feat(rag): switch Obsidian endpoint to NPU embeddings

Makefile

@@ -16,7 +16,8 @@ OPENCLAW_PORT ?= 18789
 QEMU_URI ?= qemu:///system
 LLAMA_CPP_URL ?= http://127.0.0.1:18806
 OLLAMA_URL ?= http://127.0.0.1:18807
-OLLAMA_EMBED_MODEL ?= nomic-embed-text
+OPENVINO_EMBED_URL ?= http://127.0.0.1:18817
+OPENVINO_EMBED_MODEL ?= bge-base-en-v1.5-int8-ov
 
 DC := $(COMPOSE) -f $(COMPOSE_FILE)
 COMMON_DC := $(COMPOSE) -f $(COMMON_COMPOSE_FILE)
@@ -28,7 +29,7 @@ REQUIRE_CONFIRM = test "$(CONFIRM)" = "yes" || { echo "This target changes VM/ga
 REQUIRE_INSTANCE = test -n "$(OPENCLAW_HOST)" -a -n "$(OPENCLAW_DOMAIN)" || { echo "Unknown OpenClaw HOST=$(HOST) in $(OPENCLAW_REGISTRY)"; exit 2; }
 
 .DEFAULT_GOAL := help
-.PHONY: help config ps status local-ai-health ollama-embed-health up down restart pull build logs shell clean \
+.PHONY: help config ps status local-ai-health openvino-embed-health up down restart pull build logs shell clean \
 	api-up api-down api-restart api-init api-init-force api-health api-dedup api-logs \
 	voice-up voice-gpu voice-cpu voice-down voice-build voice-logs \
 	search-up search-down automation-up automation-down n8n-logs \
@@ -53,7 +54,7 @@ ps: ## Show root Docker Compose service status.
 
 status: ps local-ai-health ## Show Docker service status plus host-side local AI endpoints.
 
-local-ai-health: ## Check host-side llama.cpp LLM and Ollama embeddings endpoints.
+local-ai-health: ## Check host-side llama.cpp LLM, Ollama fallback, and OpenVINO NPU embeddings endpoints.
 	@printf "\nHost-side local AI endpoints:\n"
 	@printf "llama.cpp (%s): " "$(LLAMA_CPP_URL)"; \
 		if curl -fsS --max-time 3 "$(LLAMA_CPP_URL)/v1/models" >/tmp/swarm-llama-models.json 2>/dev/null; then \
@@ -62,14 +63,17 @@ local-ai-health: ## Check host-side llama.cpp LLM and Ollama embeddings endpoint
 			printf "FAILED\n"; \
 		fi
 	@printf "ollama.service: "; systemctl --user is-active ollama.service 2>/dev/null || true
-	@printf "Ollama API (%s): " "$(OLLAMA_URL)"; \
+	@printf "Ollama fallback API (%s): " "$(OLLAMA_URL)"; \
 		curl -fsS --max-time 3 "$(OLLAMA_URL)/api/version" 2>/dev/null | jq -r '"OK version=" + .version' || printf "FAILED\n"
+	@printf "openvino-embeddings.service: "; systemctl --user is-active openvino-embeddings.service 2>/dev/null || true
+	@printf "OpenVINO NPU embeddings (%s): " "$(OPENVINO_EMBED_URL)"; \
+		curl -fsS --max-time 3 "$(OPENVINO_EMBED_URL)/healthz" 2>/dev/null | jq -r '"OK model=" + .model + " device=" + .device' || printf "FAILED\n"
 
-ollama-embed-health: ## Smoke-test Ollama embeddings using OLLAMA_EMBED_MODEL=nomic-embed-text.
-	@curl -fsS --max-time 20 "$(OLLAMA_URL)/api/embed" \
+openvino-embed-health: ## Smoke-test OpenVINO NPU embeddings using OPENVINO_EMBED_MODEL=bge-base-en-v1.5-int8-ov.
+	@curl -fsS --max-time 20 "$(OPENVINO_EMBED_URL)/v1/embeddings" \
 		-H 'Content-Type: application/json' \
-		-d '{"model":"$(OLLAMA_EMBED_MODEL)","input":"socket check"}' \
-		| jq -r '"embeddings=" + ((.embeddings // []) | length | tostring) + " dim=" + (((.embeddings // [[]])[0] // []) | length | tostring)'
+		-d '{"model":"$(OPENVINO_EMBED_MODEL)","input":"socket check"}' \
+		| jq -r '"embeddings=" + ((.data // []) | length | tostring) + " dim=" + (((.data // [{embedding: []}])[0].embedding // []) | length | tostring) + " npu_busy_delta_us=" + ((.npu_busy_delta_us // 0) | tostring)'
 
 up: ## Start root compose services. Use PROFILE=api,voice,search,automation or SERVICE=name.
 	@if [ -n "$(PROFILE)" ]; then \
@@ -137,23 +141,23 @@ api-dedup: ## Remove duplicate LiteLLM model DB entries.
 api-logs: ## Follow LiteLLM logs.
 	$(DC) logs -f --tail="$(LOGS_TAIL)" litellm litellm-db litellm-init
 
-voice-up: ## Start all voice services.
+voice-up: ## Start default voice services: NPU Whisper and Kokoro TTS.
 	$(DC) --profile voice up -d
 
