Bubble bullshit

.env.example

@@ -20,10 +20,51 @@ COLLECTION_NAME=codebase
 
 # Embeddings
 EMBEDDING_MODEL=BAAI/bge-base-en-v1.5
-EMBEDDING_PROVIDER=fastembed
 # Optional repo tag attached to each payload
 REPO_NAME=workspace
 
+# ---------------------------------------------------------------------------
+# Embedding Service Configuration (Shared ONNX for scale)
+# ---------------------------------------------------------------------------
+# EMBEDDING_PROVIDER: local | remote
+#   local  = Use in-process ONNX (default, high memory per worker)
+#   remote = Use shared embedding service (recommended for scale)
+EMBEDDING_PROVIDER=local
+
+# When EMBEDDING_PROVIDER=remote, calls this service
+EMBEDDING_SERVICE_URL=http://embedding:8100
+EMBEDDING_SERVICE_TIMEOUT=60
+
+# Max concurrent ONNX inferences (local mode or in embedding service)
+# Prevents memory explosion with parallel workers
+EMBED_MAX_CONCURRENT=2
+
+# Max batch size per embed request
+EMBED_MAX_BATCH=256
+
+# ---------------------------------------------------------------------------
+# ONNX CPU Optimizations (for embedding service)
+# ---------------------------------------------------------------------------
+# ONNX_THREADS: Number of threads for intra-op parallelism
+#   0 = auto (1 per physical core), or set explicit count (e.g., 4-6)
+ONNX_THREADS=0
+
+# ONNX_DISABLE_SPINNING: Disable thread spin-wait (saves CPU cycles)
+#   0 = spinning enabled (faster, burns CPU), 1 = disabled (power efficient)
+ONNX_DISABLE_SPINNING=0
+
+# EMBED_OPTIMAL_BATCH: Internal batch size for chunking large requests
+#   Sweet spot for CPU is 32-64. Too small = overhead, too large = memory pressure
+EMBED_OPTIMAL_BATCH=32
+
+# ---------------------------------------------------------------------------
+# Embedding Model Options
+# ---------------------------------------------------------------------------
+# Model options (changing model requires re-indexing!):
+#   BAAI/bge-base-en-v1.5            - Default, solid quality (768 dim, 0.21 GB)
+#   nomic-ai/nomic-embed-text-v1.5   - Faster, outperforms BGE on MTEB (768 dim, 0.13 GB)
+#   BAAI/bge-large-en-v1.5           - Higher quality, slower (1024 dim, 0.67 GB)
+#
 # Qwen3-Embedding Feature Flag (optional, experimental)
 # Enable to use Qwen3-Embedding-0.6B instead of BGE-base (requires reindex)
 # QWEN3_EMBEDDING_ENABLED=0

deploy/helm/context-engine/templates/configmap.yaml

@@ -11,7 +11,9 @@ data:
   QDRANT_URL: {{ include "context-engine.qdrantUrl" . | quote }}
   EMBEDDING_MODEL: {{ .Values.config.embeddingModel | quote }}
   EMBEDDING_PROVIDER: {{ .Values.config.embeddingProvider | quote }}
+  EMBEDDING_SERVICE_URL: {{ .Values.config.embeddingServiceUrl | quote }}
   EMBEDDING_WARMUP: "0"
+  INDEX_WORKERS: {{ .Values.config.indexWorkers | default "4" | quote }}
 
   FASTMCP_HOST: {{ .Values.config.fastmcp.host | quote }}
   FASTMCP_PORT: {{ .Values.config.fastmcp.port | quote }}

