diff --git a/AGENTS.md b/AGENTS.md
index 62f3e98..35cb411 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -57,45 +57,44 @@ The commit message should be structured as follows:
 
 ### E2E Test
 
-- **Version scope**:
-  - E2E tests cover only `vX.Y.Z` (release) and `vX.Y.Z-rc.N` (release candidate) version formats.
-  - Other version formats (e.g. dev builds, custom tags) are out of scope and should not be tested in E2E.
-
-- **Do not test resource specifications**:
-  - Do not validate individual fields of the YAML file declaring the resource (resource spec).
-  - Instead, create the resource and verify that its status reaches the expected state.
-
-- **Assume fully controlled cluster**:
-  - Do not check if components are already installed.
-  - Assume the cluster is fully controlled by the test and installed components are safe to overwrite or delete.
-
-- **Test suite layout**:
-  - Split tests by purpose under `test/e2e`, for example `test/e2e/performance` and `test/e2e/quality`.
-  - In each directory, define shared Ginkgo configuration (labels, timeouts, common hooks) in `suite_test.go`, and keep scenarios in separate `*_test.go` files.
-  - Shared configuration values must come from the `test/utils/settings` package instead of hard-coded constants in test files.
-
-- **Environment variable management**:
-  - Manage all E2E environment variables centrally in `test/e2e/envs/env_vars.go`.
-  - When a new environment variable is required:
-    - Add it to the `envVars` slice with default value, description, category, and type.
-    - Expose it via public variables (for example `TestModel`, `HFToken`) and access it only through those variables.
-    - Do not call `os.Getenv` directly in test code.
-  - Keep the documentation consistent: changes must pass the `validateEnvVars()` check.
-
-- **Resource templates and settings**:
-  - Manage Kubernetes resource specifications for Gateway, InferenceService, Jobs, and similar resources as Go templates (`.yaml.tmpl`) under `test/config/**`.
-  - Tests must read template paths and default values from constants in `test/utils/settings/constants.go`.
-  - When adding a new benchmark or performance test Job:
-    - Add the template file under an appropriate `test/config/<domain>` subdirectory.
-    - Define the corresponding path and default parameters in the `settings` package.
-
-- **Utility reuse**:
-  - Implement all cluster manipulation logic (namespace creation, Gateway create/delete, Heimdall install/uninstall, InferenceService(Template) create/delete, etc.) in the `test/utils` package and call only those helpers from tests.
-  - Follow this pattern for scenario flow:
-    - `BeforeAll`: create namespace → install Gateway → install Heimdall → create InferenceServiceTemplates → create InferenceServices → wait until they are Ready.
-    - `AfterAll`: if `envs.SkipCleanup` is `false`, clean up the above resources in reverse order.
-    - `It(...)`: render the Job template → create the Job with `kubectl create -f -` → wait for completion with `kubectl wait` → collect logs and perform domain-specific assertions.
-
-- **Makefile and workflow integration**:
-  - Provide separate Make targets per test purpose (for example `e2e-performance`, `e2e-quality`) so that CI can run them independently.
-  - GitHub Actions and other workflows should invoke these targets directly, and new test categories should follow the same pattern when adding additional targets and workflows.
+See [`test/AGENTS.md`](test/AGENTS.md).
+
+## Agent Self-Improvement
+
+After completing any non-trivial task, evaluate whether the work involved:
+- A recurring pattern that will likely appear again in future tasks, or
+- A mistake that was corrected through user feedback, or
+- A design decision that required deliberate reasoning to reach the right answer.
+
+If any of the above applies, **record it in the most relevant `AGENTS.md`** before closing the task — this file for general patterns, [`test/AGENTS.md`](test/AGENTS.md) for test-specific patterns, [`deploy/helm/AGENTS.md`](deploy/helm/AGENTS.md) for Helm chart patterns, and [`website/AGENTS.md`](website/AGENTS.md) for documentation patterns. Entries should be concise, actionable, and placed under the most relevant existing section. If no section fits, create one.
+
+The goal is to make every repeated task faster and every repeated mistake impossible.
+
+### Creating Sub-directory AGENTS.md Files
+
+When a directory accumulates enough domain-specific rules to warrant separation, create a dedicated `AGENTS.md` in that directory. Follow this checklist:
+
+1. **Create `AGENTS.md`** in the target directory with a header that links back to this root file:
+   ```markdown
+   # <Domain> — Agent Rules
+
+   Rules specific to the `<dir>/` directory. General contribution guidelines are in the root [`AGENTS.md`](/AGENTS.md).
+   ```
+
+2. **Create a `CLAUDE.md` symlink** pointing to `AGENTS.md` in the same directory. Cursor reads `CLAUDE.md` as context; the symlink ensures both tools see the same content:
+   ```shell
+   cd <dir> && ln -s AGENTS.md CLAUDE.md
+   ```
+
+3. **Move the relevant sections** from the root `AGENTS.md` (or parent `AGENTS.md`) into the new file. Replace the moved content in the parent with a one-line reference:
+   ```markdown
+   ### E2E Test
+
+   See [`test/AGENTS.md`](test/AGENTS.md).
+   ```
+
+4. **Update the Agent Self-Improvement section** in the parent to mention the new file as a recording target.
+
+## Helm Charts
+
+See [`deploy/helm/AGENTS.md`](deploy/helm/AGENTS.md) for design principles and chart development rules.
diff --git a/Makefile b/Makefile
index 14e9cbd..dcd9ef2 100644
--- a/Makefile
+++ b/Makefile
@@ -19,7 +19,11 @@ help: ## Display this help.
 
 .PHONY: helm-lint
 helm-lint: ## Lint Helm charts.
-	@helm lint ./deploy/helm/*
+	@for chart in ./deploy/helm/*; do \
+	  if [ -d "$$chart" ] && [ -f "$$chart/Chart.yaml" ]; then \
+	    helm lint "$$chart"; \
+	  fi; \
+	done
 
 .PHONY: helm-docs
 helm-docs: ## Generate Helm chart documentation.
diff --git a/deploy/helm/AGENTS.md b/deploy/helm/AGENTS.md
new file mode 100644
index 0000000..e0c386f
--- /dev/null
+++ b/deploy/helm/AGENTS.md
@@ -0,0 +1,130 @@
+# Helm Charts — Agent Rules
+
+Rules specific to the `deploy/helm/` directory. General contribution guidelines are in the root [`AGENTS.md`](/AGENTS.md).
+
+## Design Principles
+
+### Minimum Necessary Complexity
+
+- **Do not add configuration options, fields, or abstractions for hypothetical future use cases.** Only add what the current task concretely requires.
+- Before introducing a new value field, ask: "Is there a real, current use case that cannot be handled without it?" If the answer is no, omit the field and handle the edge case through documentation instead.
+- Example: when considering whether to add a `minio.externalHost` field to support cross-namespace MinIO, the right answer was to document that users can point `loki.storage.s3.endpoint` to the external host directly — no new field needed.
+
+### Documentation over Code for Edge Cases
+
+- When a behavior difference only arises in a non-default, edge-case configuration, prefer documenting the workaround over adding a dedicated code path or configuration key.
+- Reserve code changes for cases where the default path is broken or the workaround is genuinely error-prone.
+
+### Reject Designs Before They Are Built
+
+- If an initial design is heading in the wrong direction (e.g., standalone prerequisites instead of sub-chart dependencies, `enabled: false` defaults, nested config instead of top-level sections), raise the issue and redesign before writing code. Retrofitting a wrong structure is always more costly.
+
+## Helm Chart Development
+
+### Sub-chart Integration
+
+- **All infrastructure components belong as sub-chart dependencies** of `moai-inference-framework`. Do not design them as standalone prerequisites that users install separately.
+- **Enablement convention**: Every sub-chart dependency must have both a `condition:` entry in `Chart.yaml` AND `enabled: true` in the default `values.yaml`. Setting `enabled: false` as the default breaks the "install everything in one chart" philosophy. Follow the same pattern as existing components (`keda`, `lws`, `odin`, etc.).
+
+  ```yaml
+  # Chart.yaml — always add condition: and use the official repository
+  - name: vector
+    version: 0.39.0
+    repository: https://helm.vector.dev
+    condition: vector.enabled
+
+  # values.yaml — always default to true
+  vector:
+    enabled: true
+  ```
+
+- **Official repositories**: Always use the chart's official upstream repository, not a mirror.
+  - loki: `https://grafana.github.io/helm-charts`
+  - vector: `https://helm.vector.dev`
+  - minio: `https://charts.min.io`
+
+### Dynamic Service Name References
+
+- **Do not use `fullnameOverride`** to fix service names. Instead, build references using `.Release.Name` so that names are always consistent with whatever release name the user chooses.