-voice-gpu: ## Start GPU whisper server and Kokoro TTS.
-	$(DC) --profile voice up -d whisper-server-gpu kokoro-tts
+voice-gpu: ## Start manual GPU whisper fallback and Kokoro TTS.
+	$(DC) --profile voice-gpu --profile voice up -d whisper-server-gpu kokoro-tts
 
 voice-cpu: ## Start CPU whisper server and Kokoro TTS.
-	$(DC) --profile voice up -d whisper-server kokoro-tts
+	$(DC) --profile voice-cpu-backup --profile voice up -d whisper-server kokoro-tts
 
 voice-down: ## Stop voice profile services.
-	$(DC) --profile voice down
+	$(DC) --profile voice --profile voice-gpu --profile voice-cpu-backup down
 
 voice-build: ## Build the custom Blackwell CUDA whisper image.
-	$(DC) --profile voice build whisper-server-gpu
+	$(DC) --profile voice-gpu build whisper-server-gpu
 
-voice-logs: ## Follow voice service logs.
-	$(DC) logs -f --tail="$(LOGS_TAIL)" whisper-server-gpu whisper-server kokoro-tts
+voice-logs: ## Follow default voice service logs.
+	$(DC) logs -f --tail="$(LOGS_TAIL)" whisper-server-npu kokoro-tts
 
 search-up: ## Start Brave Search MCP and SearXNG.
 	$(DC) --profile search up -d

docker-compose.yaml

@@ -37,7 +37,7 @@ services:
   whisper-init:
     image: ghcr.io/ggml-org/whisper.cpp@sha256:672650b5e67f9cb86af7ac6e09dea8eac12a024086e1e5c0172fdccf336aba09
     container_name: whisper-init
-    profiles: ["voice"]
+    profiles: ["voice", "voice-cpu-backup"]
     restart: "no"
     volumes:
       - whisper-models:/app/models
@@ -54,17 +54,15 @@ services:
           fi
         done
 
-  # Primary whisper.cpp server: NVIDIA RTX 5070 Ti via CUDA (Blackwell sm_120).
-  # Uses ggml-base.bin to keep the service alive while llama-server owns most of
-  # the laptop GPU VRAM. The previous ggml-small.bin profile needed ~465 MiB
-  # contiguous CUDA memory and restarted when only ~560 MiB fragmented VRAM was
-  # free. CPU whisper-server below remains the higher-accuracy fallback.
+  # Manual GPU whisper.cpp fallback: NVIDIA RTX 5070 Ti via CUDA (Blackwell sm_120).
+  # Kept out of the normal `voice` profile because the OpenVINO NPU Whisper
+  # service is the default and this container consumes GPU resources.
   #
   # The official `ghcr.io/ggml-org/whisper.cpp:main-cuda` ships kernels only
   # for sm_75/80/86/90 and fails to init CUDA on Blackwell. We build a custom
   # image with `CMAKE_CUDA_ARCHITECTURES=120` from the local Dockerfile.
   # Build manually with: docker build -t whisper.cpp:cuda-blackwell ./whisper-cuda-blackwell
-  # Or `docker compose --profile voice build whisper-server-gpu`.
+  # Or `docker compose --profile voice-gpu build whisper-server-gpu`.
   whisper-server-gpu:
     image: whisper.cpp:cuda-blackwell
     build:
@@ -72,7 +70,7 @@ services:
       dockerfile: Dockerfile
     container_name: whisper-server-gpu
     restart: unless-stopped
-    profiles: ["voice"]
+    profiles: ["voice-gpu"]
     ports:
       - "18801:8080"
     volumes:
@@ -115,16 +113,62 @@ services:
       agentmon.role: "voice"
       agentmon.port: "18801"
 
-  # Fallback whisper.cpp server: CPU-only, medium model.
-  # Kept around for resilience — runs if the GPU server is down (driver issue,
-  # gemma takes all VRAM, custom image broken, etc.). Uses no GPU resources.
-  # ~14 s per short clip (medium-on-CPU is 90x slower than small-on-GPU above).
-  # Start with: docker compose --profile voice up -d whisper-server
+  # Experimental OpenVINO GenAI Whisper server using the Intel NPU.
+  # This is not whisper.cpp; it implements the same OpenAI-style
+  # /v1/audio/transcriptions route using OpenVINO WhisperPipeline on NPU.
+  # Host requirements: intel-npu-driver-bin installed, /dev/accel/accel0 present,
+  # and the host NPU Level Zero driver/compiler libraries mounted below.
+  whisper-server-npu:
+    image: whisper-openvino-npu:local
+    build:
+      context: ./whisper-openvino-npu
+      dockerfile: Dockerfile
+    container_name: whisper-server-npu
+    restart: unless-stopped
+    profiles: ["voice"]
+    ports:
+      - "18816:8080"
+    devices:
+      - /dev/accel/accel0:/dev/accel/accel0
+    group_add:
+      - "987" # host render group gid on willlaptop
+    environment:
+      - WHISPER_DEVICE=NPU
+      - WHISPER_MODEL_DIR=/models/whisper-tiny-fp16-ov
+      - LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu
+      - ZE_ENABLE_ALT_DRIVERS=/usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1
+    volumes:
+      - /home/will/.cache/openvino-models/whisper-tiny-fp16-ov:/models/whisper-tiny-fp16-ov:ro
+      - /usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1.32.1:/usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1.32.1:ro
+      - /usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1:/usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1:ro
+      - /usr/lib/x86_64-linux-gnu/libze_intel_npu.so:/usr/lib/x86_64-linux-gnu/libze_intel_npu.so:ro
+      - /usr/lib/x86_64-linux-gnu/libnpu_driver_compiler.so:/usr/lib/x86_64-linux-gnu/libnpu_driver_compiler.so:ro
+    healthcheck:
+      test:
+        [
+          "CMD-SHELL",
+          "curl -f http://localhost:8080/health >/dev/null 2>&1 || exit 1",
+        ]
+      interval: 30s
+      timeout: 5s
+      start_period: 30s
+      retries: 3
+    labels:
+      agentmon.monitor: "true"
+      agentmon.role: "voice"
+      agentmon.port: "18816"
+
+  # Manual fallback whisper.cpp server: CPU-only, medium model.
+  # Kept around for resilience — runs if the NPU/GPU servers are down. Uses no
+  # accelerator resources, but is slow (~14 s per short clip).
+  # Disabled from the normal `voice` profile now that `whisper-server-npu` is
+  # the trial default. Start manually with:
+  #   docker compose --profile voice-cpu-backup up -d whisper-server
   whisper-server:
     image: ghcr.io/ggml-org/whisper.cpp@sha256:672650b5e67f9cb86af7ac6e09dea8eac12a024086e1e5c0172fdccf336aba09
     container_name: whisper-server
     restart: unless-stopped
-    profiles: ["voice"]
+    profiles: ["voice-cpu-backup"]
     ports:
       - "18811:8080"
     volumes:

