chore(vllm): derive the darwin vLLM version, drop the second pin

Follow-up: VLLM_VERSION was still a hardcoded string duplicating what
VLLM_METAL_VERSION already determines. Derive it at install time from
vllm-metal's own installer (vllm_v=) at the pinned tag - one source of truth,
no second value to drift. The bumper now touches only VLLM_METAL_VERSION;
the derivation is immutable per tag, so builds stay reproducible.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:opus-4.8 [Claude Code]

.github/backend-matrix.yml

@@ -4974,6 +4974,13 @@ includeDarwin:
   - backend: "kitten-tts"
     tag-suffix: "-metal-darwin-arm64-kitten-tts"
     build-type: "mps"
+  # vLLM on Apple Silicon via vllm-metal (MLX). The install is custom
+  # (backend/python/vllm/install.sh has a darwin branch); lang stays python so
+  # backend_build_darwin.yml drives it through build-darwin-python-backend ->
+  # scripts/build/python-darwin.sh, which runs the backend's install.sh.
+  - backend: "vllm"
+    tag-suffix: "-metal-darwin-arm64-vllm"
+    build-type: "mps"
   - backend: "piper"
     tag-suffix: "-metal-darwin-arm64-piper"
     build-type: "metal"

.github/bump_vllm_metal.sh

@@ -0,0 +1,55 @@
+#!/bin/bash
+# Bump the single vllm-metal pin (VLLM_METAL_VERSION) in the vLLM backend's
+# darwin (Apple Silicon) install path. The macOS/Metal build
+# (backend/python/vllm/install.sh, Darwin branch) installs vllm-metal, which is
+# version-locked to a specific vLLM source release. install.sh derives that vLLM
+# version at build time from vllm-metal's own installer (`vllm_v=`) at the pinned
+# tag, so there is only ONE value to bump here -- mirroring bump_vllm_wheel.sh,
+# which bumps the Linux cu130 wheel pin.
+#
+# This deliberately tracks vllm-project/vllm-metal, NOT vllm-project/vllm: the
+# darwin build can only use the exact vLLM version vllm-metal supports, so it may
+# lag the Linux pin (requirements-cublas13-after.txt) until vllm-metal catches up.
+set -xe
+REPO=$1   # vllm-project/vllm-metal
+FILE=$2   # backend/python/vllm/install.sh
+VAR=$3    # VLLM_METAL_VERSION (used for the workflow's output file names)
+
+if [ -z "$FILE" ] || [ -z "$REPO" ] || [ -z "$VAR" ]; then
+    echo "usage: $0 <repo> <install-file> <var-name>" >&2
+    exit 1
+fi
+
+# vllm-metal ships frequent dev releases, all flagged as non-prerelease, so
+# /releases/latest returns the newest one (with its cp312 wheel asset).
+LATEST_TAG=$(curl -sS -H "Accept: application/vnd.github+json" \
+    "https://api.github.com/repos/$REPO/releases/latest" \
+    | python3 -c "import json,sys; print(json.load(sys.stdin)['tag_name'])")
+
+# The coupled vLLM source version lives in vllm-metal's installer at that tag.
+NEW_VLLM_VERSION=$(curl -fsSL \
+    "https://raw.githubusercontent.com/$REPO/$LATEST_TAG/install.sh" \
+    | grep -oE 'vllm_v="[0-9]+\.[0-9]+\.[0-9]+"' | head -1 | cut -d'"' -f2)
+
+if [ -z "$LATEST_TAG" ] || [ -z "$NEW_VLLM_VERSION" ]; then
+    echo "Could not resolve vllm-metal tag ($LATEST_TAG) or its vllm_v ($NEW_VLLM_VERSION)." >&2
+    exit 1
+fi
+
+set +e
+CURRENT_TAG=$(grep -oE 'VLLM_METAL_VERSION="[^"]*"' "$FILE" | head -1 | cut -d'"' -f2)
+set -e
+
+# Rewrite the single pin. install.sh derives VLLM_VERSION from this tag at build
+# time, so there is nothing else to touch. peter-evans/create-pull-request opens
+# no PR on a clean tree, so a no-op rewrite (already current) is safe.
+sed -i "$FILE" \
+    -e "s|VLLM_METAL_VERSION=\"[^\"]*\"|VLLM_METAL_VERSION=\"$LATEST_TAG\"|"
+
+if [ -z "$CURRENT_TAG" ]; then
+    echo "Could not find VLLM_METAL_VERSION=\"...\" in $FILE." >&2
+    exit 0
+fi
+
+echo "vllm-metal ${CURRENT_TAG} -> ${LATEST_TAG} (builds vLLM ${NEW_VLLM_VERSION}): https://github.com/$REPO/releases/tag/${LATEST_TAG}" >> "${VAR}_message.txt"
+echo "${LATEST_TAG}" >> "${VAR}_commit.txt"

.github/workflows/bump_deps.yaml

@@ -154,3 +154,39 @@ jobs:
           branch: "update/VLLM_VERSION"
           body: ${{ steps.bump.outputs.message }}
           signoff: true
+
+  bump-vllm-metal:
+    # The darwin (Apple Silicon) vLLM build installs vllm-metal, which is locked
+    # to a specific vLLM source release. install.sh pins both VLLM_METAL_VERSION
+    # (the wheel release) and VLLM_VERSION (the vLLM it builds against); this job
+    # tracks vllm-project/vllm-metal and rewrites both atomically. Separate from
+    # bump-vllm-wheel because darwin follows vllm-metal, not vllm/vllm latest.
+    if: github.repository == 'mudler/LocalAI'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+      - name: Bump vllm-metal pin 🔧
+        id: bump
+        run: |
+          bash .github/bump_vllm_metal.sh vllm-project/vllm-metal backend/python/vllm/install.sh VLLM_METAL_VERSION
+          {
+            echo 'message<<EOF'
+            cat "VLLM_METAL_VERSION_message.txt"
+            echo EOF
+          } >> "$GITHUB_OUTPUT"
+          {
+            echo 'commit<<EOF'
+            cat "VLLM_METAL_VERSION_commit.txt"
+            echo EOF
+          } >> "$GITHUB_OUTPUT"
+          rm -rfv VLLM_METAL_VERSION_message.txt VLLM_METAL_VERSION_commit.txt
+      - name: Create Pull Request
+        uses: peter-evans/create-pull-request@v8
+        with:
+          token: ${{ secrets.UPDATE_BOT_TOKEN }}
+          push-to-fork: ci-forks/LocalAI
+          commit-message: ':arrow_up: Update vllm-project/vllm-metal (darwin)'
+          title: 'chore: :arrow_up: Update vllm-metal (darwin) to `${{ steps.bump.outputs.commit }}`'
+          branch: "update/VLLM_METAL_VERSION"
+          body: ${{ steps.bump.outputs.message }}
+          signoff: true