+
+  ```yaml
+  # templates/grafana/datasource-loki.yaml
+  url: http://{{ .Release.Name }}-loki-gateway.{{ include "common.names.namespace" . }}.svc.cluster.local
+
+  # templates/loki/credentials.yaml
+  BUCKET_HOST: {{ printf "%s-minio" .Release.Name | quote }}
+  ```
+
+- In sub-chart `customConfig` values rendered through `tpl`, use `{{ .Release.Name }}` directly — it is evaluated by the sub-chart's `tpl` call and resolves to the parent release name.
+
+  ```yaml
+  # values.yaml (vector customConfig) — .Release.Name evaluated by tpl
+  endpoint: "http://{{ .Release.Name }}-loki-gateway"
+  ```
+
+### Separation of Concerns in values.yaml
+
+- **Large infrastructure components must be top-level sections**, not nested under their consumers. For example, MinIO configuration belongs at `minio:`, not at `loki.minio:`. This allows MinIO to be independently enabled/disabled and reused by other components in the future.
+
+### MinIO Provisioning Pattern
+
+- Use the `minio/minio` chart (`https://charts.min.io`), not the bitnami chart.
+- Create buckets, users, and policies directly via the chart's top-level `buckets`, `users`, and `policies` fields (not under a `provisioning` key).
+- Create a **dedicated user per consuming service** with a policy scoped to only its bucket — do not use root credentials for service-to-service access.
+
+  ```yaml
+  minio:
+    policies:
+      - name: loki
+        statements:
+          - resources: ["arn:aws:s3:::loki/*"]
+            effect: Allow
+            actions: ["s3:*"]
+    users:
+      - accessKey: loki
+        secretKey: "loki123!"
+        policy: loki
+    buckets:
+      - name: loki
+  ```
+
+- Templates that read MinIO credentials must reference the `users` array directly:
+
+  ```yaml
+  # credentials.yaml
+  stringData:
+    AWS_ACCESS_KEY_ID:     {{ (index .Values.minio.users 0).accessKey | quote }}
+    AWS_SECRET_ACCESS_KEY: {{ (index .Values.minio.users 0).secretKey | quote }}
+  ```
+
+### Helm `tpl` Passthrough — Vector Label Syntax
+
+- The vector chart renders `customConfig` through Helm's `tpl` function (`{{ tpl (toYaml .Values.customConfig) . | indent 4 }}`). This means any `{{ }}` expression in `customConfig` is evaluated as a Go template at render time.
+- To pass **Vector's own field-template syntax** (`{{ field }}`) through `tpl` without evaluation, use Go raw string literals:
+
+  ```yaml
+  # values.yaml — correct
+  labels:
+    namespace: "{{`{{ namespace }}`}}"
+
+  # values.yaml — WRONG: tpl evaluates {{ namespace }} as a Go template function
+  labels:
+    namespace: "{{ namespace }}"
+  ```
+
+- **Before using `customConfig` with any sub-chart, always verify whether the chart applies `tpl` to it** by running `helm pull <chart> --version <ver> --untar` and inspecting the ConfigMap template.
+
+### YAML Anchors
+
+- **Do not use YAML anchors at the root level of `values.yaml`** (e.g., `_defaults: &defaults`). Helm treats unknown root-level keys as invalid and may emit warnings or errors. Instead, duplicate shared configuration explicitly for each component.
+
+### MIF Pod Label Keys
+
+When filtering or labeling logs, metrics, or other signals by MIF-specific pod attributes, use these label keys:
+
+| Concept           | Label key                    | Example value       |
+| :---------------- | :--------------------------- | :------------------ |
+| Pool              | `mif.moreh.io/pool`          | `heimdall`          |
+| Role              | `mif.moreh.io/role`          | `prefill`, `decode` |
+| App name          | `app.kubernetes.io/name`     | `vllm`              |
+| Inference service | `app.kubernetes.io/instance` | `llama-3-2-1b`      |
diff --git a/deploy/helm/CLAUDE.md b/deploy/helm/CLAUDE.md
new file mode 120000
index 0000000..47dc3e3
--- /dev/null
+++ b/deploy/helm/CLAUDE.md
@@ -0,0 +1 @@
+AGENTS.md
\ No newline at end of file
diff --git a/deploy/helm/moai-inference-framework/Chart.lock b/deploy/helm/moai-inference-framework/Chart.lock
index cb00edb..a899860 100644
--- a/deploy/helm/moai-inference-framework/Chart.lock
+++ b/deploy/helm/moai-inference-framework/Chart.lock
@@ -23,5 +23,14 @@ dependencies:
 - name: node-feature-discovery
   repository: oci://registry.k8s.io/nfd/charts
   version: 0.18.3
-digest: sha256:d7f75e788dca4192775595637ec123afa390e09eebcaef9e9c0e40ff46c23e23
-generated: "2026-02-19T16:14:15.495286+09:00"
+- name: minio
+  repository: https://charts.min.io
+  version: 5.4.0
+- name: loki
+  repository: https://grafana.github.io/helm-charts
+  version: 6.30.0
+- name: vector
+  repository: https://helm.vector.dev
+  version: 0.39.0
+digest: sha256:85af11696c630ed9ac9ef85a7a18c8b821187a76e949e535019fc5b91d929ee8
+generated: "2026-02-20T18:51:13.630416372+09:00"
diff --git a/deploy/helm/moai-inference-framework/Chart.yaml b/deploy/helm/moai-inference-framework/Chart.yaml
index 43e39f0..538d7b0 100644
--- a/deploy/helm/moai-inference-framework/Chart.yaml
+++ b/deploy/helm/moai-inference-framework/Chart.yaml
@@ -42,3 +42,15 @@ dependencies:
     version: 0.18.3
     repository: oci://registry.k8s.io/nfd/charts
     condition: nfd.enabled
+  - name: minio
+    version: 5.4.0
+    repository: https://charts.min.io
+    condition: minio.enabled
+  - name: loki
+    version: 6.30.0
+    repository: https://grafana.github.io/helm-charts
+    condition: loki.enabled
+  - name: vector
+    version: 0.39.0
+    repository: https://helm.vector.dev
+    condition: vector.enabled
diff --git a/deploy/helm/moai-inference-framework/README.md b/deploy/helm/moai-inference-framework/README.md
index 8abc6f9..bebc015 100644
--- a/deploy/helm/moai-inference-framework/README.md
+++ b/deploy/helm/moai-inference-framework/README.md
@@ -17,7 +17,10 @@ Moreh Inference Framework
 
 | Repository | Name | Version |
 |------------|------|---------|
+| https://charts.min.io | minio | 5.4.0 |
+| https://grafana.github.io/helm-charts | loki | 6.30.0 |
 | https://helm.mittwald.de | replicator(kubernetes-replicator) | 2.12.2 |
+| https://helm.vector.dev | vector | 0.39.0 |
 | https://kedacore.github.io/charts | keda | 2.18.0 |
 | https://moreh-dev.github.io/helm-charts | odin | v0.6.0 |
 | https://moreh-dev.github.io/helm-charts | odin-crd | v0.6.0 |
@@ -52,7 +55,74 @@ Moreh Inference Framework
 | global | object | `{"imagePullSecrets":[]}` | global values are shared across all sub-charts if the value's key matches. |
 | global.imagePullSecrets | list | `[]` | Image pull secrets. |
 | keda.enabled | bool | `true` | Enable kedacore/keda. Set to false if already deployed. |
+| loki.backend.extraArgs[0] | string | `"-config.expand-env=true"` |  |
+| loki.backend.extraEnvFrom[0].secretRef.name | string | `"loki-bucket"` |  |
+| loki.backend.extraEnvFrom[1].configMapRef.name | string | `"loki-bucket"` |  |
+| loki.backend.nodeSelector."node-role.kubernetes.io/control-plane" | string | `""` |  |
+| loki.backend.persistence.volumeClaimsEnabled | bool | `false` |  |
+| loki.backend.replicas | int | `1` |  |
+| loki.enabled | bool | `true` | Enable grafana/loki. |
+| loki.gateway.extraArgs[0] | string | `"-config.expand-env=true"` |  |
+| loki.gateway.extraEnvFrom[0].secretRef.name | string | `"loki-bucket"` |  |
+| loki.gateway.extraEnvFrom[1].configMapRef.name | string | `"loki-bucket"` |  |
+| loki.gateway.nodeSelector."node-role.kubernetes.io/control-plane" | string | `""` |  |
+| loki.gateway.replicas | int | `1` |  |
+| loki.loki.auth_enabled | bool | `false` |  |
+| loki.loki.commonConfig.replication_factor | int | `1` |  |
+| loki.loki.image.tag | string | `"3.5.1"` |  |
+| loki.loki.schemaConfig.configs[0].from | string | `"2024-06-24"` |  |
+| loki.loki.schemaConfig.configs[0].index.period | string | `"24h"` |  |
+| loki.loki.schemaConfig.configs[0].index.prefix | string | `"loki_index_"` |  |
+| loki.loki.schemaConfig.configs[0].object_store | string | `"s3"` |  |
+| loki.loki.schemaConfig.configs[0].schema | string | `"v13"` |  |
+| loki.loki.schemaConfig.configs[0].store | string | `"tsdb"` |  |
+| loki.loki.storage.bucketNames.admin | string | `"loki"` |  |
+| loki.loki.storage.bucketNames.chunks | string | `"loki"` |  |
+| loki.loki.storage.bucketNames.ruler | string | `"loki"` |  |
+| loki.loki.storage.s3.accessKeyId | string | `"${AWS_ACCESS_KEY_ID}"` |  |
+| loki.loki.storage.s3.endpoint | string | `"http://${BUCKET_HOST}:${BUCKET_PORT}"` |  |
+| loki.loki.storage.s3.region | string | `"${BUCKET_REGION}"` |  |
+| loki.loki.storage.s3.s3ForcePathStyle | bool | `true` |  |
+| loki.loki.storage.s3.secretAccessKey | string | `"${AWS_SECRET_ACCESS_KEY}"` |  |
