feat(rag): add OpenVINO NPU embedding services

2026-06-04 13:07:51 -07:00
parent 83d0ced08c
commit 0a6f84fbf3
6 changed files with 651 additions and 0 deletions
@@ -0,0 +1,117 @@
+#!/usr/bin/env python3
+"""RAG/embedding health HTTP wrapper for n8n.
+
+Listens on 0.0.0.0:18814 so the n8n container can call it via
+http://172.19.0.1:18814.
+
+Endpoints:
+  GET  /healthz -> service liveness
+  POST /check   -> run ~/.hermes/scripts/rag_embedding_health.py and return JSON
+"""
+
+from __future__ import annotations
+
+import http.server
+import json
+import os
+import subprocess
+import time
+from pathlib import Path
+
+PORT = int(os.environ.get("PORT", "18814"))
+CHECK_SCRIPT = Path(os.environ.get("RAG_HEALTH_SCRIPT", "/home/will/.hermes/scripts/rag_embedding_health.py"))
+TIMEOUT = int(os.environ.get("RAG_HEALTH_TIMEOUT", "180"))
+
+
+class Handler(http.server.BaseHTTPRequestHandler):
+    def do_GET(self):
+        if self.path.rstrip("/") == "/healthz":
+            self._json({"status": "ok", "service": "rag-embedding-health"})
+        else:
+            self._json({"error": "not found"}, status=404)
+
+    def do_POST(self):
+        if self.path.rstrip("/") != "/check":
+            self._json({"error": "not found"}, status=404)
+            return
+
+        started = time.time()
+        if not CHECK_SCRIPT.exists():
+            self._json(
+                {
+                    "ok": False,
+                    "status": "failed",
+                    "exitCode": 127,
+                    "output": f"RAG health script missing: {CHECK_SCRIPT}",
+                    "durationMs": 0,
+                },
+                status=200,
+            )
+            return
+
+        env = os.environ.copy()
+        env.setdefault("HERMES_HOME", "/home/will/.hermes")
+        env.setdefault("OLLAMA_BASE_URL", "http://127.0.0.1:18817")
+        env.setdefault("RAG_EMBED_MODEL", "bge-base-en-v1.5-int8-ov")
+        env.setdefault("N8N_URL", "http://127.0.0.1:18808")
+        env.setdefault("OBSIDIAN_REINDEX_URL", "http://127.0.0.1:18810")
+
+        try:
+            proc = subprocess.run(
+                [str(CHECK_SCRIPT)],
+                text=True,
+                capture_output=True,
+                timeout=TIMEOUT,
+                check=False,
+                env=env,
+            )
+            output = (proc.stdout or proc.stderr or "").strip()
+            self._json(
+                {
+                    "ok": proc.returncode == 0,
+                    "status": "ok" if proc.returncode == 0 else "failed",
+                    "exitCode": proc.returncode,
+                    "output": output[:4000],
+                    "durationMs": int((time.time() - started) * 1000),
+                },
+                status=200,
+            )
+        except subprocess.TimeoutExpired:
+            self._json(
+                {
+                    "ok": False,
+                    "status": "timeout",
+                    "exitCode": 124,
+                    "output": f"RAG/embedding health check timed out after {TIMEOUT}s",
+                    "durationMs": int((time.time() - started) * 1000),
+                },
+                status=200,
+            )
+        except Exception as exc:
+            self._json(
+                {
+                    "ok": False,
+                    "status": "error",
+                    "exitCode": 1,
+                    "output": str(exc)[:4000],
+                    "durationMs": int((time.time() - started) * 1000),
+                },
+                status=200,
+            )
+
+    def _json(self, data, status=200):
+        body = json.dumps(data, indent=2).encode("utf-8")
+        self.send_response(status)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def log_message(self, format, *args):
+        return
+
+
+if __name__ == "__main__":
+    server = http.server.HTTPServer(("0.0.0.0", PORT), Handler)
+    print(f"rag-embedding-health listening on 0.0.0.0:{PORT}", flush=True)
+    server.serve_forever()