docs/swarm-infrastructure.html

@@ -83,13 +83,13 @@
         <!-- Local services -->
         <g><rect x="965" y="165" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="165" width="210" height="80" rx="9" fill="rgba(6,78,59,.4)" stroke="#34d399" stroke-width="1.6"/><text x="1070" y="195" text-anchor="middle" class="title">LiteLLM</text><text x="1070" y="216" text-anchor="middle" class="tiny">LLM router + DB</text><text x="1070" y="234" text-anchor="middle" class="port">:18804</text></g>
         <g><rect x="965" y="275" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="275" width="210" height="80" rx="9" fill="rgba(8,51,68,.4)" stroke="#22d3ee" stroke-width="1.6"/><text x="1070" y="305" text-anchor="middle" class="title">Search</text><text x="1070" y="326" text-anchor="middle" class="tiny">SearXNG + Brave MCP</text><text x="1070" y="344" text-anchor="middle" class="port">:18803 / :18802</text></g>
-        <g><rect x="965" y="385" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="385" width="210" height="80" rx="9" fill="rgba(8,51,68,.4)" stroke="#22d3ee" stroke-width="1.6"/><text x="1070" y="415" text-anchor="middle" class="title">Voice</text><text x="1070" y="436" text-anchor="middle" class="tiny">Kokoro + Whisper</text><text x="1070" y="454" text-anchor="middle" class="port">:18805 / :18811</text></g>
+        <g><rect x="965" y="385" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="385" width="210" height="80" rx="9" fill="rgba(8,51,68,.4)" stroke="#22d3ee" stroke-width="1.6"/><text x="1070" y="415" text-anchor="middle" class="title">Voice</text><text x="1070" y="436" text-anchor="middle" class="tiny">Kokoro + Whisper</text><text x="1070" y="454" text-anchor="middle" class="port">:18805 / :18816</text></g>
         <g><rect x="965" y="555" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="555" width="210" height="80" rx="9" fill="rgba(76,29,149,.4)" stroke="#a78bfa" stroke-width="1.6"/><text x="1070" y="585" text-anchor="middle" class="title">Docker services</text><text x="1070" y="606" text-anchor="middle" class="tiny">agentmon.monitor=true</text><text x="1070" y="624" text-anchor="middle" class="port">swarm/service snapshots</text></g>
         <g><rect x="965" y="665" width="210" height="80" rx="9" fill="#0f172a"/><rect x="965" y="665" width="210" height="80" rx="9" fill="rgba(120,53,15,.3)" stroke="#fbbf24" stroke-width="1.6"/><text x="1070" y="695" text-anchor="middle" class="title">OpenClaw VMs</text><text x="1070" y="716" text-anchor="middle" class="tiny">currently dormant</text><text x="1070" y="734" text-anchor="middle" class="port">openclaw.snapshot</text></g>
         <g><rect x="965" y="775" width="210" height="60" rx="9" fill="#0f172a"/><rect x="965" y="775" width="210" height="60" rx="9" fill="rgba(76,29,149,.4)" stroke="#a78bfa" stroke-width="1.6"/><text x="1070" y="802" text-anchor="middle" class="title">Obsidian / RAG</text><text x="1070" y="822" text-anchor="middle" class="port">:27123/:27124 + ChromaDB</text></g>
 
         <!-- host local ai box -->
-        <g><rect x="280" y="675" width="190" height="100" rx="10" fill="#0f172a"/><rect x="280" y="675" width="190" height="100" rx="10" fill="rgba(76,29,149,.4)" stroke="#a78bfa" stroke-width="1.8"/><text x="375" y="706" text-anchor="middle" class="title">host local AI</text><text x="375" y="730" text-anchor="middle" class="tiny">llama.cpp :18806</text><text x="375" y="752" text-anchor="middle" class="tiny">Ollama embed :18807</text></g>
+        <g><rect x="280" y="675" width="210" height="120" rx="10" fill="#0f172a"/><rect x="280" y="675" width="210" height="120" rx="10" fill="rgba(76,29,149,.4)" stroke="#a78bfa" stroke-width="1.8"/><text x="385" y="706" text-anchor="middle" class="title">host local AI</text><text x="385" y="730" text-anchor="middle" class="tiny">llama.cpp :18806</text><text x="385" y="752" text-anchor="middle" class="tiny">Ollama fallback :18807</text><text x="385" y="774" text-anchor="middle" class="tiny">OpenVINO NPU embed :18817</text></g>
 