+| loki.loki.storage_config.tsdb_shipper.active_index_directory | string | `"/var/loki/tsdb-index"` |  |
+| loki.loki.storage_config.tsdb_shipper.cache_location | string | `"/var/loki/tsdb-cache"` |  |
+| loki.loki.storage_config.tsdb_shipper.cache_ttl | string | `"168h"` |  |
+| loki.loki.structuredConfig.compactor.delete_request_store | string | `"s3"` |  |
+| loki.loki.structuredConfig.compactor.retention_enabled | bool | `true` |  |
+| loki.loki.structuredConfig.limits_config.ingestion_burst_size_mb | int | `60` |  |
+| loki.loki.structuredConfig.limits_config.ingestion_rate_mb | int | `30` |  |
+| loki.loki.structuredConfig.limits_config.max_entries_limit_per_query | int | `50000` |  |
+| loki.loki.structuredConfig.limits_config.max_query_series | int | `10000` |  |
+| loki.loki.structuredConfig.limits_config.per_stream_rate_limit | string | `"30MB"` |  |
+| loki.loki.structuredConfig.limits_config.per_stream_rate_limit_burst | string | `"60MB"` |  |
+| loki.loki.structuredConfig.limits_config.retention_period | string | `"2160h"` |  |
+| loki.loki.structuredConfig.limits_config.split_queries_by_interval | string | `"24h"` |  |
+| loki.read.extraArgs[0] | string | `"-config.expand-env=true"` |  |
+| loki.read.extraEnvFrom[0].secretRef.name | string | `"loki-bucket"` |  |
+| loki.read.extraEnvFrom[1].configMapRef.name | string | `"loki-bucket"` |  |
+| loki.read.nodeSelector."node-role.kubernetes.io/control-plane" | string | `""` |  |
+| loki.read.replicas | int | `1` |  |
+| loki.write.extraArgs[0] | string | `"-config.expand-env=true"` |  |
+| loki.write.extraEnvFrom[0].secretRef.name | string | `"loki-bucket"` |  |
+| loki.write.extraEnvFrom[1].configMapRef.name | string | `"loki-bucket"` |  |
+| loki.write.nodeSelector."node-role.kubernetes.io/control-plane" | string | `""` |  |
+| loki.write.persistence.volumeClaimsEnabled | bool | `false` |  |
+| loki.write.replicas | int | `1` |  |
 | lws.enabled | bool | `true` | Enable kubernetes-sigs/lws. Set to false if already deployed. |
+| minio.buckets[0].name | string | `"loki"` |  |
+| minio.enabled | bool | `true` | Enable minio/minio as the S3-compatible object storage backend for Loki. Set to false if MinIO is already deployed; in that case, configure loki storage to point to the existing MinIO service. |
+| minio.mode | string | `"standalone"` |  |
+| minio.persistence.enabled | bool | `false` |  |
+| minio.policies[0].name | string | `"loki"` |  |
+| minio.policies[0].statements[0].actions[0] | string | `"s3:*"` |  |
+| minio.policies[0].statements[0].effect | string | `"Allow"` |  |
+| minio.policies[0].statements[0].resources[0] | string | `"arn:aws:s3:::loki/*"` |  |
+| minio.resources.requests.memory | string | `"2Gi"` |  |
+| minio.rootPassword | string | `"minio123!"` | MinIO root password. Override with a strong password in production. |
+| minio.rootUser | string | `"minio"` | MinIO root user. |
+| minio.users[0].accessKey | string | `"loki"` |  |
+| minio.users[0].policy | string | `"loki"` |  |
+| minio.users[0].secretKey | string | `"loki123!"` | Password for the loki MinIO user. Override with a strong password in production. |
 | nameOverride | string | `""` | Chart name override. |
 | namespaceOverride | string | `""` | Namespace override. |
 | nfd.enabled | bool | `true` | Enable kubernetes-sigs/node-feature-discovery. Set to false if already deployed. |
@@ -81,6 +151,40 @@ Moreh Inference Framework
 | prometheus-stack.thanosRuler.enabled | bool | `false` |  |
 | prometheus-stack.windowsMonitoring.enabled | bool | `false` |  |
 | replicator.enabled | bool | `true` | Enable mittwald/kubernetes-replicator. Set to false if already deployed. |
+| vector.customConfig.api.address | string | `"0.0.0.0:8686"` |  |
+| vector.customConfig.api.enabled | bool | `true` |  |
+| vector.customConfig.data_dir | string | `"/vector-data"` |  |
+| vector.customConfig.sinks.loki.encoding.codec | string | `"json"` |  |
+| vector.customConfig.sinks.loki.endpoint | string | `"http://{{ .Release.Name }}-loki-gateway"` |  |
+| vector.customConfig.sinks.loki.inputs[0] | string | `"mif_log_transform"` |  |
+| vector.customConfig.sinks.loki.labels.app | string | `"{{`{{ app }}`}}"` |  |
+| vector.customConfig.sinks.loki.labels.inference_service | string | `"{{`{{ inference_service }}`}}"` |  |
+| vector.customConfig.sinks.loki.labels.level | string | `"{{`{{ level }}`}}"` |  |
+| vector.customConfig.sinks.loki.labels.namespace | string | `"{{`{{ namespace }}`}}"` |  |
+| vector.customConfig.sinks.loki.labels.node_name | string | `"{{`{{ node_name }}`}}"` |  |
+| vector.customConfig.sinks.loki.labels.pool_name | string | `"{{`{{ pool_name }}`}}"` |  |
+| vector.customConfig.sinks.loki.labels.role | string | `"{{`{{ role }}`}}"` |  |
+| vector.customConfig.sinks.loki.type | string | `"loki"` |  |
+| vector.customConfig.sources.mif_logs.extra_label_selector | string | `"mif.moreh.io/log.collect=true"` |  |
+| vector.customConfig.sources.mif_logs.type | string | `"kubernetes_logs"` |  |
+| vector.customConfig.transforms.mif_log_transform.inputs[0] | string | `"mif_logs"` |  |
+| vector.customConfig.transforms.mif_log_transform.source | string | `".namespace          = .kubernetes.pod_namespace\n.node_name          = \"$VECTOR_SELF_NODE_NAME\"\n.app                = get(.kubernetes.pod_labels, [\"app.kubernetes.io/name\"])      ?? \"\"\n.inference_service  = get(.kubernetes.pod_labels, [\"app.kubernetes.io/instance\"])  ?? \"\"\n.pool_name          = get(.kubernetes.pod_labels, [\"mif.moreh.io/pool\"])           ?? \"\"\n.role               = get(.kubernetes.pod_labels, [\"mif.moreh.io/role\"])           ?? \"\"\n\nlog_format = get(.kubernetes.pod_labels, [\"mif.moreh.io/log.format\"]) ?? \"\"\n\nif log_format == \"json\" {\n  structured, err = parse_json(.message)\n  if err == null {\n    . = merge!(., structured)\n    msg, err = get(., [\"msg\"])\n    if err == null {\n      .message = msg\n      del(.msg)\n    }\n    time, err = get(., [\"time\"])\n    if err == null {\n      .timestamp = time\n      del(.time)\n    }\n  }\n}\n\ndel(.file)\ndel(.source_type)\ndel(.stream)\ndel(.kubernetes)\n"` |  |
+| vector.customConfig.transforms.mif_log_transform.type | string | `"remap"` |  |
+| vector.enabled | bool | `true` | Enable vector/vector as a DaemonSet log collector. |
+| vector.role | string | `"Agent"` |  |
+| vector.tolerations[0].effect | string | `"NoExecute"` |  |
+| vector.tolerations[0].key | string | `"node.kubernetes.io/unschedulable"` |  |
+| vector.tolerations[0].operator | string | `"Exists"` |  |
+| vector.tolerations[0].tolerationSeconds | int | `5` |  |
+| vector.tolerations[1].effect | string | `"NoSchedule"` |  |
+| vector.tolerations[1].key | string | `"node-role.kubernetes.io/compute"` |  |
+| vector.tolerations[1].operator | string | `"Equal"` |  |
+| vector.tolerations[1].value | string | `"true"` |  |
+| vector.tolerations[2].effect | string | `"NoSchedule"` |  |
+| vector.tolerations[2].key | string | `"amd.com/gpu"` |  |
+| vector.tolerations[2].operator | string | `"Exists"` |  |
+| vector.updateStrategy.rollingUpdate.maxUnavailable | int | `10` |  |
+| vector.updateStrategy.type | string | `"RollingUpdate"` |  |
 
 ----------------------------------------------
 Autogenerated from chart metadata using [helm-docs v1.14.2](https://github.com/norwoodj/helm-docs/releases/v1.14.2)
diff --git a/deploy/helm/moai-inference-framework/charts/loki-6.30.0.tgz b/deploy/helm/moai-inference-framework/charts/loki-6.30.0.tgz
new file mode 100644
index 0000000..b5e7322
Binary files /dev/null and b/deploy/helm/moai-inference-framework/charts/loki-6.30.0.tgz differ
diff --git a/deploy/helm/moai-inference-framework/charts/minio-5.4.0.tgz b/deploy/helm/moai-inference-framework/charts/minio-5.4.0.tgz
new file mode 100644
index 0000000..22f8d73
Binary files /dev/null and b/deploy/helm/moai-inference-framework/charts/minio-5.4.0.tgz differ
diff --git a/deploy/helm/moai-inference-framework/charts/vector-0.39.0.tgz b/deploy/helm/moai-inference-framework/charts/vector-0.39.0.tgz
new file mode 100644
index 0000000..9c71305
Binary files /dev/null and b/deploy/helm/moai-inference-framework/charts/vector-0.39.0.tgz differ
diff --git a/deploy/helm/moai-inference-framework/templates/grafana/datasource-loki.yaml b/deploy/helm/moai-inference-framework/templates/grafana/datasource-loki.yaml
new file mode 100644
index 0000000..87852f5
--- /dev/null
+++ b/deploy/helm/moai-inference-framework/templates/grafana/datasource-loki.yaml
@@ -0,0 +1,26 @@
+{{- $ps := index .Values "prometheus-stack" }}
+{{- if and .Values.loki.enabled $ps.enabled $ps.grafana.enabled }}
+---
+# Grafana datasource ConfigMap for Loki.