backend/cpp/ik-llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-IK_LLAMA_VERSION?=6c00e87ac84404af588ad2e65935bd6f079c696f
+IK_LLAMA_VERSION?=7ccf1d209588962b96eacca325b37e9b3e8faf5e
 LLAMA_REPO?=https://github.com/ikawrakow/ik_llama.cpp
 
 CMAKE_ARGS?=

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=73618f27a801c0b8614ceaf3547d3c2a99baae14
+LLAMA_VERSION?=be4a6a63eb2b848e19c277bdcf2bd399e8af76d9
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=

backend/go/crispasr/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # CrispASR version (release tag)
 CRISPASR_REPO?=https://github.com/CrispStrobe/CrispASR
-CRISPASR_VERSION?=63b57289255267edf66e43e33bc3911e04a2e92d
+CRISPASR_VERSION?=96b2a6ee31d30389fed8a7ef1a54239b75231ddc
 SO_TARGET?=libgocrispasr.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/go/parakeet-cpp/Makefile

@@ -1,6 +1,6 @@
 # parakeet-cpp backend Makefile.
 #
-# Upstream pin lives below as PARAKEET_VERSION?=db755a78d39f789bb7d4e3935158a9e8105dbe36
+# Upstream pin lives below as PARAKEET_VERSION?=89f5e2977b4d8bccd45e7bcc6f2ef7c4ed49e89a
 # (.github/bump_deps.sh) can find and update it - matches the
 # whisper.cpp / ds4 / vibevoice-cpp convention.
 #
@@ -15,7 +15,7 @@
 # That's what the L0 smoke test uses. The default target below does the
 # proper clone-at-pin + cmake build so CI doesn't need a side-checkout.
 
-PARAKEET_VERSION?=db755a78d39f789bb7d4e3935158a9e8105dbe36
+PARAKEET_VERSION?=89f5e2977b4d8bccd45e7bcc6f2ef7c4ed49e89a
 PARAKEET_REPO?=https://github.com/mudler/parakeet.cpp
 
 GOCMD?=go

backend/go/whisper/Makefile

@@ -8,7 +8,7 @@ JOBS?=$(shell nproc --ignore=1)
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=bae6bc02b1940bbfb87b6a0299c565e563b916d1
+WHISPER_CPP_VERSION?=43d78af5be58f41d6ffbc227d608f104577741ea
 SO_TARGET?=libgowhisper.so
 
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF

backend/index.yaml

@@ -645,6 +645,7 @@
     nvidia-cuda-13: "cuda13-vllm"
     nvidia-l4t-cuda-13: "cuda13-nvidia-l4t-arm64-vllm"
     cpu: "cpu-vllm"
+    metal: "metal-vllm"
 - &sglang
   name: "sglang"
   license: apache-2.0
@@ -2927,6 +2928,17 @@
     nvidia-cuda-13: "cuda13-vllm-development"
     nvidia-l4t-cuda-13: "cuda13-nvidia-l4t-arm64-vllm-development"
     cpu: "cpu-vllm-development"
+    metal: "metal-vllm-development"
+- !!merge <<: *vllm
+  name: "metal-vllm"
+  uri: "quay.io/go-skynet/local-ai-backends:latest-metal-darwin-arm64-vllm"
+  mirrors:
+    - localai/localai-backends:latest-metal-darwin-arm64-vllm
+- !!merge <<: *vllm
+  name: "metal-vllm-development"
+  uri: "quay.io/go-skynet/local-ai-backends:master-metal-darwin-arm64-vllm"
+  mirrors:
+    - localai/localai-backends:master-metal-darwin-arm64-vllm
 - !!merge <<: *vllm
   name: "cuda12-vllm"
   uri: "quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-12-vllm"

backend/python/vllm/install.sh

@@ -43,6 +43,24 @@ if [ "x${BUILD_PROFILE}" == "xcublas13" ]; then
     EXTRA_PIP_INSTALL_FLAGS+=" --index-strategy=unsafe-best-match"
 fi
 
+# Apple Silicon (Metal/MLX) via vllm-metal.
+# vllm-metal (github.com/vllm-project/vllm-metal) brings vLLM to macOS on Apple
+# Silicon: it registers through vLLM's platform-plugin entry point
+# (metal -> vllm_metal:register), MetalPlatform activates, and the vLLM v1
+# AsyncLLM engine runs on the GPU through MLX. LocalAI's backend.py is UNCHANGED
+# on darwin — AsyncEngineArgs(...) -> AsyncLLMEngine.from_engine_args transparently
+# resolves to the MLX engine (proven on a real M4 / macOS 26.5 against Qwen3-0.6B).
+#
+# vllm-metal REQUIRES Python 3.12, so force the portable CPython before the venv
+# is created (ensureVenv reads PYTHON_VERSION/PYTHON_PATCH/PY_STANDALONE_TAG).
+# The patch + standalone tag mirror the l4t13 cp312 pin — a known-good
+# python-build-standalone release that also ships an aarch64-apple-darwin asset.
+if [ "$(uname -s)" = "Darwin" ]; then
+    PYTHON_VERSION="3.12"
+    PYTHON_PATCH="12"
+    PY_STANDALONE_TAG="20251120"
+fi
+
 # JetPack 7 / L4T arm64 vllm + torch wheels come straight from PyPI now
 # (torch 2.11+ ships aarch64 + cu130 manylinux wheels and vllm 0.20+ ships
 # an aarch64 wheel pinned to that torch). They're cp312-only, so bump the
@@ -57,11 +75,92 @@ if [ "x${BUILD_PROFILE}" == "xl4t13" ]; then
     PY_STANDALONE_TAG="20251120"
 fi
 