         <!-- legend -->
         <g transform="translate(40,820)">
@@ -104,7 +104,7 @@
     </div>
     <div class="cards">
       <div class="info"><h3>Monitoring model</h3><ul><li>• n8n direct probes critical ports</li><li>• agentmon aggregates Docker/OpenClaw snapshots</li><li>• n8n polls agentmon for stale/degraded state</li></ul></div>
-      <div class="info"><h3>Operational endpoints</h3><ul><li>• n8n: 127.0.0.1:18808</li><li>• agentmon query/UI: 8081 / 8082</li><li>• local LLM/embed: 18806 / 18807</li></ul></div>
+      <div class="info"><h3>Operational endpoints</h3><ul><li>• n8n: 127.0.0.1:18808</li><li>• agentmon query/UI: 8081 / 8082</li><li>• local LLM/embed: 18806 / 18817</li><li>• Ollama fallback: 18807</li></ul></div>
       <div class="info"><h3>Source paths</h3><ul><li>• Swarm repo: ~/lab/swarm</li><li>• Agentmon repo: ~/lab/agentmon</li><li>• Workflows: swarm-common/n8n-workflows</li></ul></div>
     </div>
     <div class="footer">Generated as repo documentation. Open locally in a browser; no JavaScript, all SVG inline.</div>