+# The grafana-sidecar discovers ConfigMaps labelled grafana_datasource=1 and
+# provisions them as datasources automatically.
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: {{ include "common.names.name" . }}-datasource-loki
+  namespace: {{ include "common.names.namespace" . }}
+  labels:
+    grafana_datasource: "1"
+    {{- include "mif.labels" . | nindent 4 }}
+data:
+  loki-datasource.yaml: |
+    apiVersion: 1
+    datasources:
+      - name: Loki
+        type: loki
+        access: proxy
+        url: http://{{ .Release.Name }}-loki-gateway.{{ include "common.names.namespace" . }}.svc.cluster.local
+        isDefault: false
+        jsonData:
+          maxLines: 5000
+{{- end }}
diff --git a/deploy/helm/moai-inference-framework/templates/loki/credentials.yaml b/deploy/helm/moai-inference-framework/templates/loki/credentials.yaml
new file mode 100644
index 0000000..a026a04
--- /dev/null
+++ b/deploy/helm/moai-inference-framework/templates/loki/credentials.yaml
@@ -0,0 +1,29 @@
+{{- if .Values.loki.enabled }}
+---
+# ConfigMap consumed by Loki components via extraEnvFrom.
+# The values are referenced in loki.storage.s3 as ${BUCKET_HOST} etc.,
+# which Loki resolves at startup because each component runs with -config.expand-env=true.
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: loki-bucket
+  namespace: {{ include "common.names.namespace" . }}
+  labels:
+    {{- include "mif.labels" . | nindent 4 }}
+data:
+  BUCKET_HOST:   {{ printf "%s-minio" .Release.Name | quote }}
+  BUCKET_PORT:   "9000"
+  BUCKET_REGION: ""
+  BUCKET_NAME:   "loki"
+---
+apiVersion: v1
+kind: Secret
+metadata:
+  name: loki-bucket
+  namespace: {{ include "common.names.namespace" . }}
+  labels:
+    {{- include "mif.labels" . | nindent 4 }}
+stringData:
+  AWS_ACCESS_KEY_ID:     {{ (index .Values.minio.users 0).accessKey | quote }}
+  AWS_SECRET_ACCESS_KEY: {{ (index .Values.minio.users 0).secretKey | quote }}
+{{- end }}
diff --git a/deploy/helm/moai-inference-framework/values.yaml b/deploy/helm/moai-inference-framework/values.yaml
index d9c19e3..59ded47 100644
--- a/deploy/helm/moai-inference-framework/values.yaml
+++ b/deploy/helm/moai-inference-framework/values.yaml
@@ -86,7 +86,7 @@ replicator:
 nfd:
   # -- Enable kubernetes-sigs/node-feature-discovery. Set to false if already deployed.
   enabled: true
-  
+
   worker:
     # -- NFD Worker Tolerations to allow NFD workers to deploy to GPU nodes
     tolerations:
@@ -104,6 +104,245 @@ nfd:
         operator: "Exists"
         effect: "NoSchedule"
 
+minio:
+  # -- Enable minio/minio as the S3-compatible object storage backend for Loki.
+  # Set to false if MinIO is already deployed; in that case, configure loki storage
+  # to point to the existing MinIO service.
+  enabled: true
+
+  mode: standalone
+
+  # -- MinIO root user.
+  rootUser: minio
+  # -- MinIO root password. Override with a strong password in production.
+  rootPassword: "minio123!"
+
+  persistence:
+    enabled: false
+
+  # S3 policy granting the loki user access to the loki bucket only.
+  policies:
+    - name: loki
+      statements:
+        - resources:
+            - "arn:aws:s3:::loki/*"
+          effect: Allow
+          actions:
+            - "s3:*"
+
+  # Dedicated loki user with restricted S3 access (not the root user).
+  # AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY in the loki-bucket Secret are
+  # set from this user's credentials.
+  users:
+    - accessKey: loki
+      # -- Password for the loki MinIO user. Override with a strong password in production.
+      secretKey: "loki123!"
+      policy: loki
+
+  buckets:
+    - name: loki
+
+  resources:
+    requests:
+      memory: 2Gi
+
+loki:
+  # -- Enable grafana/loki.
+  enabled: true
+
+  loki:
+    auth_enabled: false
+    commonConfig:
+      replication_factor: 1
+    image:
+      tag: "3.5.1"
+    schemaConfig:
+      configs:
+        - from: "2024-06-24"
+          store: tsdb
+          object_store: s3
+          schema: v13
+          index:
+            prefix: loki_index_
+            period: 24h
+    storage:
+      bucketNames:
+        chunks: loki
+        ruler: loki
+        admin: loki
+      s3:
+        s3ForcePathStyle: true
+        # ${...} references are resolved by Loki at startup via -config.expand-env=true.
+        endpoint: "http://${BUCKET_HOST}:${BUCKET_PORT}"
+        region: "${BUCKET_REGION}"
+        accessKeyId: "${AWS_ACCESS_KEY_ID}"
+        secretAccessKey: "${AWS_SECRET_ACCESS_KEY}"
+    storage_config:
+      tsdb_shipper:
+        active_index_directory: /var/loki/tsdb-index
+        cache_location: /var/loki/tsdb-cache
+        cache_ttl: 168h
+    structuredConfig:
+      compactor:
+        retention_enabled: true
+        delete_request_store: s3
+      limits_config:
+        max_entries_limit_per_query: 50000
+        split_queries_by_interval: 24h
+        ingestion_rate_mb: 30
+        ingestion_burst_size_mb: 60
+        per_stream_rate_limit: 30MB
+        per_stream_rate_limit_burst: 60MB
+        max_query_series: 10000
+        retention_period: 2160h # 90 days
+
+  # Each component runs with -config.expand-env=true so that ${BUCKET_HOST} etc.
+  # in the Loki config are expanded from the loki-bucket ConfigMap and Secret.