+# ===================== Apple Silicon (Metal/MLX) =====================
+# Reproduce vllm-metal's upstream installer
+# (curl -fsSL https://raw.githubusercontent.com/vllm-project/vllm-metal/main/install.sh)
+# but INTO LocalAI's managed venv (ensureVenv) instead of a throwaway
+# ~/.venv-vllm-metal, so the backend integrates with LocalAI's venv lifecycle
+# (portable CPython, _makeVenvPortable relocation, runtime activation). The
+# normal CUDA/CPU installRequirements is skipped on darwin — there is no
+# macOS/arm64 vLLM wheel on PyPI; vLLM is built from source and the MLX engine
+# is layered on by the vllm-metal wheel.
+if [ "$(uname -s)" = "Darwin" ]; then
+    # Create/activate the portable 3.12 venv. On darwin USE_PIP=true and
+    # PORTABLE_PYTHON=true (set by scripts/build/python-darwin.sh), so this is a
+    # `python -m venv` based, relocatable venv.
+    ensureVenv
+
+    # vllm-metal's installer drives everything through `uv`: building vLLM from
+    # the CPU requirements needs `--index-strategy unsafe-best-match` (mixes the
+    # pytorch CPU channel with PyPI), a flag plain pip does not have. The darwin
+    # venv is pip-based, so bootstrap uv into it. uv honours $VIRTUAL_ENV (set by
+    # libbackend's _activateVenv) and installs into THIS venv — same pattern the
+    # intel branch below relies on.
+    pip install uv
+
+    # The ONLY darwin version pin -- AUTO-BUMPED by .github/bump_vllm_metal.sh,
+    # which tracks vllm-project/vllm-metal releases (NOT vllm/vllm latest). Keep
+    # it as a plain double-quoted assignment on its own line so the bumper's sed
+    # can rewrite it. Darwin therefore follows vllm-metal and can lag the Linux
+    # vllm pin (requirements-cublas13-after.txt, bumped independently against
+    # vllm/vllm) until vllm-metal supports a newer vLLM.
+    VLLM_METAL_VERSION="v0.3.0.dev20260622062346"
+
+    # The coupled vLLM source version is whatever this vllm-metal release builds
+    # against -- it declares it in its own installer as `vllm_v=`. Derive it from
+    # the PINNED tag rather than hardcoding a second value that could drift. The
+    # tag is immutable, so this stays reproducible across rebuilds.
+    VLLM_VERSION=$(curl -fsSL "https://raw.githubusercontent.com/vllm-project/vllm-metal/${VLLM_METAL_VERSION}/install.sh" \
+        | grep -oE 'vllm_v="[0-9]+\.[0-9]+\.[0-9]+"' | head -n1 | cut -d'"' -f2)
+    if [ -z "${VLLM_VERSION}" ]; then
+        echo "ERROR: could not derive the vLLM version from vllm-metal ${VLLM_METAL_VERSION}" >&2
+        exit 1
+    fi
+    echo "vllm-metal ${VLLM_METAL_VERSION} builds against vLLM ${VLLM_VERSION}"
+
+    _vllm_src=$(mktemp -d)
+    trap 'rm -rf "${_vllm_src}"' EXIT
+    pushd "${_vllm_src}"
+        # 1) Build vLLM ${VLLM_VERSION} from the release source tarball against
+        #    the CPU requirements. vllm-metal layers its MLX platform plugin on
+        #    top of this exact build.
+        curl -fsSL -o "vllm-${VLLM_VERSION}.tar.gz" \
+            "https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}.tar.gz"
+        tar -xzf "vllm-${VLLM_VERSION}.tar.gz"
+        pushd "vllm-${VLLM_VERSION}"
+            uv pip install -r requirements/cpu.txt --index-strategy unsafe-best-match
+            # -Wno-parentheses: clang on macOS treats one of vLLM's C++ warnings
+            # as an error without it (matches the upstream installer's CXXFLAGS).
+            CXXFLAGS="-Wno-parentheses" uv pip install .
+        popd
+    popd
+
+    # 2) Install the prebuilt vllm-metal wheel from the PINNED release
+    #    (${VLLM_METAL_VERSION}). It pulls mlx / mlx-metal as deps and registers
+    #    the `metal` platform plugin that backend.py resolves to at engine-init
+    #    time. Pinning the tag (vs releases/latest) keeps the wheel and the vLLM
+    #    source build above reproducible and coupled; .github/bump_vllm_metal.sh
+    #    advances both together.
+    _metal_wheel_url=$(curl -fsSL "https://api.github.com/repos/vllm-project/vllm-metal/releases/tags/${VLLM_METAL_VERSION}" \
+        | grep -oE '"browser_download_url"[[:space:]]*:[[:space:]]*"[^"]+\.whl"' \
+        | head -n1 | sed -E 's/.*"(https[^"]+)".*/\1/')
+    if [ -z "${_metal_wheel_url}" ]; then
+        echo "ERROR: could not resolve a vllm-metal wheel URL for release ${VLLM_METAL_VERSION}" >&2
+        exit 1
+    fi
+    echo "Installing vllm-metal wheel: ${_metal_wheel_url}"
+    uv pip install "${_metal_wheel_url}"
+
+    # Generate the gRPC stubs (backend_pb2*). installRequirements normally does
+    # this via runProtogen at the end; we skipped installRequirements on darwin,
+    # so call it explicitly here.
+    runProtogen
+
 # Intel XPU has no upstream-published vllm wheels, so we always build vllm
 # from source against torch-xpu and replace the default triton with
 # triton-xpu (matching torch 2.11). Mirrors the upstream procedure:
 # https://github.com/vllm-project/vllm/blob/main/docs/getting_started/installation/gpu.xpu.inc.md
-if [ "x${BUILD_TYPE}" == "xintel" ]; then
+elif [ "x${BUILD_TYPE}" == "xintel" ]; then
     # Hide requirements-intel-after.txt so installRequirements doesn't
     # try `pip install vllm` (would either fail or grab a non-XPU wheel).
     _intel_after="${backend_dir}/requirements-intel-after.txt"