docs/swarm-infrastructure.md

@@ -32,9 +32,10 @@ local AI/search/voice services
         +--> SearXNG :18803
         +--> Brave MCP :18802
         +--> llama.cpp :18806
-        +--> Ollama embeddings :18807
+        +--> Ollama embeddings :18807 (legacy/CPU fallback)
+        +--> OpenVINO NPU embeddings :18817
         +--> Kokoro TTS :18805
-        +--> Whisper :18811
+        +--> Whisper NPU :18816
 ```
 
 See also:
@@ -115,15 +116,16 @@ Docker services:
 - `searxng` — `:18803`, local metasearch
 - `brave-search` — `:18802`, Brave Search MCP server
 - `kokoro-tts` — `:18805`, local TTS
-- `whisper-server` — `:18811`, local transcription
+- `whisper-server-npu` — `:18816`, OpenVINO NPU local transcription
 - `n8n-agent` — `:18808`, automation
 
 Host/user services:
 
 - `llama-server.service` — `:18806`, local llama.cpp OpenAI-compatible LLM
-- `ollama.service` — `:18807`, embeddings API
+- `ollama.service` — `:18807`, legacy/CPU embeddings API fallback
+- `openvino-embeddings.service` — `:18817`, OpenVINO NPU embeddings API (`/v1/embeddings`, `/api/embed`, `/api/embeddings`)
 - `docker-health-endpoint.service` — `:18809`, read-only container health for n8n
-- `obsidian-reindex-endpoint.service` — `:18810`, Obsidian/RAG reindex trigger
+- `obsidian-reindex-endpoint.service` — `:18810`, Obsidian/RAG reindex trigger; default collection `obsidian_bge_npu` using OpenVINO NPU embeddings
 - `url-content-extractor.service` — `:18812`, YouTube/PDF/web extraction
 - `voice-memo-processor.service` — `:18813`, voice memo processing
 - `rag-embedding-health.service` — `:18814`, RAG/embedding health wrapper
@@ -142,8 +144,9 @@ Local REST API:
 RAG/vector store:
 
 - ChromaDB path: `~/.hermes/data/rag-search/chroma/`
-- Reindex state/progress: `~/.hermes/data/rag-search/obsidian_index_state.json` and `obsidian_reindex_progress.json`
-- Embeddings backend: Ollama on `:18807`, normally `nomic-embed-text`
+- Reindex state/progress: active BGE/NPU state in `~/.hermes/data/rag-search/obsidian_bge_npu_index_state.json` and `obsidian_bge_npu_reindex_progress.json`; legacy Ollama state in `obsidian_index_state.json` remains for comparison/fallback.
+- Active RAG query/reindex embedding backend: OpenVINO NPU embeddings service on `:18817`, currently `bge-base-en-v1.5-int8-ov`, collection `obsidian_bge_npu`.
+- Legacy comparison/fallback collection: `obsidian`, built with Ollama on `:18807` using `nomic-embed-text`.
 - Reindex endpoint: `POST :18810/reindex` for incremental updates, `POST :18810/reindex?full=true` for full semantic rebuilds, `GET :18810/semantic-health` to verify vectors plus a search smoke test.
 
 ## Monitoring model

scripts/docker-health-server.py

@@ -24,7 +24,7 @@ CONTAINERS = [
     "litellm-db",
     "n8n-agent",
     "searxng",
-    "whisper-server",
+    "whisper-server-npu",
 ]
 
 

scripts/obsidian-reindex-server.py

@@ -26,14 +26,20 @@ from urllib.parse import parse_qs, urlparse
 
 PORT = int(os.environ.get("PORT", 18810))
 REINDEX_TIMEOUT = int(os.environ.get("REINDEX_TIMEOUT", "1800"))
+RAG_COLLECTION = os.environ.get("RAG_COLLECTION", "obsidian").strip() or "obsidian"
+RAG_EMBED_MODEL = os.environ.get("RAG_EMBED_MODEL", "nomic-embed-text").strip() or "nomic-embed-text"
+OLLAMA_BASE_URL = (os.environ.get("OLLAMA_BASE_URL") or "http://127.0.0.1:18807").rstrip("/")
 
 REINDEX_SCRIPT = str(
     Path.home()
     / ".hermes/skills/note-taking/rag-search/scripts/reindex_obsidian.sh"
 )
-STATE_FILE = (
-    Path.home() / ".hermes/data/rag-search/obsidian_index_state.json"
-)
+STATE_FILE = Path(
+    os.environ.get("RAG_STATE_FILE")
+    or Path.home() / ".hermes/data/rag-search" / (
+        "obsidian_index_state.json" if RAG_COLLECTION == "obsidian" else f"{RAG_COLLECTION}_index_state.json"
+    )
+).expanduser()
 SEARCH_SCRIPT = str(Path.home() / ".hermes/skills/note-taking/rag-search/scripts/search.py")
 VENV_PYTHON = str(Path.home() / ".hermes/skills/note-taking/rag-search/venv/bin/python")
 
@@ -50,11 +56,16 @@ def run_reindex(full: bool = False) -> dict:
         cmd = [REINDEX_SCRIPT]
         if full:
             cmd.append("--full")
+        env = os.environ.copy()
+        env.setdefault("RAG_COLLECTION", RAG_COLLECTION)
+        env.setdefault("RAG_EMBED_MODEL", RAG_EMBED_MODEL)
+        env.setdefault("OLLAMA_BASE_URL", OLLAMA_BASE_URL)
         result = subprocess.run(
             cmd,
             capture_output=True,
             text=True,
             timeout=REINDEX_TIMEOUT,
+            env=env,
         )
         if result.returncode != 0:
             return {
@@ -97,12 +108,16 @@ def run_semantic_search(query: str, top_k: int = 5) -> dict:
     if not query:
         return {"ok": False, "error": "query is required", "results": []}
     top_k = max(1, min(int(top_k or 5), 20))
+    env = os.environ.copy()
+    env.setdefault("RAG_COLLECTION", RAG_COLLECTION)
+    env.setdefault("RAG_EMBED_MODEL", RAG_EMBED_MODEL)
+    env.setdefault("OLLAMA_BASE_URL", OLLAMA_BASE_URL)
     result = subprocess.run(
         [
             VENV_PYTHON if Path(VENV_PYTHON).exists() else sys.executable,
             SEARCH_SCRIPT,
             "--index",
-            "obsidian",
+            RAG_COLLECTION,
             "--top-k",
             str(top_k),
             "--raw",
@@ -111,6 +126,7 @@ def run_semantic_search(query: str, top_k: int = 5) -> dict:
         capture_output=True,
         text=True,
         timeout=90,
+        env=env,
     )
     if result.returncode != 0:
         return {
@@ -125,7 +141,7 @@ def run_semantic_search(query: str, top_k: int = 5) -> dict:
     return {
         "ok": True,
         "query": query,
-        "index": payload.get("index", "obsidian"),
+        "index": payload.get("index", RAG_COLLECTION),
         "top_k": top_k,
         "result_count": len(results),
         "results": results,
@@ -144,7 +160,8 @@ def semantic_health() -> dict:
                 "note_count",
                 "vector_count",
                 "collection",
-                "chroma_path",
+                "embedding_backend",
+                "embedding_model",
                 "last_full_index",
                 "last_incremental_index",
             )

scripts/openvino-embeddings-server.py

