 240 times in a row. Reroute `location=<path>` into the
+        # dedicated `path=` kwarg the embedded-mode wants.
+        import qdrant_client as _qc
+        _orig_qcinit = _qc.AsyncQdrantClient.__init__
+        def _patched_qcinit(self, location=None, **_kw):
+            if (location and location != ":memory:"
+                    and "://" not in str(location)
+                    and not _kw.get("path")):
+                _kw["path"] = location
+                location = None
+            _orig_qcinit(self, location=location, **_kw)
+        _qc.AsyncQdrantClient.__init__ = _patched_qcinit
         store = QdrantStore(
             location=str(store_path),
             collection_name=args.pipeline_name,

package/dist/modelLoader.js

@@ -68,7 +68,17 @@ const FAMILY_OVERRIDES = [
     { pattern: /qwen3.*vl.*\b8b\b/i, family: "qwen3", tier: "agentic-capable", thinking: "on", reason: "qwen3-vl-8b: agentic-capable, full thinking" },
     // ── Qwen3 — capable (≥8B) BEFORE the smaller non-vl 4b pattern, so
     //    14b / 32b don't get caught by the 4b regex's stem match. ──
-    { pattern: /qwen3.*\b(8b|14b|32b)\b/i, family: "qwen3", tier: "agentic-capable", thinking: "on", reason: "qwen3 ≥8b: agentic-capable" },
+    //    Mainline qwen3 releases ship 8B, 14B, 32B. Custom/finetuned
+    //    27B/30B/72B variants exist (e.g. qwen3.6-27b on LM Studio) — we
+    //    enumerate them so `family="qwen3"` propagates and model.py can
+    //    apply chat_template_kwargs.enable_thinking=false when needed.
+    //    Without the explicit 27b entry, qwen3.6-27b falls through to
+    //    the param-count path with `family=null`, the disable_thinking
+    //    branch in model.py is skipped, and the model produces
+    //    thinking-only responses that the OpenAI formatter drops (run
+    //    1ecad7cb hit this — Copywriter ended in 11.5KB thinking block,
+    //    ReportWriter emailed [CONTENT NOT AVAILABLE]).
+    { pattern: /qwen3.*\b(8b|14b|27b|30b|32b|72b)\b/i, family: "qwen3", tier: "agentic-capable", thinking: "off", reason: "qwen3 ≥8b: agentic-capable, thinking disabled (model is large enough to reason without explicit <think> blocks; thinking-only responses get stripped by the OpenAI formatter and lose data to next agent)" },
     // ── Qwen3 — small variants ──
     { pattern: /qwen3.*\b1\.7b\b/i, family: "qwen3", tier: "text-only", thinking: "off", reason: "qwen3-1.7b: param count below the agentic 7B floor; thinking disabled to prevent <think> loops" },
     { pattern: /qwen3.*\b0\.5b\b/i, family: "qwen3", tier: "text-only", thinking: "off", reason: "qwen3-0.5b: too small for agentic work" },
@@ -214,7 +224,7 @@ export function classifyModel(id, lmstudioMeta) {
 // Bump whenever FAMILY_OVERRIDES, tier thresholds, or any classifier
 // logic changes in a way that could re-tier an existing cached model.
 // Cached profiles with a mismatched version are ignored and re-classified.
-const PROFILE_SCHEMA_VERSION = 2;
+const PROFILE_SCHEMA_VERSION = 3;
 const LMSTUDIO_BASE = process.env.LMSTUDIO_BASE_URL || "http://127.0.0.1:1234";
 const OLLAMA_BASE = process.env.OLLAMA_BASE_URL || "http://127.0.0.1:11434";
 // JIT-load timeout. 8B-class quantised models on M-series Macs typically

package/localRagIngest.py

@@ -206,6 +206,25 @@ async def _ingest(args) -> int:
         from agentscope.rag import QdrantStore, SimpleKnowledge
         store_path = _resolve_store_path(args.pipeline_name)
         store_path.mkdir(parents=True, exist_ok=True)
+        # ── Fix qdrant-client's location-vs-path ambiguity ───────────────
+        # agentscope's QdrantStore only exposes `location=` and forwards
+        # it verbatim to `AsyncQdrantClient(location=...)`. When `location`
+        # is a filesystem path (not `:memory:` and not a URL), qdrant-client
+        # tries to parse it as a host:port → getaddrinfo → fails with
+        # `[Errno 8] nodename nor servname provided` on Mac / `[Errno
+        # 11001] getaddrinfo failed` on Windows. Per-file embedding then
+        # crashes 240 times in a row. Reroute `location=<path>` into the
+        # dedicated `path=` kwarg the embedded-mode wants.
+        import qdrant_client as _qc
+        _orig_qcinit = _qc.AsyncQdrantClient.__init__
+        def _patched_qcinit(self, location=None, **_kw):
+            if (location and location != ":memory:"
+                    and "://" not in str(location)
+                    and not _kw.get("path")):
+                _kw["path"] = location
+                location = None
+            _orig_qcinit(self, location=location, **_kw)
+        _qc.AsyncQdrantClient.__init__ = _patched_qcinit
         store = QdrantStore(
             location=str(store_path),
             collection_name=args.pipeline_name,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@melaya/runner",
-  "version": "1.0.46",
+  "version": "1.0.48",
   "description": "Run Melaya AI pipelines locally with your own LM Studio or Ollama models",
   "license": "UNLICENSED",
   "private": false,