+  gateway:
+    replicas: 1
+    extraArgs:
+      - -config.expand-env=true
+    extraEnvFrom:
+      - secretRef:
+          name: loki-bucket
+      - configMapRef:
+          name: loki-bucket
+    nodeSelector:
+      node-role.kubernetes.io/control-plane: ""
+
+  read:
+    replicas: 1
+    extraArgs:
+      - -config.expand-env=true
+    extraEnvFrom:
+      - secretRef:
+          name: loki-bucket
+      - configMapRef:
+          name: loki-bucket
+    nodeSelector:
+      node-role.kubernetes.io/control-plane: ""
+
+  write:
+    replicas: 1
+    extraArgs:
+      - -config.expand-env=true
+    extraEnvFrom:
+      - secretRef:
+          name: loki-bucket
+      - configMapRef:
+          name: loki-bucket
+    nodeSelector:
+      node-role.kubernetes.io/control-plane: ""
+    persistence:
+      volumeClaimsEnabled: false
+
+  backend:
+    replicas: 1
+    extraArgs:
+      - -config.expand-env=true
+    extraEnvFrom:
+      - secretRef:
+          name: loki-bucket
+      - configMapRef:
+          name: loki-bucket
+    nodeSelector:
+      node-role.kubernetes.io/control-plane: ""
+    persistence:
+      volumeClaimsEnabled: false
+
+vector:
+  # -- Enable vector/vector as a DaemonSet log collector.
+  enabled: true
+
+  role: Agent
+
+  tolerations:
+    - key: node.kubernetes.io/unschedulable
+      operator: Exists
+      effect: NoExecute
+      tolerationSeconds: 5
+    - key: node-role.kubernetes.io/compute
+      operator: Equal
+      value: "true"
+      effect: NoSchedule
+    - key: amd.com/gpu
+      operator: Exists
+      effect: NoSchedule
+
+  updateStrategy:
+    type: RollingUpdate
+    rollingUpdate:
+      maxUnavailable: 10
+
+  customConfig:
+    data_dir: /vector-data
+    api:
+      enabled: true
+      address: "0.0.0.0:8686"
+    sources:
+      mif_logs:
+        type: kubernetes_logs
+        # Only collect logs from pods that explicitly opt in to MIF log collection.
+        extra_label_selector: "mif.moreh.io/log.collect=true"
+    transforms:
+      mif_log_transform:
+        type: remap
+        inputs:
+          - mif_logs
+        source: |
+          .namespace          = .kubernetes.pod_namespace
+          .node_name          = "$VECTOR_SELF_NODE_NAME"
+          .app                = get(.kubernetes.pod_labels, ["app.kubernetes.io/name"])      ?? ""
+          .inference_service  = get(.kubernetes.pod_labels, ["app.kubernetes.io/instance"])  ?? ""
+          .pool_name          = get(.kubernetes.pod_labels, ["mif.moreh.io/pool"])           ?? ""
+          .role               = get(.kubernetes.pod_labels, ["mif.moreh.io/role"])           ?? ""
+
+          log_format = get(.kubernetes.pod_labels, ["mif.moreh.io/log.format"]) ?? ""
+
+          if log_format == "json" {
+            structured, err = parse_json(.message)
+            if err == null {
+              . = merge!(., structured)
+              msg, err = get(., ["msg"])
+              if err == null {
+                .message = msg
+                del(.msg)
+              }
+              time, err = get(., ["time"])
+              if err == null {
+                .timestamp = time
+                del(.time)
+              }
+            }
+          }
+
+          del(.file)
+          del(.source_type)
+          del(.stream)
+          del(.kubernetes)
+    sinks:
+      loki:
+        type: loki
+        # {{ .Release.Name }} is evaluated by the vector chart's tpl rendering,
+        # producing the correct release-prefixed Loki gateway service name.
+        endpoint: "http://{{ .Release.Name }}-loki-gateway"
+        inputs:
+          - mif_log_transform
+        encoding:
+          codec: json
+        labels:
+          # {{`{{ field }}`}} is Go raw-string escaping required because the
+          # Vector chart renders customConfig through Helm's tpl function.
+          # tpl sees {{`{{ namespace }}`}} → evaluates raw string → outputs {{ namespace }}
+          # which is Vector's own field-template syntax.
+          namespace: "{{`{{ namespace }}`}}"
+          inference_service: "{{`{{ inference_service }}`}}"
+          pool_name: "{{`{{ pool_name }}`}}"
+          role: "{{`{{ role }}`}}"
+          node_name: "{{`{{ node_name }}`}}"
+          app: "{{`{{ app }}`}}"
+          level: "{{`{{ level }}`}}"
+
 ecrTokenRefresher:
   # -- Enable ECR token refresher.
   enabled: true
diff --git a/test/AGENTS.md b/test/AGENTS.md
new file mode 100644
index 0000000..1bcb608
--- /dev/null
+++ b/test/AGENTS.md
@@ -0,0 +1,48 @@
+# Test — Agent Rules
+
+Rules specific to the `test/` directory. General contribution guidelines are in the root [`AGENTS.md`](/AGENTS.md).
+
+## E2E Test
+
+- **Version scope**:
+  - E2E tests cover only `vX.Y.Z` (release) and `vX.Y.Z-rc.N` (release candidate) version formats.
+  - Other version formats (e.g. dev builds, custom tags) are out of scope and should not be tested in E2E.
+
+- **Do not test resource specifications**:
+  - Do not validate individual fields of the YAML file declaring the resource (resource spec).
+  - Instead, create the resource and verify that its status reaches the expected state.
+
+- **Assume fully controlled cluster**:
+  - Do not check if components are already installed.
+  - Assume the cluster is fully controlled by the test and installed components are safe to overwrite or delete.
+
+- **Test suite layout**:
+  - Split tests by purpose under `test/e2e`, for example `test/e2e/performance` and `test/e2e/quality`.
+  - In each directory, define shared Ginkgo configuration (labels, timeouts, common hooks) in `suite_test.go`, and keep scenarios in separate `*_test.go` files.
+  - Shared configuration values must come from the `test/utils/settings` package instead of hard-coded constants in test files.
+
+- **Environment variable management**:
+  - Manage all E2E environment variables centrally in `test/e2e/envs/env_vars.go`.
+  - When a new environment variable is required:
+    - Add it to the `envVars` slice with default value, description, category, and type.
+    - Expose it via public variables (for example `TestModel`, `HFToken`) and access it only through those variables.
+    - Do not call `os.Getenv` directly in test code.
+  - Keep the documentation consistent: changes must pass the `validateEnvVars()` check.
+
+- **Resource templates and settings**:
+  - Manage Kubernetes resource specifications for Gateway, InferenceService, Jobs, and similar resources as Go templates (`.yaml.tmpl`) under `test/config/**`.
+  - Tests must read template paths and default values from constants in `test/utils/settings/constants.go`.
+  - When adding a new benchmark or performance test Job:
+    - Add the template file under an appropriate `test/config/<domain>` subdirectory.
+    - Define the corresponding path and default parameters in the `settings` package.
+
+- **Utility reuse**:
+  - Implement all cluster manipulation logic (namespace creation, Gateway create/delete, Heimdall install/uninstall, InferenceService(Template) create/delete, etc.) in the `test/utils` package and call only those helpers from tests.
+  - Follow this pattern for scenario flow:
+    - `BeforeAll`: create namespace → install Gateway → install Heimdall → create InferenceServiceTemplates → create InferenceServices → wait until they are Ready.
+    - `AfterAll`: if `envs.SkipCleanup` is `false`, clean up the above resources in reverse order.
+    - `It(...)`: render the Job template → create the Job with `kubectl create -f -` → wait for completion with `kubectl wait` → collect logs and perform domain-specific assertions.
+
+- **Makefile and workflow integration**:
+  - Provide separate Make targets per test purpose (for example `e2e-performance`, `e2e-quality`) so that CI can run them independently.
+  - GitHub Actions and other workflows should invoke these targets directly, and new test categories should follow the same pattern when adding additional targets and workflows.
diff --git a/test/CLAUDE.md b/test/CLAUDE.md
new file mode 120000
index 0000000..47dc3e3
--- /dev/null
+++ b/test/CLAUDE.md
@@ -0,0 +1 @@
+AGENTS.md
\ No newline at end of file
diff --git a/website/AGENTS.md b/website/AGENTS.md
index 4135368..f6655d4 100644
--- a/website/AGENTS.md
+++ b/website/AGENTS.md
@@ -1,6 +1,6 @@
 # Website (Docusaurus) — Agent Rules
 
-This file defines rules for contributors and automation agents working in `website/`.
+Rules specific to the `website/` directory. General contribution guidelines are in the root [`AGENTS.md`](/AGENTS.md).
 
 ## 1. Structure & Metadata
 
@@ -30,8 +30,8 @@ This file defines rules for contributors and automation agents working in `websi
 https://docusaurus.io/docs/markdown-features/tabs
 
 ```mdx
-import Tabs from '@theme/Tabs';
-import TabItem from '@theme/TabItem';
+import Tabs from "@theme/Tabs";
+import TabItem from "@theme/TabItem";
 
 <Tabs groupId="fruits" queryString>
   <TabItem value="apple" label="Apple" default>
@@ -57,6 +57,13 @@ image:
 
 - title: ` ```<language> title="<title>"`
 - highlight: ` ```<language> {1,4-6}`
+- **Expected output**: Blocks showing command output must always specify both a language type and a title on the same opening fence. Use `shell` for terminal output:
+  ````mdx
+  ```shell title="Expected output (one pod per node, all `Running`)"
+  NAME           READY   STATUS    RESTARTS   AGE
+  vector-xxxxx   1/1     Running   0          2m
+  ```
+  ````
 - **Variables**:
   - Format as `<variableName>` (camelCase, no quotes).