@@ -0,0 +1,236 @@
+#!/usr/bin/env python3
+"""OpenVINO GenAI embedding HTTP service for Will's local swarm stack.
+
+Default port: 18817
+Default model: OpenVINO/bge-base-en-v1.5-int8-ov, cached under ~/.cache/openvino-models/
+Default device: NPU
+
+Exposes a deliberately small compatibility surface:
+  GET  /healthz
+  GET  /api/tags                       # Ollama-ish model listing for health scripts
+  POST /api/embed                      # Ollama-ish batched embeddings
+  POST /api/embeddings                 # Ollama-ish single embedding
+  POST /v1/embeddings                  # OpenAI-compatible embeddings response
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+import threading
+import time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+from typing import Any
+
+import openvino as ov
+import openvino_genai as ovg
+
+DEFAULT_MODEL_NAME = "bge-base-en-v1.5-int8-ov"
+DEFAULT_MODEL_DIR = Path.home() / ".cache/openvino-models" / DEFAULT_MODEL_NAME
+DEFAULT_PORT = 18817
+NPU_BUSY_FILE = Path("/sys/class/accel/accel0/device/npu_busy_time_us")
+
+
+def npu_busy_time_us() -> int | None:
+    try:
+        return int(NPU_BUSY_FILE.read_text().strip())
+    except Exception:
+        return None
+
+
+class EmbeddingService:
+    def __init__(self, model_dir: Path, model_name: str, device: str, max_length: int) -> None:
+        self.model_dir = model_dir
+        self.model_name = model_name
+        self.device = device
+        self.max_length = max_length
+        self.loaded_at = time.time()
+        self.lock = threading.Lock()
+        self.embedding_dim: int | None = None
+
+        if not self.model_dir.exists():
+            raise FileNotFoundError(f"model directory not found: {self.model_dir}")
+
+        core = ov.Core()
+        self.available_devices = list(core.available_devices)
+        if self.device not in self.available_devices:
+            raise RuntimeError(f"OpenVINO device {self.device!r} unavailable; available={self.available_devices}")
+
+        # Intel NPU currently needs static shape for this embedding pipeline.
+        # batch_size=1 is intentional: multi-input requests are served by looping
+        # one text at a time, keeping the model shape acceptable to NPUW.
+        cfg = ovg.TextEmbeddingPipeline.Config()
+        cfg.max_length = int(max_length)
+        cfg.pad_to_max_length = True
+        cfg.batch_size = 1
+        self.pipeline = ovg.TextEmbeddingPipeline(self.model_dir, self.device, cfg)
+
+    def embed_one(self, text: str, *, purpose: str = "query") -> dict[str, Any]:
+        text = str(text or "")
+        if not text.strip():
+            raise ValueError("embedding input text is empty")
+        if purpose not in {"query", "document"}:
+            raise ValueError("embedding purpose must be 'query' or 'document'")
+        before = npu_busy_time_us()
+        started = time.perf_counter()
+        # TextEmbeddingPipeline is a native object; serialize calls until proven
+        # safe under concurrent NPU use. Tiny silicon clown-car avoidance clause.
+        with self.lock:
+            if purpose == "document":
+                # batch_size=1 means embed_documents must receive exactly one doc.
+                vec = self.pipeline.embed_documents([text])[0]
+            else:
+                vec = self.pipeline.embed_query(text)
+        after = npu_busy_time_us()
+        vector = [float(x) for x in vec]
+        self.embedding_dim = len(vector)
+        return {
+            "embedding": vector,
+            "dim": len(vector),
+            "purpose": purpose,
+            "duration_ms": round((time.perf_counter() - started) * 1000, 3),
+            "npu_busy_delta_us": None if before is None or after is None else after - before,
+        }
+
+    def health(self) -> dict[str, Any]:
+        return {
+            "status": "ok",
+            "service": "openvino-embeddings",
+            "model": self.model_name,
+            "model_dir": str(self.model_dir),
+            "device": self.device,
+            "available_devices": self.available_devices,
+            "embedding_dim": self.embedding_dim,
+            "max_length": self.max_length,
+            "uptime_s": round(time.time() - self.loaded_at, 3),
+            "npu_busy_time_us": npu_busy_time_us(),
+        }
+
+
+def normalize_input(value: Any) -> list[str]:
+    if isinstance(value, str):
+        return [value]
+    if isinstance(value, list):
+        texts = [str(item) for item in value]
+        if texts:
+            return texts
+    raise ValueError("input must be a non-empty string or list of strings")
+
+
+class Handler(BaseHTTPRequestHandler):
+    server_version = "OpenVINOEmbeddings/0.1"
+
+    @property
+    def svc(self) -> EmbeddingService:
+        return self.server.embedding_service  # type: ignore[attr-defined]
+
+    def do_GET(self) -> None:
+        path = self.path.split("?", 1)[0].rstrip("/") or "/"
+        if path in {"/", "/healthz", "/readyz"}:
+            self.write_json(self.svc.health())
+        elif path == "/api/tags":
+            self.write_json({"models": [{"name": self.svc.model_name, "model": self.svc.model_name}]})
+        elif path == "/v1/models":
+            self.write_json({"object": "list", "data": [{"id": self.svc.model_name, "object": "model", "owned_by": "local"}]})
+        else:
+            self.write_json({"error": "not found"}, status=404)
+
+    def do_POST(self) -> None:
+        path = self.path.split("?", 1)[0].rstrip("/") or "/"
+        try:
+            payload = self.read_json()
+            if path == "/api/embed":
+                texts = normalize_input(payload.get("input"))
+                purpose = str(payload.get("purpose") or payload.get("task") or "document")
+                results = [self.svc.embed_one(text, purpose=purpose) for text in texts]
+                self.write_json({
+                    "model": payload.get("model") or self.svc.model_name,
+                    "embeddings": [item["embedding"] for item in results],
+                    "embedding_dim": results[0]["dim"] if results else None,
+                    "purpose": purpose,
+                    "npu_busy_delta_us": sum((item.get("npu_busy_delta_us") or 0) for item in results),