backend/python/vllm/requirements-cublas13-after.txt

@@ -4,4 +4,7 @@
 # instead — the cublas13 case in install.sh adds --index-strategy=unsafe-best-match
 # so uv consults this index alongside PyPI.
 --extra-index-url https://wheels.vllm.ai/0.23.0/cu130
+# VERSION COUPLING: darwin/Apple-Silicon builds use vllm-metal (see install.sh),
+# which pins this exact vLLM version. Bumping vllm here means coordinating with a
+# vllm-metal release that supports the new version, or macOS/Metal builds break.
 vllm==0.23.0

core/application/config_file_watcher.go

@@ -215,6 +215,7 @@ func readRuntimeSettingsJson(startupAppConfig config.ApplicationConfig) fileHand
 		envBackendGalleries := slices.Equal(appConfig.BackendGalleries, startupAppConfig.BackendGalleries)
 		envAutoloadGalleries := appConfig.AutoloadGalleries == startupAppConfig.AutoloadGalleries
 		envAutoloadBackendGalleries := appConfig.AutoloadBackendGalleries == startupAppConfig.AutoloadBackendGalleries
+		envPIIDefaultDetectors := slices.Equal(appConfig.PIIDefaultDetectors, startupAppConfig.PIIDefaultDetectors)
 		envAgentJobRetentionDays := appConfig.AgentJobRetentionDays == startupAppConfig.AgentJobRetentionDays
 		envForceEvictionWhenBusy := appConfig.ForceEvictionWhenBusy == startupAppConfig.ForceEvictionWhenBusy
 		envLRUEvictionMaxRetries := appConfig.LRUEvictionMaxRetries == startupAppConfig.LRUEvictionMaxRetries
@@ -335,6 +336,15 @@ func readRuntimeSettingsJson(startupAppConfig config.ApplicationConfig) fileHand
 			if settings.AutoloadBackendGalleries != nil && !envAutoloadBackendGalleries {
 				appConfig.AutoloadBackendGalleries = *settings.AutoloadBackendGalleries
 			}
+			if settings.PIIDefaultDetectors != nil && !envPIIDefaultDetectors {
+				// Request-side default redaction reads this live via
+				// ResolvePIIPolicy, so a file edit takes effect on the next chat
+				// request. The MITM listener resolves its per-host detector map
+				// once at start, so a raw file edit reaches cloud-proxy traffic
+				// only after a restart or a POST /api/settings (which rebuilds
+				// the listener) — the admin UI uses the latter.
+				appConfig.PIIDefaultDetectors = append([]string(nil), (*settings.PIIDefaultDetectors)...)
+			}
 			if settings.AutoUpgradeBackends != nil {
 				appConfig.AutoUpgradeBackends = *settings.AutoUpgradeBackends
 			}

core/application/runtime_settings_branding_test.go

@@ -109,6 +109,52 @@ var _ = Describe("loadRuntimeSettingsFromFile", func() {
 		})
 	})
 
+	// Instance-wide default PII detectors. The file is the only source (no
+	// env var), and the loader runs immediately before startMITMIfConfigured,
+	// so a regression here means the cloud-proxy MITM listener resolves an
+	// empty detector set at boot and forwards intercepted traffic unredacted —
+	// even though pii_default_detectors is on disk and the MITM model has PII
+	// enabled. It also breaks request-side default redaction the same way.
+	Describe("PII default detectors", func() {
+		It("loads pii_default_detectors from the file", func() {
+			cfg := &config.ApplicationConfig{DynamicConfigsDir: seedSettings(`{"pii_default_detectors": ["privacy-filter-nemotron", "secret-filter"]}`)}
+			loadRuntimeSettingsFromFile(cfg)
+			Expect(cfg.PIIDefaultDetectors).To(Equal([]string{"privacy-filter-nemotron", "secret-filter"}))
+		})
+
+		It("does not override an env/CLI-set value (LOCALAI_PII_DEFAULT_DETECTORS)", func() {
+			cfg := &config.ApplicationConfig{
+				DynamicConfigsDir:   seedSettings(`{"pii_default_detectors": ["from-file"]}`),
+				PIIDefaultDetectors: []string{"from-env"}, // simulate WithPIIDefaultDetectors(env)
+			}
+			loadRuntimeSettingsFromFile(cfg)
+			Expect(cfg.PIIDefaultDetectors).To(Equal([]string{"from-env"}), "env var must win over the persisted file value")
+		})
+	})
+
+	// The live file watcher applies pii_default_detectors on a runtime change
+	// the same way it handles galleries/threads/etc.: env-set values (current
+	// == startup snapshot) are left alone, otherwise the file value is applied
+	// to the live config so request-side default redaction picks it up without
+	// a restart.
+	Describe("file watcher: pii_default_detectors", func() {
+		It("applies a changed file value to the live config", func() {
+			startup := config.ApplicationConfig{} // no env baseline
+			live := &config.ApplicationConfig{PIIDefaultDetectors: []string{"old"}}
+			handler := readRuntimeSettingsJson(startup)
+			Expect(handler([]byte(`{"pii_default_detectors":["new-a","new-b"]}`), live)).To(Succeed())
+			Expect(live.PIIDefaultDetectors).To(Equal([]string{"new-a", "new-b"}))
+		})
+
+		It("leaves an env-controlled value untouched", func() {
+			startup := config.ApplicationConfig{PIIDefaultDetectors: []string{"from-env"}}
+			live := &config.ApplicationConfig{PIIDefaultDetectors: []string{"from-env"}}
+			handler := readRuntimeSettingsJson(startup)
+			Expect(handler([]byte(`{"pii_default_detectors":["from-file"]}`), live)).To(Succeed())
+			Expect(live.PIIDefaultDetectors).To(Equal([]string{"from-env"}), "env-controlled detectors must not be overwritten by the file")
+		})
+	})
+
 	// The Agent Pool block has a mix of zero and non-zero defaults
 	// (Enabled=true, EmbeddingModel="granite-...", MaxChunkingSize=400,
 	// VectorEngine="chromem", AgentHubURL="https://agenthub.localai.io").