   - Highlight lines containing variables in code blocks (e.g., ` ```yaml {2} `).
@@ -96,3 +103,9 @@ This is a warning
 ## 4. Content Guidelines
 
 - **Inference Deployment**: When documenting deployment of inference services (e.g., vLLM, SGLang), instructions MUST use the `InferenceService` resource with a preset.
+
+- **No duplicate installation steps**: Operation or feature docs must not repeat the values file example or `helm upgrade` command that already appears in `getting-started/prerequisites.mdx`. Instead, link directly to the relevant section:
+  ```mdx
+  See [Prerequisites](../getting-started/prerequisites.mdx#moai-inference-framework) for the required values and install command.
+  ```
+  Duplication causes the two pages to diverge whenever the chart version or values change.
diff --git a/website/CLAUDE.md b/website/CLAUDE.md
new file mode 120000
index 0000000..47dc3e3
--- /dev/null
+++ b/website/CLAUDE.md
@@ -0,0 +1 @@
+AGENTS.md
\ No newline at end of file
diff --git a/website/docs/getting-started/prerequisites.mdx b/website/docs/getting-started/prerequisites.mdx
index 67293a5..f1f36da 100644
--- a/website/docs/getting-started/prerequisites.mdx
+++ b/website/docs/getting-started/prerequisites.mdx
@@ -4,8 +4,8 @@ sidebar_label: Prerequisites
 sidebar_position: 1
 ---
 
-import Tabs from '@theme/Tabs';
-import TabItem from '@theme/TabItem';
+import Tabs from "@theme/Tabs";
+import TabItem from "@theme/TabItem";
 
 This document introduces the prerequisites for the MoAI Inference Framework and provides instructions on how to install them.
 
@@ -74,7 +74,9 @@ ecrTokenRefresher:
     secretAccessKey: <AWS_SECRET_ACCESS_KEY>
 ```
 
-In addition, if dependencies such as keda, kube-prometheus-stack, lws are already installed in your cluster, you should skip their installation by setting the corresponding values to `false` in the `moai-inference-framework-values.yaml` file. Refer to [moai-inference-framework README](https://github.com/moreh-dev/mif/tree/main/deploy/helm/moai-inference-framework) to see the full list of dependencies.
+:::info
+If dependencies such as `keda`, `kube-prometheus-stack`, or `lws` are already installed in your cluster, skip their installation by setting the corresponding values to `false`. Refer to [moai-inference-framework README](https://github.com/moreh-dev/mif/tree/main/deploy/helm/moai-inference-framework) for the full list of dependencies.
+:::
 
 Then, deploy the `moai-inference-framework` chart using the following command:
 
@@ -136,7 +138,7 @@ deviceConfig:
   spec:
     driver:
       enable: true
-      version: '6.4.3'
+      version: "6.4.3"
       blacklist: true
       image: <registry>/amdgpu-driver
       imageRegistrySecret:
@@ -279,7 +281,7 @@ spec:
       app.kubernetes.io/instance: rdma-shared-device-plugin
   updateStrategy:
     rollingUpdate:
-      maxUnavailable: '30%'
+      maxUnavailable: "30%"
   template:
     metadata:
       labels:
@@ -415,8 +417,8 @@ Create a `istiod-values.yaml` file and install the Istio control plane.
 ```yaml title="istiod-values.yaml"
 pilot:
   env:
-    PILOT_ENABLE_ALPHA_GATEWAY_API: 'true'
-    ENABLE_GATEWAY_API_INFERENCE_EXTENSION: 'true'
+    PILOT_ENABLE_ALPHA_GATEWAY_API: "true"
+    ENABLE_GATEWAY_API_INFERENCE_EXTENSION: "true"
 ```
 
 ```shell
diff --git a/website/docs/operations/log-collection.mdx b/website/docs/operations/log-collection.mdx
new file mode 100644
index 0000000..f3a5bfb
--- /dev/null
+++ b/website/docs/operations/log-collection.mdx
@@ -0,0 +1,263 @@
+---
+title: Log Collection (Loki + Vector)
+sidebar_label: Log Collection
+sidebar_position: 3
+---
+
+This document explains how to enable centralized log collection for the MoAI Inference Framework using [Loki](https://grafana.com/oss/loki/) (log aggregation) and [Vector](https://vector.dev/) (log collection agent).
+
+## Overview
+
+```
+Inference Service Pods
+        │  (container logs)
+        ▼
+   Vector DaemonSet           ← runs on every node
+        │  transforms + labels logs
+        │  (namespace, inference_service, pool_name, …)
+        ▼
+   Loki Gateway               ← stores logs in MinIO (S3)
+        │
+        ▼
+   Grafana (LogQL)            ← search and visualize logs
+```
+
+### Labels available for log search
+
+| Label               | Source                                                                             | Example value           |
+| :------------------ | :--------------------------------------------------------------------------------- | :---------------------- |
+| `namespace`         | `kubernetes.pod_namespace`                                                         | `default`               |
+| `inference_service` | pod label `app.kubernetes.io/instance`                                             | `llama-3-2-1b`          |
+| `pool_name`         | pod label `mif.moreh.io/pool`                                                      | `heimdall`              |
+| `role`              | pod label `mif.moreh.io/role`                                                      | `prefill`, `decode`     |
+| `app`               | pod label `app.kubernetes.io/name`                                                 | `vllm`                  |
+| `node_name`         | `VECTOR_SELF_NODE_NAME` env var (injected by Vector)                               | `gpu-node-01`           |
+| `level`             | parsed from JSON log field `level` (pods with `mif.moreh.io/log.format=json` only) | `info`, `warn`, `error` |
+
+---
+
+## Prerequisites
+
+- The `moai-inference-framework` Helm chart installed (or being installed).
+
+:::info
+MinIO, Loki, and Vector are all **enabled by default** in the `moai-inference-framework` chart. No additional configuration is required to get started.
+:::
+
+---
+
+## Installation
+
+Log collection is installed as part of the `moai-inference-framework` Helm chart. See [Prerequisites](../getting-started/prerequisites.mdx#moai-inference-framework) for the required values and install command.
+
+---
+
+## Verifying the installation
+
+Check that all Loki components are running.
+
+```shell
+kubectl get pods -n mif -l app.kubernetes.io/name=loki
+```
+
+```shell title="Expected output (all pods Running)"
+NAME                           READY   STATUS    RESTARTS   AGE
+loki-backend-0                 1/1     Running   0          2m
+loki-gateway-xxxxxxxxx-xxxxx   1/1     Running   0          2m
+loki-read-xxxxxxxxx-xxxxx      1/1     Running   0          2m
+loki-write-0                   1/1     Running   0          2m
+```
+
+Check that Vector is running on all nodes.
+
+```shell
+kubectl get pods -n mif -l app.kubernetes.io/name=vector
+```
+
+```shell title="Expected output (one pod per node, all Running)"
+NAME           READY   STATUS    RESTARTS   AGE
+vector-xxxxx   1/1     Running   0          2m
+vector-yyyyy   1/1     Running   0          2m
+```
+
+Check Vector logs to confirm it is shipping to Loki without errors.
+
+```shell
+kubectl logs -n mif -l app.kubernetes.io/name=vector --tail=50
+```
+
+---
+
+## Enabling log collection for a pod
+
+Vector collects logs only from pods that explicitly opt in. Two pod labels control this behavior.
+
+### Opt-in label
+
+Add the `mif.moreh.io/log.collect=true` label to a pod to include its logs in Vector's collection. Pods without this label are ignored entirely.
+
+```yaml
+metadata:
+  labels:
+    mif.moreh.io/log.collect: "true"
+```
+
+### Log format label
+
+Add the `mif.moreh.io/log.format=json` label to enable structured JSON log parsing for a pod. When set, Vector parses each log line as JSON and promotes the following fields:
+
+| JSON field | Mapped to             |
+| :--------- | :-------------------- |
+| `msg`      | `message`             |
+| `time`     | `timestamp`           |
+| `level`    | `level` (Loki label)  |
+| others     | merged into the event |
+
+Without this label, the log line is forwarded as-is without any JSON parsing.
+
+```yaml
+metadata:
+  labels:
+    mif.moreh.io/log.collect: "true"
+    mif.moreh.io/log.format: "json"
+```
+
+:::info
+The `level` Loki label is only populated for pods with `mif.moreh.io/log.format=json`. For plain-text pods, `level` remains empty.
+:::
+
+---
+
+## Searching logs in Grafana
+
+Open Grafana → **Explore** → select the **Loki** datasource.