+                    "durations_ms": [item["duration_ms"] for item in results],
+                })
+            elif path == "/api/embeddings":
+                text = payload.get("prompt") or payload.get("input")
+                result = self.svc.embed_one(str(text or ""), purpose="query")
+                self.write_json({
+                    "model": payload.get("model") or self.svc.model_name,
+                    "embedding": result["embedding"],
+                    "embedding_dim": result["dim"],
+                    "npu_busy_delta_us": result["npu_busy_delta_us"],
+                    "duration_ms": result["duration_ms"],
+                })
+            elif path == "/v1/embeddings":
+                texts = normalize_input(payload.get("input"))
+                purpose = str(payload.get("purpose") or payload.get("task") or "query")
+                results = [self.svc.embed_one(text, purpose=purpose) for text in texts]
+                self.write_json({
+                    "object": "list",
+                    "model": payload.get("model") or self.svc.model_name,
+                    "data": [
+                        {"object": "embedding", "index": idx, "embedding": item["embedding"]}
+                        for idx, item in enumerate(results)
+                    ],
+                    "usage": {"prompt_tokens": 0, "total_tokens": 0},
+                    "embedding_dim": results[0]["dim"] if results else None,
+                    "purpose": purpose,
+                    "npu_busy_delta_us": sum((item.get("npu_busy_delta_us") or 0) for item in results),
+                    "durations_ms": [item["duration_ms"] for item in results],
+                })
+            else:
+                self.write_json({"error": "not found"}, status=404)
+        except ValueError as exc:
+            self.write_json({"error": str(exc)}, status=400)
+        except Exception as exc:
+            self.write_json({"error": f"{type(exc).__name__}: {exc}"}, status=500)
+
+    def read_json(self) -> dict[str, Any]:
+        length = int(self.headers.get("Content-Length") or 0)
+        body = self.rfile.read(length).decode("utf-8", "replace") if length else "{}"
+        payload = json.loads(body or "{}")
+        if not isinstance(payload, dict):
+            raise ValueError("JSON body must be an object")
+        return payload
+
+    def write_json(self, payload: dict[str, Any], status: int = 200) -> None:
+        body = json.dumps(payload, ensure_ascii=False).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: str, *args: Any) -> None:  # noqa: A002 - stdlib override name
+        print(f"{self.address_string()} - {format % args}", file=sys.stderr, flush=True)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", default=os.environ.get("OPENVINO_EMBED_HOST", "0.0.0.0"))
+    parser.add_argument("--port", type=int, default=int(os.environ.get("OPENVINO_EMBED_PORT", DEFAULT_PORT)))
+    parser.add_argument("--model-dir", default=os.environ.get("OPENVINO_EMBED_MODEL_DIR", str(DEFAULT_MODEL_DIR)))
+    parser.add_argument("--model-name", default=os.environ.get("OPENVINO_EMBED_MODEL", DEFAULT_MODEL_NAME))
+    parser.add_argument("--device", default=os.environ.get("OPENVINO_EMBED_DEVICE", "NPU"))
+    parser.add_argument("--max-length", type=int, default=int(os.environ.get("OPENVINO_EMBED_MAX_LENGTH", "512")))
+    args = parser.parse_args()
+
+    service = EmbeddingService(Path(args.model_dir).expanduser(), args.model_name, args.device, args.max_length)
+    httpd = ThreadingHTTPServer((args.host, args.port), Handler)
+    httpd.embedding_service = service  # type: ignore[attr-defined]
+    print(
+        f"openvino-embeddings listening on {args.host}:{args.port} "
+        f"model={args.model_name} device={args.device}",
+        flush=True,
+    )
+    try:
+        httpd.serve_forever()
+    except KeyboardInterrupt:
+        pass
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/rag-embedding-health-server.py

@@ -51,7 +51,8 @@ class Handler(http.server.BaseHTTPRequestHandler):
 
         env = os.environ.copy()
         env.setdefault("HERMES_HOME", "/home/will/.hermes")
-        env.setdefault("OLLAMA_BASE_URL", "http://127.0.0.1:18807")
+        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")
 

scripts/voice-memo-processor.py

@@ -32,7 +32,7 @@ AUDIO_DIR = os.path.join(tempfile.gettempdir(), "voice-memo-audio")
 os.makedirs(AUDIO_DIR, exist_ok=True)
 
 # Service endpoints (from host perspective)
-WHISPER_URL = os.environ.get("WHISPER_URL", "http://127.0.0.1:18811/v1/audio/transcriptions")
+WHISPER_URL = os.environ.get("WHISPER_URL", "http://127.0.0.1:18816/v1/audio/transcriptions")
 LLM_URL = os.environ.get("LLM_URL", "http://127.0.0.1:18806/v1/chat/completions")
 KOKORO_URL = os.environ.get("KOKORO_URL", "http://127.0.0.1:18805/v1/audio/speech")
 

scripts/voice-memo-service.py

@@ -7,7 +7,7 @@ from http.server import HTTPServer, BaseHTTPRequestHandler
 from pathlib import Path
 
 PORT = int(os.environ.get("VOICE_MEMO_PORT", "18813"))
-WHISPER_URL = os.environ.get("WHISPER_BASE_URL", "http://127.0.0.1:18811")
+WHISPER_URL = os.environ.get("WHISPER_BASE_URL", "http://127.0.0.1:18816")
 LLM_URL = os.environ.get("LLAMA_CPP_BASE_URL", "http://127.0.0.1:18806")
 KOKORO_URL = os.environ.get("KOKORO_BASE_URL", "http://127.0.0.1:18805")
 TELEGRAM_BOT_TOKEN = os.environ.get("TELEGRAM_BOT_TOKEN", "")

swarm-common/obsidian-reindex-endpoint.service

@@ -0,0 +1,16 @@
+[Unit]
+Description=Obsidian Vault Reindex Endpoint
+After=network.target
+
+[Service]
+Type=simple
+ExecStart=/usr/bin/python3 /home/will/lab/swarm/scripts/obsidian-reindex-server.py
+Restart=on-failure
+RestartSec=5
+Environment=PORT=18810
+Environment=RAG_COLLECTION=obsidian_bge_npu
+Environment=RAG_EMBED_MODEL=bge-base-en-v1.5-int8-ov
+Environment=OLLAMA_BASE_URL=http://127.0.0.1:18817
+
+[Install]
+WantedBy=default.target

swarm-common/openvino-embeddings.service

@@ -0,0 +1,19 @@
+[Unit]
+Description=OpenVINO NPU Embeddings HTTP Service (port 18817)
+After=network.target
+
+[Service]
+Type=simple
+WorkingDirectory=/home/will/lab/swarm
+ExecStart=/home/will/.venvs/npu/bin/python /home/will/lab/swarm/scripts/openvino-embeddings-server.py