core/application/startup.go

@@ -750,6 +750,20 @@ func loadRuntimeSettingsFromFile(options *config.ApplicationConfig) {
 		options.MITMListen = *settings.MITMListen
 	}
 
+	// Instance-wide default PII detectors. LOCALAI_PII_DEFAULT_DETECTORS (via
+	// WithPIIDefaultDetectors) wins when set; otherwise the file is the source
+	// — apply it only when the env/CLI left the value empty, mirroring the
+	// "env > file" precedence used for the other fields. This must land before
+	// startMITMIfConfigured (called right after this loader): the cloud-proxy
+	// listener resolves each intercept host's detectors once at start via
+	// ResolvePIIPolicy, and a MITM model that names no detectors of its own
+	// falls back to these defaults. Without it the listener (and request-side
+	// default redaction) starts with an empty detector set and forwards
+	// traffic unredacted even though pii_default_detectors is on disk.
+	if settings.PIIDefaultDetectors != nil && len(options.PIIDefaultDetectors) == 0 {
+		options.PIIDefaultDetectors = append([]string(nil), (*settings.PIIDefaultDetectors)...)
+	}
+
 	// Backend upgrade flags
 	if settings.AutoUpgradeBackends != nil {
 		if !options.AutoUpgradeBackends {

core/cli/run.go

@@ -181,6 +181,8 @@ type RunCMD struct {
 	// Cloud-proxy MITM listener (off by default).
 	MITMListen string `env:"LOCALAI_MITM_LISTEN" help:"Address (host:port) for the cloudproxy MITM listener. Empty = disabled. Clients set HTTPS_PROXY=http://<this>:<port>. Intercept hosts are declared per-model via the model YAML mitm.hosts: block; create one from the Add Model UI." group:"middleware"`
 	MITMCADir  string `env:"LOCALAI_MITM_CA_DIR" type:"path" help:"Directory holding the MITM proxy CA cert + key. Defaults to <data-path>/mitm-ca." group:"middleware"`
+
+	PIIDefaultDetectors []string `env:"LOCALAI_PII_DEFAULT_DETECTORS" help:"Instance-wide default PII/secret detector model names applied to any PII-enabled model (chiefly cloud-proxy / MITM models) that names no pii.detectors of its own. Comma-separated, e.g. privacy-filter-nemotron,secret-filter. Takes precedence over the value persisted via the Middleware UI." group:"middleware"`
 }
 
 func (r *RunCMD) Run(ctx *cliContext.Context) error {
@@ -243,6 +245,7 @@ func (r *RunCMD) Run(ctx *cliContext.Context) error {
 		config.WithAPIAddress(r.Address),
 		config.WithMITMListen(r.MITMListen),
 		config.WithMITMCADir(r.MITMCADir),
+		config.WithPIIDefaultDetectors(r.PIIDefaultDetectors),
 		config.WithAgentJobRetentionDays(r.AgentJobRetentionDays),
 		config.WithLlamaCPPTunnelCallback(func(tunnels []string) {
 			tunnelEnvVar := strings.Join(tunnels, ",")

core/config/application_config.go

@@ -712,6 +712,18 @@ func WithMITMCADir(dir string) AppOption {
 	}
 }
 
+// WithPIIDefaultDetectors sets the instance-wide default PII/secret detector
+// model names applied to any PII-enabled model (chiefly cloud-proxy / MITM
+// models) that names no pii.detectors of its own. CLI/env:
+// LOCALAI_PII_DEFAULT_DETECTORS. Empty leaves the value to
+// runtime_settings.json / the Middleware UI; a non-empty value takes
+// precedence over the file (env > file).
+func WithPIIDefaultDetectors(detectors []string) AppOption {
+	return func(o *ApplicationConfig) {
+		o.PIIDefaultDetectors = detectors
+	}
+}
+
 func WithDynamicConfigDir(dynamicConfigsDir string) AppOption {
 	return func(o *ApplicationConfig) {
 		o.DynamicConfigsDir = dynamicConfigsDir