deploy/helm/context-engine/templates/embedding-service.yaml

@@ -0,0 +1,91 @@
+{{- if eq .Values.config.embeddingProvider "remote" }}
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: {{ include "context-engine.fullname" . }}-embedding
+  labels:
+    {{- include "context-engine.labels" . | nindent 4 }}
+    app.kubernetes.io/component: embedding
+spec:
+  replicas: {{ .Values.embedding.replicas | default 2 }}
+  selector:
+    matchLabels:
+      {{- include "context-engine.selectorLabels" . | nindent 6 }}
+      app.kubernetes.io/component: embedding
+  template:
+    metadata:
+      labels:
+        {{- include "context-engine.selectorLabels" . | nindent 8 }}
+        app.kubernetes.io/component: embedding
+    spec:
+      containers:
+        - name: embedding
+          image: "{{ .Values.embedding.image.repository | default "context-engine-embedding" }}:{{ .Values.embedding.image.tag | default "latest" }}"
+          ports:
+            - containerPort: 8100
+          env:
+            - name: EMBEDDING_MODEL
+              value: {{ .Values.config.embeddingModel | quote }}
+            - name: EMBED_MAX_CONCURRENT
+              value: {{ .Values.embedding.maxConcurrent | default "2" | quote }}
+            - name: EMBED_OPTIMAL_BATCH
+              value: {{ .Values.embedding.optimalBatch | default "32" | quote }}
+            - name: ONNX_THREADS
+              value: {{ .Values.embedding.onnxThreads | default "4" | quote }}
+            - name: ONNX_DISABLE_SPINNING
+              value: "1"
+            - name: OMP_NUM_THREADS
+              value: {{ .Values.embedding.onnxThreads | default "4" | quote }}
+            - name: MKL_NUM_THREADS
+              value: {{ .Values.embedding.onnxThreads | default "4" | quote }}
+          resources:
+            {{- toYaml .Values.embedding.resources | nindent 12 }}
+          readinessProbe:
+            httpGet:
+              path: /health
+              port: 8100
+            initialDelaySeconds: 30
+            periodSeconds: 10
+          livenessProbe:
+            httpGet:
+              path: /health
+              port: 8100
+            initialDelaySeconds: 60
+            periodSeconds: 30
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: embedding
+  labels:
+    {{- include "context-engine.labels" . | nindent 4 }}
+spec:
+  selector:
+    {{- include "context-engine.selectorLabels" . | nindent 4 }}
+    app.kubernetes.io/component: embedding
+  ports:
+    - port: 8100
+      targetPort: 8100
+---
+{{- if .Values.embedding.autoscaling.enabled }}
+apiVersion: autoscaling/v2
+kind: HorizontalPodAutoscaler
+metadata:
+  name: {{ include "context-engine.fullname" . }}-embedding
+spec:
+  scaleTargetRef:
+    apiVersion: apps/v1
+    kind: Deployment
+    name: {{ include "context-engine.fullname" . }}-embedding
+  minReplicas: {{ .Values.embedding.autoscaling.minReplicas | default 2 }}
+  maxReplicas: {{ .Values.embedding.autoscaling.maxReplicas | default 10 }}
+  metrics:
+    - type: Resource
+      resource:
+        name: cpu
+        target:
+          type: Utilization
+          averageUtilization: {{ .Values.embedding.autoscaling.targetCPU | default 70 }}
+{{- end }}
+{{- end }}
+

deploy/helm/context-engine/values.yaml

@@ -110,6 +110,37 @@ qdrant:
     initialDelaySeconds: 5
     periodSeconds: 5
 
+# -----------------------------------------------------------------------------
+# Embedding Service Configuration (shared ONNX model)
+# -----------------------------------------------------------------------------
+embedding:
+  # -- Number of replicas
+  replicas: 2
+  # -- Image configuration
+  image:
+    repository: context-engine-embedding
+    tag: latest
+  # -- Max concurrent embeddings per replica
+  maxConcurrent: 2
+  # -- Optimal batch size for CPU cache
+  optimalBatch: 32
+  # -- ONNX threads per replica
+  onnxThreads: 4
+  # -- Resource requests and limits
+  resources:
+    requests:
+      cpu: "2"
+      memory: 4Gi
+    limits:
+      cpu: "4"
+      memory: 6Gi
+  # -- Autoscaling configuration
+  autoscaling:
+    enabled: true
+    minReplicas: 2
+    maxReplicas: 10
+    targetCPU: 70
+
 # -----------------------------------------------------------------------------
 # MCP Indexer HTTP Configuration
 # -----------------------------------------------------------------------------
@@ -463,10 +494,14 @@ config:
   # -- Qdrant URL (auto-generated if not set)
   qdrantUrl: ""
   # -- Embedding model
-  embeddingModel: BAAI/bge-base-en-v1.5
-  # -- Embedding provider
-  embeddingProvider: fastembed
-
+  embeddingModel: nomic-ai/nomic-embed-text-v1.5
+  # -- Embedding provider (remote = shared service, fastembed = local)
+  embeddingProvider: remote
+  # -- Embedding service URL (when provider=remote)
+  embeddingServiceUrl: http://embedding:8100
+  # -- Index workers (parallel file processing)
+  indexWorkers: 4
+
   # -- FastMCP settings
   fastmcp:
     host: "0.0.0.0"

deploy/kubernetes/configmap.yaml

@@ -15,8 +15,10 @@ data:
   CTX_SUMMARY_CHARS: '0'
   CURRENT_REPO: ''
   DECODER_MAX_TOKENS: '4000'
-  EMBEDDING_MODEL: BAAI/bge-base-en-v1.5
-  EMBEDDING_PROVIDER: fastembed
+  EMBEDDING_MODEL: nomic-ai/nomic-embed-text-v1.5
+  EMBEDDING_PROVIDER: remote
+  EMBEDDING_SERVICE_URL: http://embedding:8100
+  INDEX_WORKERS: "4"
   EMBEDDING_WARMUP: '0'
   FASTMCP_HOST: 0.0.0.0
   FASTMCP_HTTP_HEALTH_PORT: '18002'