+Restart=on-failure
+RestartSec=5
+Environment=OPENVINO_EMBED_PORT=18817
+Environment=OPENVINO_EMBED_HOST=0.0.0.0
+Environment=OPENVINO_EMBED_DEVICE=NPU
+Environment=OPENVINO_EMBED_MODEL=bge-base-en-v1.5-int8-ov
+Environment=OPENVINO_EMBED_MODEL_DIR=/home/will/.cache/openvino-models/bge-base-en-v1.5-int8-ov
+Environment=OPENVINO_EMBED_MAX_LENGTH=512
+
+[Install]
+WantedBy=default.target

swarm-common/rag-embedding-health.service

@@ -9,6 +9,8 @@ Restart=on-failure
 RestartSec=5
 Environment=PORT=18814
 Environment=RAG_HEALTH_TIMEOUT=180
+Environment=OLLAMA_BASE_URL=http://127.0.0.1:18817
+Environment=RAG_EMBED_MODEL=bge-base-en-v1.5-int8-ov
 
 [Install]
 WantedBy=default.target

whisper-openvino-npu/Dockerfile

@@ -0,0 +1,31 @@
+FROM python:3.14-slim
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1 \
+    LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu \
+    ZE_ENABLE_ALT_DRIVERS=/usr/lib/x86_64-linux-gnu/libze_intel_npu.so.1
+
+RUN apt-get update \
+    && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+       ffmpeg libze1 ca-certificates curl \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN python -m pip install --upgrade pip \
+    && python -m pip install \
+       fastapi==0.126.0 \
+       uvicorn[standard]==0.38.0 \
+       python-multipart==0.0.22 \
+       openvino==2026.2.0 \
+       openvino-genai==2026.2.0.0 \
+       soundfile==0.13.1 \
+       numpy==2.4.6
+
+WORKDIR /app
+COPY server.py /app/server.py
+
+EXPOSE 8080
+HEALTHCHECK --interval=30s --timeout=5s --start-period=30s --retries=3 \
+  CMD curl -fsS http://localhost:8080/health >/dev/null || exit 1
+
+CMD ["uvicorn", "server:app", "--host", "0.0.0.0", "--port", "8080"]

whisper-openvino-npu/server.py

@@ -0,0 +1,147 @@
+import os
+import subprocess
+import tempfile
+import threading
+import time
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import openvino as ov
+import openvino_genai as ov_genai
+import soundfile as sf
+from fastapi import FastAPI, File, Form, UploadFile
+from fastapi.responses import JSONResponse, PlainTextResponse
+
+MODEL_DIR = Path(os.environ.get("WHISPER_MODEL_DIR", "/models/whisper-tiny-fp16-ov"))
+DEVICE = os.environ.get("WHISPER_DEVICE", "NPU")
+BUSY_PATH = Path("/sys/class/accel/accel0/device/npu_busy_time_us")
+
+app = FastAPI(title="OpenVINO NPU Whisper server", version="0.1.0")
+_lock = threading.Lock()
+_pipe = None
+_core = None
+
+
+def busy_us() -> Optional[int]:
+    try:
+        return int(BUSY_PATH.read_text().strip())
+    except Exception:
+        return None
+
+
+def get_core():
+    global _core
+    if _core is None:
+        _core = ov.Core()
+    return _core
+
+
+def get_pipe():
+    global _pipe
+    if _pipe is None:
+        _pipe = ov_genai.WhisperPipeline(str(MODEL_DIR), DEVICE)
+    return _pipe
+
+
+def load_audio(upload_path: Path) -> tuple[np.ndarray, int]:
+    """Decode arbitrary uploaded audio to mono 16 kHz float32 using ffmpeg + soundfile."""
+    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as wav:
+        wav_path = Path(wav.name)
+    try:
+        subprocess.run(
+            [
+                "ffmpeg",
+                "-nostdin",
+                "-hide_banner",
+                "-loglevel",
+                "error",
+                "-y",
+                "-i",
+                str(upload_path),
+                "-ac",
+                "1",
+                "-ar",
+                "16000",
+                "-f",
+                "wav",
+                str(wav_path),
+            ],
+            check=True,
+        )
+        audio, sr = sf.read(wav_path, dtype="float32")
+        if audio.ndim > 1:
+            audio = audio.mean(axis=1)
+        return audio, int(sr)
+    finally:
+        try:
+            wav_path.unlink()
+        except FileNotFoundError:
+            pass
+
+
+@app.get("/")
+def root():
+    return PlainTextResponse("OpenVINO NPU Whisper server\n")
+
+
+@app.get("/health")
+def health():
+    try:
+        core = get_core()
+        devices = core.available_devices
+        npu_name = core.get_property("NPU", "FULL_DEVICE_NAME") if "NPU" in devices else None
+        return {
+            "ok": "NPU" in devices,
+            "device": DEVICE,
+            "devices": devices,
+            "npu": npu_name,
+            "model_dir": str(MODEL_DIR),
+            "model_exists": MODEL_DIR.exists(),
+            "npu_busy_time_us": busy_us(),
+        }
+    except Exception as e:
+        return JSONResponse(status_code=500, content={"ok": False, "error": f"{type(e).__name__}: {e}"})
+
+
+@app.post("/v1/audio/transcriptions")
+async def transcriptions(
+    file: UploadFile = File(...),
+    model: Optional[str] = Form(default=None),
+    language: Optional[str] = Form(default=None),
+    response_format: Optional[str] = Form(default="json"),
+):
+    suffix = Path(file.filename or "audio").suffix or ".audio"
+    with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp:
+        upload_path = Path(tmp.name)
+        tmp.write(await file.read())
+
+    before = busy_us()
+    t0 = time.perf_counter()
+    try:
+        audio, sr = load_audio(upload_path)
+        # OpenVINO GenAI WhisperPipeline appears stateful for Whisper generation on
+        # this stack: reusing one pipeline produced unstable language detection on
+        # repeated short clips. Recreate per request for correctness; OpenVINO's
+        # compiled-cache path keeps warm init reasonably fast.
+        with _lock:
+            pipe = ov_genai.WhisperPipeline(str(MODEL_DIR), DEVICE)
+            result = pipe.generate(audio)
+        text = str(result).strip()
+        elapsed = time.perf_counter() - t0
+        after = busy_us()
+        if response_format == "text":
+            return PlainTextResponse(text)
+        return {
+            "text": text,
+            "duration_seconds": round(elapsed, 4),
+            "sample_rate": sr,
+            "device": DEVICE,
+            "model": model or MODEL_DIR.name,
+            "npu_busy_delta_us": None if before is None or after is None else after - before,
+        }
+    finally:
+        try:
+            upload_path.unlink()
+        except FileNotFoundError:
+            pass