core/http/endpoints/localai/agent_collections.go

@@ -70,7 +70,7 @@ func UploadToCollectionEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		file, err := c.FormFile("file")
 		if err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": "file required"})
@@ -116,7 +116,7 @@ func ListCollectionEntriesEndpoint(app *application.Application) echo.HandlerFun
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		entries, err := svc.ListCollectionEntriesForUser(userID, c.Param("name"))
+		entries, err := svc.ListCollectionEntriesForUser(userID, decodedParam(c, "name"))
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -139,7 +139,7 @@ func GetCollectionEntryContentEndpoint(app *application.Application) echo.Handle
 		if err != nil {
 			entry = entryParam
 		}
-		content, chunkCount, err := svc.GetCollectionEntryContentForUser(userID, c.Param("name"), entry)
+		content, chunkCount, err := svc.GetCollectionEntryContentForUser(userID, decodedParam(c, "name"), entry)
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -164,7 +164,7 @@ func SearchCollectionEndpoint(app *application.Application) echo.HandlerFunc {
 		if err := c.Bind(&payload); err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": err.Error()})
 		}
-		results, err := svc.SearchCollectionForUser(userID, c.Param("name"), payload.Query, payload.MaxResults)
+		results, err := svc.SearchCollectionForUser(userID, decodedParam(c, "name"), payload.Query, payload.MaxResults)
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -182,7 +182,7 @@ func ResetCollectionEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		if err := svc.ResetCollectionForUser(userID, c.Param("name")); err != nil {
+		if err := svc.ResetCollectionForUser(userID, decodedParam(c, "name")); err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 			}
@@ -202,7 +202,7 @@ func DeleteCollectionEntryEndpoint(app *application.Application) echo.HandlerFun
 		if err := c.Bind(&payload); err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": err.Error()})
 		}
-		remaining, err := svc.DeleteCollectionEntryForUser(userID, c.Param("name"), payload.Entry)
+		remaining, err := svc.DeleteCollectionEntryForUser(userID, decodedParam(c, "name"), payload.Entry)
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -230,7 +230,7 @@ func AddCollectionSourceEndpoint(app *application.Application) echo.HandlerFunc
 		if payload.UpdateInterval < 1 {
 			payload.UpdateInterval = 60
 		}
-		if err := svc.AddCollectionSourceForUser(userID, c.Param("name"), payload.URL, payload.UpdateInterval); err != nil {
+		if err := svc.AddCollectionSourceForUser(userID, decodedParam(c, "name"), payload.URL, payload.UpdateInterval); err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 			}
@@ -250,7 +250,7 @@ func RemoveCollectionSourceEndpoint(app *application.Application) echo.HandlerFu
 		if err := c.Bind(&payload); err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": err.Error()})
 		}
-		if err := svc.RemoveCollectionSourceForUser(userID, c.Param("name"), payload.URL); err != nil {
+		if err := svc.RemoveCollectionSourceForUser(userID, decodedParam(c, "name"), payload.URL); err != nil {
 			return c.JSON(http.StatusInternalServerError, map[string]string{"error": err.Error()})
 		}
 		return c.JSON(http.StatusOK, map[string]string{"status": "ok"})
@@ -267,7 +267,7 @@ func GetCollectionEntryRawFileEndpoint(app *application.Application) echo.Handle
 		if err != nil {
 			entry = entryParam
 		}
-		fpath, err := svc.GetCollectionEntryFilePathForUser(userID, c.Param("name"), entry)
+		fpath, err := svc.GetCollectionEntryFilePathForUser(userID, decodedParam(c, "name"), entry)
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
@@ -282,7 +282,7 @@ func ListCollectionSourcesEndpoint(app *application.Application) echo.HandlerFun
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		sources, err := svc.ListCollectionSourcesForUser(userID, c.Param("name"))
+		sources, err := svc.ListCollectionSourcesForUser(userID, decodedParam(c, "name"))
 		if err != nil {
 			if strings.Contains(err.Error(), "not found") {
 				return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})

core/http/endpoints/localai/agent_collections_param_test.go

@@ -0,0 +1,49 @@
+package localai
+
+import (
+	"net/http"
+	"net/http/httptest"
+
+	"github.com/labstack/echo/v4"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+// Regression for #10443: agent/collection names carry a "legacy-api-key:"
+// prefix, so the ':' is percent-encoded as %3A in the request path. Echo routes
+// such paths via URL.RawPath and stores the path-param value still escaped, so
+// handlers must URL-decode it before looking the collection up in the store -
+// otherwise the lookup sees "legacy-api-key%3ALiteraryResearch" and 404s.
+var _ = Describe("decodedParam", func() {
+	var e *echo.Echo
+
+	BeforeEach(func() {
+		e = echo.New()
+	})
+
+	// route runs a request through Echo's real router so the path param is
+	// populated exactly as it would be in production, then returns the decoded
+	// value the handler would observe.
+	route := func(rawPath string) string {
+		var got string
+		e.GET("/api/agents/collections/:name/upload", func(c echo.Context) error {
+			got = decodedParam(c, "name")
+			return c.NoContent(http.StatusOK)
+		})
+		req := httptest.NewRequest(http.MethodGet, rawPath, nil)
+		rec := httptest.NewRecorder()
+		e.ServeHTTP(rec, req)
+		Expect(rec.Code).To(Equal(http.StatusOK))
+		return got
+	}
+
+	It("decodes a percent-encoded colon in the collection name", func() {
+		got := route("/api/agents/collections/legacy-api-key%3ALiteraryResearch/upload")
+		Expect(got).To(Equal("legacy-api-key:LiteraryResearch"))
+	})
+
+	It("leaves an unencoded name untouched", func() {
+		got := route("/api/agents/collections/PlainCollection/upload")
+		Expect(got).To(Equal("PlainCollection"))
+	})
+})

core/http/endpoints/localai/agents.go

@@ -6,6 +6,7 @@ import (
 	"io"
 	"maps"
 	"net/http"
+	"net/url"
 	"os"
 	"path/filepath"
 	"slices"
@@ -33,6 +34,22 @@ func getUserID(c echo.Context) string {
 	return user.ID
 }
 
