@tiens.nguyen/gonext-local-worker 1.0.157 → 1.0.158

package/gonext-local-worker.mjs

@@ -1610,6 +1610,7 @@ async function runAgentChatJob(job) {
       ragAwsAccessKeyId: payload?.ragAwsAccessKeyId ?? "",
       ragAwsSecretAccessKey: payload?.ragAwsSecretAccessKey ?? "",
       ragEmbedModel: payload?.ragEmbedModel ?? "",
+      ragEmbedUrl: payload?.ragEmbedUrl ?? "",
       ragTopK: payload?.ragTopK ?? 6,
     });
     // 30 min max for an agent run: multi-step ReAct on a 14B/31B with a cold prompt

package/gonext_agent_chat.py

@@ -1265,44 +1265,57 @@ def _rag_chunk_text(text: str, rel_path: str, ext: str):
     return chunks
 
 
-def _ollama_embed(ollama_root: str, model: str, texts: list) -> list:
-    """Embed a batch of texts via Ollama /api/embed (falls back to /api/embeddings)."""
-    root = (ollama_root or "").rstrip("/")
+def _embed(base: str, model: str, texts: list) -> list:
+    """Embed a batch of texts. Auto-detects the endpoint from `base` (a bare host:port,
+    an .../v1, or an Ollama host): tries OpenAI-compatible /v1/embeddings first (MLX
+    servers, modern Ollama), then Ollama /api/embed, then legacy /api/embeddings."""
+    # Normalize to a bare root (strip any /v1, /api, /api/embed(dings) suffix).
+    root = re.sub(r"/(v1|api/embeddings|api/embed|api)/?$", "", (base or "").rstrip("/")).rstrip("/")
     if not root:
-        raise RuntimeError("No Ollama endpoint available for embeddings.")
+        raise RuntimeError("No embedding endpoint configured (set the RAG embedding server URL).")
     ctx = _ssl_context()
-    # Preferred: batch /api/embed (recent Ollama).
-    try:
+    last_err = None
+
+    def _post(url, payload, timeout):
         req = urllib.request.Request(
-            f"{root}/api/embed",
-            data=json.dumps({"model": model, "input": texts}).encode("utf-8"),
-            headers={"Content-Type": "application/json"},
-            method="POST",
+            url, data=json.dumps(payload).encode("utf-8"),
+            headers={"Content-Type": "application/json"}, method="POST",
         )
-        with urllib.request.urlopen(req, timeout=300, context=ctx) as resp:
-            data = json.loads(resp.read().decode("utf-8", "replace"))
+        with urllib.request.urlopen(req, timeout=timeout, context=ctx) as resp:
+            return json.loads(resp.read().decode("utf-8", "replace"))
+
+    # 1) OpenAI-compatible /v1/embeddings — MLX servers on a port, and modern Ollama.
+    try:
+        data = _post(f"{root}/v1/embeddings", {"model": model, "input": texts}, 300)
+        items = data.get("data")
+        if isinstance(items, list) and items and isinstance(items[0], dict) and "embedding" in items[0]:
+            return [it["embedding"] for it in items]
+    except urllib.error.HTTPError as e:
+        if e.code not in (400, 404, 405, 501):
+            raise RuntimeError(f"Embeddings failed (HTTP {e.code}): {e.read().decode('utf-8', 'replace')[:200]}")
+        last_err = f"/v1/embeddings HTTP {e.code}"
+    except urllib.error.URLError as e:
+        raise RuntimeError(f"Embedding server unreachable at {root}: {getattr(e, 'reason', e)}")
+
+    # 2) Ollama batch /api/embed.
+    try:
+        data = _post(f"{root}/api/embed", {"model": model, "input": texts}, 300)
         embs = data.get("embeddings")
         if isinstance(embs, list) and embs and isinstance(embs[0], list):
             return embs
     except urllib.error.HTTPError as e:
-        if e.code != 404:
-            detail = e.read().decode("utf-8", "replace")[:200]
-            raise RuntimeError(f"Ollama embed failed (HTTP {e.code}): {detail}")
-    except urllib.error.URLError as e:
-        raise RuntimeError(f"Ollama embed unreachable: {getattr(e, 'reason', e)}")
-    # Fallback: per-text /api/embeddings (older Ollama).
+        if e.code not in (400, 404, 405, 501):
+            raise RuntimeError(f"Ollama embed failed (HTTP {e.code}): {e.read().decode('utf-8', 'replace')[:200]}")
+        last_err = f"/api/embed HTTP {e.code}"
+
+    # 3) Legacy Ollama /api/embeddings (per text).
     out = []
     for t in texts:
-        req = urllib.request.Request(
-            f"{root}/api/embeddings",
-            data=json.dumps({"model": model, "prompt": t}).encode("utf-8"),
-            headers={"Content-Type": "application/json"},
-            method="POST",
-        )
-        with urllib.request.urlopen(req, timeout=120, context=ctx) as resp:
-            data = json.loads(resp.read().decode("utf-8", "replace"))
+        data = _post(f"{root}/api/embeddings", {"model": model, "prompt": t}, 120)
         out.append(data.get("embedding") or [])
-    return out
+    if any(out):
+        return out
+    raise RuntimeError(f"No embedding endpoint responded at {root} ({last_err or 'unknown'}).")
 