+
+### By namespace
+
+```promql
+{namespace="default"}
+```
+
+### By inference service name
+
+```promql
+{inference_service="llama-3-2-1b"}
+```
+
+### By pool name
+
+```promql
+{pool_name="heimdall"}
+```
+
+### By role
+
+```promql
+{role="decode"}
+```
+
+### Combined filter
+
+```promql
+{namespace="default", inference_service="llama-3-2-1b", role="prefill"} |= "error"
+```
+
+### Filter by log level
+
+```promql
+{namespace="default", level="error"}
+```
+
+---
+
+## Architecture details
+
+### Loki
+
+| Property          | Value                                          |
+| :---------------- | :--------------------------------------------- |
+| Helm chart        | `moreh/loki` v6.30.0 (upstream grafana/loki)   |
+| App version       | 3.5.1                                          |
+| Storage backend   | S3 (MinIO), TSDB index                         |
+| Retention         | 90 days (2160 h)                               |
+| Ingestion limit   | 30 MB/s, 60 MB burst                           |
+| Max entries/query | 50 000                                         |
+| Deployment        | Distributed (gateway / read / write / backend) |
+
+### Vector
+
+| Property    | Value                                                                     |
+| :---------- | :------------------------------------------------------------------------ |
+| Helm chart  | `vector/vector` v0.39.0                                                   |
+| Deployment  | DaemonSet (Agent mode, one pod per node)                                  |
+| Log source  | Pods labelled `mif.moreh.io/log.collect=true` (`kubernetes_logs`)         |
+| Log format  | JSON parsing applied only to pods labelled `mif.moreh.io/log.format=json` |
+| Tolerations | unschedulable, compute, `amd.com/gpu`                                     |
+
+### MinIO
+
+| Property         | Value                                                        |
+| :--------------- | :----------------------------------------------------------- |
+| Helm chart       | `minio/minio` v5.4.0                                         |
+| Mode             | Standalone                                                   |
+| Bucket           | `loki` (created via post-install Job on startup)             |
+| Loki credentials | Dedicated `loki` user with S3 policy scoped to `loki` bucket |
+| Resources        | 1 CPU / 2 Gi memory (requests), 1.5 CPU / 3 Gi (limits)      |
+| Persistence      | emptyDir (ephemeral by default)                              |
+| Deployment       | Single pod                                                   |
+
+### Component naming
+
+Service names are derived from the Helm release name. With the default release name `mif`:
+
+| Service      | Name (same-namespace access) |
+| :----------- | :--------------------------- |
+| MinIO        | `mif-minio`                  |
+| Loki gateway | `mif-loki-gateway`           |
+| Loki read    | `mif-loki-read`              |
+| Loki write   | `mif-loki-write`             |
+
+Vector connects to Loki using the release-prefixed service name since all components are co-located in the same namespace.
+
+---
+
+## Using an external MinIO
+
+If MinIO is already deployed outside this chart, set `minio.enabled: false`. The `loki-bucket` ConfigMap and Secret are still generated from `minio.users[0].*` values regardless of whether the sub-chart is enabled. The `BUCKET_HOST` is derived from the Helm release name (`<release>-minio`).
+
+**Same namespace** — if the existing MinIO service name matches `<release>-minio`, no additional configuration is needed. Otherwise, set `minio.fullnameOverride` to override the name used as `BUCKET_HOST`:
+
+```yaml title="moai-inference-framework-values.yaml" {3,5-6}
+minio:
+  enabled: false
+  fullnameOverride: <existing-minio-service-name>
+  users:
+    - accessKey: <lokiAccessKey>
+      secretKey: <lokiSecretKey>
+      policy: loki
+```
+
+**Different namespace** — set `minio.fullnameOverride` to the FQDN so that Loki can resolve it cross-namespace:
+
+```yaml title="moai-inference-framework-values.yaml" {5-6}
+minio:
+  enabled: false
+  fullnameOverride: minio.minio.svc.cluster.local
+  users:
+    - accessKey: <lokiAccessKey>
+      secretKey: <lokiSecretKey>
+      policy: loki
+```
+
+---
+
+## Disabling log collection
+
+```yaml title="moai-inference-framework-values.yaml"
+minio:
+  enabled: false
+loki:
+  enabled: false
+vector:
+  enabled: false
+```
diff --git a/website/docusaurus.config.ts b/website/docusaurus.config.ts
index 477c0cb..e2f43fe 100644
--- a/website/docusaurus.config.ts
+++ b/website/docusaurus.config.ts
@@ -82,7 +82,7 @@ const config: Config = {
       ],
     },
     prism: {
-      additionalLanguages: ['bash', 'toml', 'yaml'],
+      additionalLanguages: ['bash', 'toml', 'yaml', 'promql'],
       theme: themes.nightOwlLight,
       darkTheme: themes.vsDark,
     },
diff --git a/website/package-lock.json b/website/package-lock.json
index fa9aa1f..e032ba0 100644
--- a/website/package-lock.json
+++ b/website/package-lock.json
@@ -278,6 +278,7 @@
       "resolved": "https://registry.npmjs.org/@algolia/client-search/-/client-search-5.48.0.tgz",
       "integrity": "sha512-RB9bKgYTVUiOcEb5bOcZ169jiiVW811dCsJoLT19DcbbFmU4QaK0ghSTssij35QBQ3SCOitXOUrHcGgNVwS7sQ==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@algolia/client-common": "5.48.0",
         "@algolia/requester-browser-xhr": "5.48.0",
@@ -444,6 +445,7 @@
       "resolved": "https://registry.npmjs.org/@babel/core/-/core-7.29.0.tgz",
       "integrity": "sha512-CGOfOJqWjg2qW/Mb6zNsDm+u5vFQ8DxXfbM09z69p5Z6+mE1ikP2jUXw+j42Pf1XTYED2Rni5f95npYeuwMDQA==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@babel/code-frame": "^7.29.0",
         "@babel/generator": "^7.29.0",
@@ -2479,6 +2481,7 @@
         }
       ],
       "license": "MIT",
+      "peer": true,
       "engines": {
         "node": ">=18"
       },
@@ -2501,6 +2504,7 @@
         }
       ],
       "license": "MIT",
+      "peer": true,
       "engines": {
         "node": ">=18"
       }
@@ -2610,6 +2614,7 @@
       "resolved": "https://registry.npmjs.org/postcss-selector-parser/-/postcss-selector-parser-7.1.1.tgz",
       "integrity": "sha512-orRsuYpJVw8LdAwqqLykBj9ecS5/cRHlI5+nvTo8LcCKmzDmqVORXtOIYEEQuL9D4BxtA1lm5isAqzQZCoQ6Eg==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "cssesc": "^3.0.0",
         "util-deprecate": "^1.0.2"
@@ -3031,6 +3036,7 @@
       "resolved": "https://registry.npmjs.org/postcss-selector-parser/-/postcss-selector-parser-7.1.1.tgz",
       "integrity": "sha512-orRsuYpJVw8LdAwqqLykBj9ecS5/cRHlI5+nvTo8LcCKmzDmqVORXtOIYEEQuL9D4BxtA1lm5isAqzQZCoQ6Eg==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "cssesc": "^3.0.0",
         "util-deprecate": "^1.0.2"
@@ -3830,6 +3836,7 @@
       "resolved": "https://registry.npmjs.org/@docusaurus/core/-/core-3.9.2.tgz",
       "integrity": "sha512-HbjwKeC+pHUFBfLMNzuSjqFE/58+rLVKmOU3lxQrpsxLBOGosYco/Q0GduBb0/jEMRiyEqjNT/01rRdOMWq5pw==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@docusaurus/babel": "3.9.2",
         "@docusaurus/bundler": "3.9.2",
@@ -4011,6 +4018,7 @@
       "resolved": "https://registry.npmjs.org/@docusaurus/plugin-content-docs/-/plugin-content-docs-3.9.2.tgz",