+// decodedParam returns the named path parameter, URL-decoding it.
+//
+// Echo routes a request via URL.RawPath whenever the path contains
+// percent-encoded characters (e.g. %3A for ':'), and in that case stores the
+// matched path-param value raw/escaped. Agent and collection names carry a
+// "legacy-api-key:" prefix, so the ':' arrives as %3A and the raw param no
+// longer matches the stored name. Callers must unescape before lookups.
+// Falls back to the raw value if it isn't valid percent-encoding.
+func decodedParam(c echo.Context, name string) string {
+	raw := c.Param(name)
+	if decoded, err := url.PathUnescape(raw); err == nil {
+		return decoded
+	}
+	return raw
+}
+
 // isAdminUser returns true if the authenticated user has admin role.
 func isAdminUser(c echo.Context) bool {
 	user := auth.GetUser(c)
@@ -127,7 +144,7 @@ func GetAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 
 		statuses := svc.ListAgentsForUser(userID)
 		active, exists := statuses[name]
@@ -142,7 +159,7 @@ func UpdateAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		var cfg state.AgentConfig
 		if err := c.Bind(&cfg); err != nil {
 			return c.JSON(http.StatusBadRequest, map[string]string{"error": err.Error()})
@@ -161,7 +178,7 @@ func DeleteAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		if err := svc.DeleteAgentForUser(userID, name); err != nil {
 			return c.JSON(http.StatusInternalServerError, map[string]string{"error": err.Error()})
 		}
@@ -173,7 +190,7 @@ func GetAgentConfigEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		cfg := svc.GetAgentConfigForUser(userID, name)
 		if cfg == nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": "Agent not found"})
@@ -186,7 +203,7 @@ func PauseAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		if err := svc.PauseAgentForUser(userID, c.Param("name")); err != nil {
+		if err := svc.PauseAgentForUser(userID, decodedParam(c, "name")); err != nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 		}
 		return c.JSON(http.StatusOK, map[string]string{"status": "ok"})
@@ -197,7 +214,7 @@ func ResumeAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		if err := svc.ResumeAgentForUser(userID, c.Param("name")); err != nil {
+		if err := svc.ResumeAgentForUser(userID, decodedParam(c, "name")); err != nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 		}
 		return c.JSON(http.StatusOK, map[string]string{"status": "ok"})
@@ -208,7 +225,7 @@ func GetAgentStatusEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 
 		history := svc.GetAgentStatusForUser(userID, name)
 		if history == nil {
@@ -241,7 +258,7 @@ func GetAgentObservablesEndpoint(app *application.Application) echo.HandlerFunc
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 
 		history, err := svc.GetAgentObservablesForUser(userID, name)
 		if err != nil {
@@ -261,7 +278,7 @@ func ClearAgentObservablesEndpoint(app *application.Application) echo.HandlerFun
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		if err := svc.ClearAgentObservablesForUser(userID, name); err != nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})
 		}
@@ -273,7 +290,7 @@ func ChatWithAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		var payload struct {
 			Message string `json:"message"`
 		}
@@ -302,7 +319,7 @@ func AgentSSEEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 
 		// Try local SSE manager first
 		manager := svc.GetSSEManagerForUser(userID, name)
@@ -334,7 +351,7 @@ func ExportAgentEndpoint(app *application.Application) echo.HandlerFunc {
 	return func(c echo.Context) error {
 		svc := app.AgentPoolService()
 		userID := effectiveUserID(c)
-		name := c.Param("name")
+		name := decodedParam(c, "name")
 		data, err := svc.ExportAgentForUser(userID, name)
 		if err != nil {
 			return c.JSON(http.StatusNotFound, map[string]string{"error": err.Error()})

core/http/endpoints/localai/settings.go

@@ -271,7 +271,14 @@ func UpdateSettingsEndpoint(app *application.Application) echo.HandlerFunc {
 			}
 		}
 
-		if settings.MITMListen != nil {
+		// Rebuild the MITM listener when its address OR the instance-wide
+		// default detectors change. The per-host detector map is resolved once
+		// at listener start (startMITMLocked → ResolvePIIPolicy), so a
+		// default-detector change is otherwise invisible to cloud-proxy traffic
+		// until the next restart — an admin toggling a default detector would
+		// see no redaction. RestartMITM is a no-op when the listener is
+		// disabled (empty address).
+		if settings.MITMListen != nil || settings.PIIDefaultDetectors != nil {
 			if err := app.RestartMITM(); err != nil {
 				xlog.Error("Failed to restart MITM proxy", "error", err)
 				return c.JSON(http.StatusInternalServerError, schema.SettingsResponse{

core/http/endpoints/localai/settings_test.go

@@ -146,6 +146,24 @@ var _ = Describe("Settings endpoints", func() {
 		Expect(*ondisk.PIIDefaultDetectors).To(Equal([]string{"det-a"}))
 	})
 
+	// The MITM listener resolves its per-host PII detectors once at start
+	// (startMITMLocked → ResolvePIIPolicy), and the handler used to restart it
+	// only when mitm_listen changed. So an admin toggling a default detector
+	// (the Middleware detector table POSTs only pii_default_detectors) left
+	// cloud-proxy traffic unredacted until the next reboot. A
+	// pii_default_detectors change must now rebuild the listener.
+	It("rebuilds the MITM listener when only pii_default_detectors changes", func() {
+		rec := post(`{"mitm_listen":"127.0.0.1:0"}`)
+		Expect(rec.Code).To(Equal(http.StatusOK), rec.Body.String())
+		srv1 := app.MITMServer()
+		Expect(srv1).ToNot(BeNil(), "listener should be running after mitm_listen is set")
+
+		rec = post(`{"pii_default_detectors":["det-a"]}`)
+		Expect(rec.Code).To(Equal(http.StatusOK), rec.Body.String())
+		Expect(app.MITMServer()).ToNot(BeIdenticalTo(srv1),
+			"a default-detector change must restart the listener so it picks up the new detectors")
+	})
+
 	// Residual #9125: enabling the watchdog from a cold (off) state via the
 	// React master toggle must start the live watchdog immediately, without a
 	// restart. The toggle posts watchdog_idle_enabled/busy_enabled=true while

core/http/react-ui/public/locales/id/admin.json

@@ -45,7 +45,7 @@
   },
   "scheduling": {
     "title": "Penjadwalan",
-    "subtitle": "Aturan penempatan model dan replika di seluruh klaster"
+    "subtitle": "Aturan penempatan model dan replika di seluruh kluster"
   },
   "p2p": {
     "title": "Komputasi AI Terdistribusi",
@@ -86,4 +86,4 @@
     "title": "Penjelajah",
     "subtitle": "Jelajahi file dan konfigurasi"
   }
-}
+}

core/http/react-ui/public/locales/id/chat.json

@@ -72,7 +72,7 @@
   "actions": {
     "copy": "Salin",
     "regenerate": "Hasilkan ulang",
-    "jumpToLatest": "Jump to latest"
+    "jumpToLatest": "Lompat ke terbaru"
   },
   "streaming": {
     "transferring": "Mentransfer model...",
@@ -115,4 +115,4 @@
     "clearAll": "Hapus semua",
     "deleteAllTitle": "Hapus semua percakapan"
   }
-}
+}

core/http/react-ui/public/locales/id/common.json

@@ -1,8 +1,8 @@
 {
   "unsaved": {
-    "title": "Discard unsaved changes?",
-    "message": "You have unsaved changes that will be lost if you leave this page.",
-    "leave": "Leave"
+    "title": "Buang perubahan yang belum disimpan?",
+    "message": "Anda memiliki perubahan yang belum disimpan. Perubahan tersebut akan hilang jika Anda meninggalkan halaman ini.",
+    "leave": "Tinggalkan Halaman"
   },
   "actions": {
     "save": "Simpan",