 
 def _cosine(a: list, b: list) -> float:
@@ -1509,20 +1522,22 @@ def run_agent_chat(cfg):
     rag_akid = (cfg.get("ragAwsAccessKeyId") or "").strip()
     rag_secret = (cfg.get("ragAwsSecretAccessKey") or "").strip()
     rag_embed_model = (cfg.get("ragEmbedModel") or "nomic-embed-text").strip()
+    rag_embed_url = (cfg.get("ragEmbedUrl") or "").strip()
     try:
         rag_top_k = int(cfg.get("ragTopK") or 6)
     except (TypeError, ValueError):
         rag_top_k = 6
     rag_top_k = max(1, min(50, rag_top_k))
-    # Embeddings run on the Ollama coding server (same box as the coder). Strip the
-    # OpenAI-compat /v1 suffix to reach the native /api/embed root.
-    rag_ollama_root = re.sub(r"/v1/?$", "", (coding_base_url or agent_base_url or "").rstrip("/"))
+    # Embedding server: an explicit MLX/Ollama endpoint from Settings if given, else the
+    # agent's Ollama coding server. _embed() auto-detects OpenAI-compat (/v1/embeddings,
+    # e.g. an MLX server on a port) vs Ollama (/api/embed) — so a bare host:port works.
+    rag_embed_base = rag_embed_url or (coding_base_url or agent_base_url or "")
     try:
         import boto3 as _boto3_probe  # noqa: F401,PLC0415
         _boto3_ok = True
     except Exception:  # noqa: BLE001
         _boto3_ok = False
-    _RAG_AVAILABLE = bool(rag_enabled and rag_akid and rag_secret and rag_ollama_root and _boto3_ok)
+    _RAG_AVAILABLE = bool(rag_enabled and rag_akid and rag_secret and rag_embed_base and _boto3_ok)
     if rag_enabled and not _boto3_ok:
         _log("RAG requested but boto3 is not installed in the worker python — RAG tools disabled")
 
@@ -2678,7 +2693,7 @@ def run_agent_chat(cfg):
                     return "No indexable text files found at that path."
                 for i in range(0, len(records), 64):
                     batch = records[i:i + 64]
-                    vecs = _ollama_embed(rag_ollama_root, rag_embed_model, [r["text"] for r in batch])
+                    vecs = _embed(rag_embed_base, rag_embed_model, [r["text"] for r in batch])
                     for r, v in zip(batch, vecs):
                         r["embedding"] = v
                     _emit({"type": "step", "text": f"Embedded {min(i + 64, len(records))}/{len(records)} chunks…"})
@@ -2718,7 +2733,7 @@ def run_agent_chat(cfg):
             try:
                 import time as _t
                 records = _rag_chunk_text(text, "user-note", ".txt")
-                vecs = _ollama_embed(rag_ollama_root, rag_embed_model, [r["text"] for r in records])
+                vecs = _embed(rag_embed_base, rag_embed_model, [r["text"] for r in records])
                 for r, v in zip(records, vecs):
                     r["embedding"] = v
                 client, bucket, prefix = _rag_s3_and_loc()
@@ -2749,7 +2764,7 @@ def run_agent_chat(cfg):
                 chunks = _rag_load_chunks(client, bucket, base)
                 if not chunks:
                     return f"No knowledge base found for {source_url}. Run rag_index first."
-                qvec = _ollama_embed(rag_ollama_root, rag_embed_model, [query])[0]
+                qvec = _embed(rag_embed_base, rag_embed_model, [query])[0]
                 scored = sorted(
                     ((_cosine(qvec, c.get("embedding") or []), c) for c in chunks),
                     key=lambda x: x[0], reverse=True,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.157",
+  "version": "1.0.158",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",