       "integrity": "sha512-C5wZsGuKTY8jEYsqdxhhFOe1ZDjH0uIYJ9T/jebHwkyxqnr4wW0jTkB72OMqNjsoQRcb0JN3PcSeTwFlVgzCZg==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@docusaurus/core": "3.9.2",
         "@docusaurus/logger": "3.9.2",
@@ -5047,6 +5055,7 @@
       "resolved": "https://registry.npmjs.org/@mdx-js/react/-/react-3.1.1.tgz",
       "integrity": "sha512-f++rKLQgUVYDAtECQ6fn/is15GkEH9+nZPM3MS0RcxVqoTfawHvDlSCH7JbMhAM6uJ32v3eXLvLmLvjGu7PTQw==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@types/mdx": "^2.0.0"
       },
@@ -5519,6 +5528,7 @@
       "resolved": "https://registry.npmjs.org/@svgr/core/-/core-8.1.0.tgz",
       "integrity": "sha512-8QqtOQT5ACVlmsvKOJNEaWmRPmcojMOzCz4Hs2BGG/toAp/K38LcsMRyLp349glq5AzJbCEeimEoxaX6v/fLrA==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@babel/core": "^7.21.3",
         "@svgr/babel-preset": "8.1.0",
@@ -6140,6 +6150,7 @@
       "resolved": "https://registry.npmjs.org/@types/react/-/react-19.2.13.tgz",
       "integrity": "sha512-KkiJeU6VbYbUOp5ITMIc7kBfqlYkKA5KhEHVrGMmUUMt7NeaZg65ojdPk+FtNrBAOXNVM5QM72jnADjM+XVRAQ==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "csstype": "^3.2.2"
       }
@@ -6488,6 +6499,7 @@
       "resolved": "https://registry.npmjs.org/acorn/-/acorn-8.15.0.tgz",
       "integrity": "sha512-NZyJarBfL7nWwIq+FDL6Zp/yHEhePMNnnJ0y3qfieCrmNvYct8uvtiV41UvlSe6apAfk0fY1FbWx+NwfmpvtTg==",
       "license": "MIT",
+      "peer": true,
       "bin": {
         "acorn": "bin/acorn"
       },
@@ -6555,6 +6567,7 @@
       "resolved": "https://registry.npmjs.org/ajv/-/ajv-8.17.1.tgz",
       "integrity": "sha512-B/gBuNg5SiMTrPkC+A2+cW0RszwxYmn6VYxB/inlBStS5nx6xHIt/ehKRhIMhqusl7a8LjQoZnjCs5vhwxOQ1g==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "fast-deep-equal": "^3.1.3",
         "fast-uri": "^3.0.1",
@@ -6600,6 +6613,7 @@
       "resolved": "https://registry.npmjs.org/algoliasearch/-/algoliasearch-5.48.0.tgz",
       "integrity": "sha512-aD8EQC6KEman6/S79FtPdQmB7D4af/etcRL/KwiKFKgAE62iU8c5PeEQvpvIcBPurC3O/4Lj78nOl7ZcoazqSw==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@algolia/abtesting": "1.14.0",
         "@algolia/client-abtesting": "5.48.0",
@@ -7076,6 +7090,7 @@
         }
       ],
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "baseline-browser-mapping": "^2.9.0",
         "caniuse-lite": "^1.0.30001759",
@@ -8088,6 +8103,7 @@
       "resolved": "https://registry.npmjs.org/postcss-selector-parser/-/postcss-selector-parser-7.1.1.tgz",
       "integrity": "sha512-orRsuYpJVw8LdAwqqLykBj9ecS5/cRHlI5+nvTo8LcCKmzDmqVORXtOIYEEQuL9D4BxtA1lm5isAqzQZCoQ6Eg==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "cssesc": "^3.0.0",
         "util-deprecate": "^1.0.2"
@@ -8407,6 +8423,7 @@
       "resolved": "https://registry.npmjs.org/cytoscape/-/cytoscape-3.33.1.tgz",
       "integrity": "sha512-iJc4TwyANnOGR1OmWhsS9ayRS3s+XQ185FmuHObThD+5AeJCakAAbWv8KimMTt08xCCLNgneQwFp+JRJOr9qGQ==",
       "license": "MIT",
+      "peer": true,
       "engines": {
         "node": ">=0.10"
       }
@@ -8828,6 +8845,7 @@
       "resolved": "https://registry.npmjs.org/d3-selection/-/d3-selection-3.0.0.tgz",
       "integrity": "sha512-fmTRWbNMmsmWq6xJV8D19U/gw/bwrHfNXxrIN+HfZgnzqTHp9jOmKMhsTUjXOJnZOdZY9Q28y4yebKzqDKlxlQ==",
       "license": "ISC",
+      "peer": true,
       "engines": {
         "node": ">=12"
       }
@@ -10019,6 +10037,7 @@
       "resolved": "https://registry.npmjs.org/ajv/-/ajv-6.12.6.tgz",
       "integrity": "sha512-j3fVLgvTo527anyYyJOGTYJbG+vnnQYvE0m5mmkc1TK+nxAppkCLMIL0aZ4dblVCNoGShhm+kzE4ZUykBoMg4g==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "fast-deep-equal": "^3.1.1",
         "fast-json-stable-stringify": "^2.0.0",
@@ -14602,6 +14621,7 @@
       "resolved": "https://registry.npmjs.org/ajv/-/ajv-6.12.6.tgz",
       "integrity": "sha512-j3fVLgvTo527anyYyJOGTYJbG+vnnQYvE0m5mmkc1TK+nxAppkCLMIL0aZ4dblVCNoGShhm+kzE4ZUykBoMg4g==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "fast-deep-equal": "^3.1.1",
         "fast-json-stable-stringify": "^2.0.0",
@@ -15190,6 +15210,7 @@
         }
       ],
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "nanoid": "^3.3.11",
         "picocolors": "^1.1.1",
@@ -16093,6 +16114,7 @@
       "resolved": "https://registry.npmjs.org/postcss-selector-parser/-/postcss-selector-parser-7.1.1.tgz",
       "integrity": "sha512-orRsuYpJVw8LdAwqqLykBj9ecS5/cRHlI5+nvTo8LcCKmzDmqVORXtOIYEEQuL9D4BxtA1lm5isAqzQZCoQ6Eg==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "cssesc": "^3.0.0",
         "util-deprecate": "^1.0.2"
@@ -16918,6 +16940,7 @@
       "resolved": "https://registry.npmjs.org/react/-/react-18.3.1.tgz",
       "integrity": "sha512-wS+hAgJShR0KhEvPJArfuPVN1+Hz1t0Y6n5jLrGQbkb4urgPE/0Rve+1kMB1v/oWgHgm4WIcV+i7F2pTVj+2iQ==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "loose-envify": "^1.1.0"
       },
@@ -16930,6 +16953,7 @@
       "resolved": "https://registry.npmjs.org/react-dom/-/react-dom-18.3.1.tgz",
       "integrity": "sha512-5m4nQKp+rZRb09LNH59GM4BxTh9251/ylbKIbpe7TpGxfJ+9kv6BLkLBXIjjspbgbnIBNqlI23tRnTWT0snUIw==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "loose-envify": "^1.1.0",
         "scheduler": "^0.23.2"
@@ -16986,6 +17010,7 @@
       "resolved": "https://registry.npmjs.org/@docusaurus/react-loadable/-/react-loadable-6.0.0.tgz",
       "integrity": "sha512-YMMxTUQV/QFSnbgrP3tjDzLHRg7vsbMn8e9HAa8o/1iXoiomo48b7sk/kkmWEuWNDPJVlKSJRB6Y2fHqdJk+SQ==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@types/react": "*"
       },
@@ -17014,6 +17039,7 @@
       "resolved": "https://registry.npmjs.org/react-router/-/react-router-5.3.4.tgz",
       "integrity": "sha512-Ys9K+ppnJah3QuaRiLxk+jDWOR1MekYQrlytiXxC1RyfbdsZkS5pvKAzCCr031xHixZwpnsYNT5xysdFHQaYsA==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@babel/runtime": "^7.12.13",
         "history": "^4.9.0",
@@ -18826,7 +18852,8 @@
       "version": "2.8.1",
       "resolved": "https://registry.npmjs.org/tslib/-/tslib-2.8.1.tgz",
       "integrity": "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w==",
-      "license": "0BSD"
+      "license": "0BSD",
+      "peer": true
     },
     "node_modules/tsyringe": {
       "version": "4.10.0",
@@ -18907,6 +18934,7 @@
       "integrity": "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw==",
       "devOptional": true,
       "license": "Apache-2.0",
+      "peer": true,
       "bin": {
         "tsc": "bin/tsc",
         "tsserver": "bin/tsserver"
@@ -19262,6 +19290,7 @@
       "resolved": "https://registry.npmjs.org/ajv/-/ajv-6.12.6.tgz",
       "integrity": "sha512-j3fVLgvTo527anyYyJOGTYJbG+vnnQYvE0m5mmkc1TK+nxAppkCLMIL0aZ4dblVCNoGShhm+kzE4ZUykBoMg4g==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "fast-deep-equal": "^3.1.1",
         "fast-json-stable-stringify": "^2.0.0",
@@ -19509,6 +19538,7 @@
       "resolved": "https://registry.npmjs.org/webpack/-/webpack-5.105.1.tgz",
       "integrity": "sha512-Gdj3X74CLJJ8zy4URmK42W7wTZUJrqL+z8nyGEr4dTN0kb3nVs+ZvjbTOqRYPD7qX4tUmwyHL9Q9K6T1seW6Yw==",
       "license": "MIT",
+      "peer": true,
       "dependencies": {
         "@types/eslint-scope": "^3.7.7",
         "@types/estree": "^1.0.8",