deploy/kubernetes/embedding-service.yaml

@@ -0,0 +1,106 @@
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: embedding-service
+  namespace: context-engine
+  labels:
+    app: embedding-service
+spec:
+  replicas: 2
+  selector:
+    matchLabels:
+      app: embedding-service
+  template:
+    metadata:
+      labels:
+        app: embedding-service
+    spec:
+      containers:
+        - name: embedding
+          image: context-engine-embedding:latest
+          ports:
+            - containerPort: 8100
+          env:
+            - name: EMBEDDING_MODEL
+              valueFrom:
+                configMapKeyRef:
+                  name: context-engine-config
+                  key: EMBEDDING_MODEL
+            - name: EMBED_MAX_CONCURRENT
+              value: "2"
+            - name: EMBED_MAX_BATCH
+              value: "256"
+            - name: EMBED_OPTIMAL_BATCH
+              value: "32"
+            - name: ONNX_THREADS
+              value: "4"
+            - name: ONNX_DISABLE_SPINNING
+              value: "1"
+            - name: OMP_NUM_THREADS
+              value: "4"
+            - name: MKL_NUM_THREADS
+              value: "4"
+            - name: HF_HOME
+              value: /tmp/huggingface
+            - name: FASTEMBED_CACHE_PATH
+              value: /tmp/fastembed
+          resources:
+            requests:
+              memory: "4Gi"
+              cpu: "2"
+            limits:
+              memory: "6Gi"
+              cpu: "4"
+          readinessProbe:
+            httpGet:
+              path: /health
+              port: 8100
+            initialDelaySeconds: 30
+            periodSeconds: 10
+          livenessProbe:
+            httpGet:
+              path: /health
+              port: 8100
+            initialDelaySeconds: 60
+            periodSeconds: 30
+          volumeMounts:
+            - name: embedding-cache
+              mountPath: /tmp/huggingface
+      volumes:
+        - name: embedding-cache
+          emptyDir:
+            sizeLimit: 2Gi
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: embedding
+  namespace: context-engine
+spec:
+  selector:
+    app: embedding-service
+  ports:
+    - port: 8100
+      targetPort: 8100
+  type: ClusterIP
+---
+apiVersion: autoscaling/v2
+kind: HorizontalPodAutoscaler
+metadata:
+  name: embedding-service-hpa
+  namespace: context-engine
+spec:
+  scaleTargetRef:
+    apiVersion: apps/v1
+    kind: Deployment
+    name: embedding-service
+  minReplicas: 2
+  maxReplicas: 10
+  metrics:
+    - type: Resource
+      resource:
+        name: cpu
+        target:
+          type: Utilization
+          averageUtilization: 70
+

deploy/kubernetes/indexer-services.yaml

@@ -71,6 +71,21 @@ spec:
             configMapKeyRef:
               name: context-engine-config
               key: EMBEDDING_MODEL
+        - name: EMBEDDING_PROVIDER
+          valueFrom:
+            configMapKeyRef:
+              name: context-engine-config
+              key: EMBEDDING_PROVIDER
+        - name: EMBEDDING_SERVICE_URL
+          valueFrom:
+            configMapKeyRef:
+              name: context-engine-config
+              key: EMBEDDING_SERVICE_URL
+        - name: INDEX_WORKERS
+          valueFrom:
+            configMapKeyRef:
+              name: context-engine-config
+              key: INDEX_WORKERS
         - name: HF_HOME
           value: /work/models/hf-cache
         - name: XDG_CACHE_HOME
@@ -209,6 +224,21 @@ spec:
             configMapKeyRef:
               name: context-engine-config
               key: EMBEDDING_MODEL
+        - name: EMBEDDING_PROVIDER
+          valueFrom:
+            configMapKeyRef:
+              name: context-engine-config
+              key: EMBEDDING_PROVIDER
+        - name: EMBEDDING_SERVICE_URL
+          valueFrom:
+            configMapKeyRef:
+              name: context-engine-config
+              key: EMBEDDING_SERVICE_URL
+        - name: INDEX_WORKERS
+          valueFrom:
+            configMapKeyRef:
+              name: context-engine-config
+              key: INDEX_WORKERS
         - name: HF_HOME
           value: /work/models/hf-cache
         - name: XDG_CACHE_HOME