core/http/react-ui/public/locales/id/home.json

@@ -7,15 +7,15 @@
   "resourceGpu": "GPU",
   "resourceRam": "RAM",
   "greeting": {
-    "morning": "Good morning",
-    "afternoon": "Good afternoon",
-    "evening": "Good evening",
-    "night": "Working late"
+    "morning": "Selamat pagi",
+    "afternoon": "Selamat siang",
+    "evening": "Selamat malam",
+    "night": "Selamat lembur"
   },
   "statusLine": {
-    "modelsLoaded_one": "{{count}} model loaded",
-    "modelsLoaded_other": "{{count}} models loaded",
-    "noModelsLoaded": "No models loaded",
+    "modelsLoaded_one": "{{count}} model dimuat",
+    "modelsLoaded_other": "{{count}} model dimuat",
+    "noModelsLoaded": "Tidak ada model yang dimuat",
     "nodes_one": "{{count}} node",
     "nodes_other": "{{count}} nodes"
   },
@@ -79,14 +79,14 @@
   },
   "connect": {
     "title": "Satu endpoint, semua API",
-    "subtitle": "LocalAI menyediakan API miliknya sendiri yang lengkap — pembuatan gambar & video, depth, deteksi objek, reranking, audio, pengenalan wajah & suara, serta suara realtime melalui WebRTC dan WebSocket. Di atas itu, lapisan kompatibilitas drop-in membuat aplikasi apa pun yang dibuat untuk OpenAI, Anthropic, Ollama, atau OpenAI Responses bekerja tanpa perubahan.",
+    "subtitle": "LocalAI menyediakan API miliknya sendiri yang lengkap — pembuatan gambar & video, depth, deteksi objek, reranking, audio, pengenalan wajah & suara, serta suara realtime melalui WebRTC dan WebSocket. Selain itu, lapisan kompatibilitas drop-in membuat aplikasi apa pun yang dibuat untuk OpenAI, Anthropic, Ollama, atau OpenAI Responses bekerja tanpa perubahan.",
     "nativeTitle": "API native",
     "compatTitle": "Kompatibilitas drop-in",
     "apiReference": "Referensi API lengkap",
     "copy": "Salin",
     "copied": "Disalin",
-    "browse": "Browse the API",
-    "hide": "Hide endpoints",
-    "dismiss": "Dismiss"
+    "browse": "Jelajahi API",
+    "hide": "Sembunyikan endpoint",
+    "dismiss": "Abaikan"
   }
 }

core/http/react-ui/public/locales/id/media.json

@@ -5,7 +5,7 @@
       "video": "Video",
       "tts": "TTS",
       "sound": "Suara",
-      "transform": "Transform"
+      "transform": "Transformasi"
     }
   },
   "image": {
@@ -30,7 +30,7 @@
       "refImagesAdded_other": "{{count}} gambar ditambahkan"
     },
     "actions": {
-      "view": "View",
+      "view": "Lihat",
       "generate": "Hasilkan",
       "generating": "Menghasilkan..."
     },
@@ -153,4 +153,4 @@
     "clearConfirm": "Hapus",
     "cleared": "Riwayat dihapus"
   }
-}
+}

core/http/react-ui/public/locales/id/nav.json

@@ -19,11 +19,11 @@
     "operate": "Operasikan"
   },
   "operate": {
-    "inference": "Inference",
-    "cluster": "Cluster",
-    "observability": "Observability",
-    "access": "Access",
-    "system": "System"
+    "inference": "Inferensi",
+    "cluster": "Kluster",
+    "observability": "Observabilitas",
+    "access": "Akses",
+    "system": "Sistem"
   },
   "items": {
     "home": "Beranda",
@@ -64,7 +64,7 @@
     "copyright": "© 2023-{{year}} {{author}}"
   },
   "console": {
-    "automation": "Otomasi",
+    "automation": "Automasi",
     "training": "Pelatihan"
   }
 }

core/services/nodes/inflight.go

@@ -19,25 +19,40 @@ import (
 // Per-replica: a single tracker instance is bound to (nodeID, modelName, replicaIndex).
 // The router constructs one tracker per Route() result, so each in-flight tick lands
 // on the correct row even when multiple replicas of the same model live on the same node.
+//
+// Embedding only grpc.ControlBackend (not the whole grpc.Backend) is what makes
+// the in-flight accounting safe by construction: the control-plane methods pass
+// through untracked, while every grpc.InferenceBackend method must be declared
+// explicitly below to satisfy grpc.Backend. Adding an inference method to the
+// interface therefore breaks this file's build (see the var assertion below)
+// until it is wrapped with track() - so a new inference path can't be added
+// without an in-flight accounting decision.
 type InFlightTrackingClient struct {
-	grpc.Backend // embed for passthrough of untracked methods
-	registry     InFlightTracker
-	nodeID       string
-	modelName    string
-	replicaIndex int
+	grpc.ControlBackend                       // passthrough for control-plane / streaming-constructor methods
+	inner               grpc.InferenceBackend // tracked inference methods delegate here
+	registry            InFlightTracker
+	nodeID              string
+	modelName           string
+	replicaIndex        int
 
 	firstOnce       sync.Once // guards onFirstComplete
 	onFirstComplete func()    // called once after the first tracked inference call completes
 }
 
+// Compile-time contract: *InFlightTrackingClient must implement the FULL backend
+// surface. Because it embeds only ControlBackend, this fails to compile if any
+// InferenceBackend method is left unwrapped.
+var _ grpc.Backend = (*InFlightTrackingClient)(nil)
+
 // NewInFlightTrackingClient wraps a gRPC backend client with in-flight tracking.
 func NewInFlightTrackingClient(inner grpc.Backend, registry InFlightTracker, nodeID, modelName string, replicaIndex int) *InFlightTrackingClient {
 	return &InFlightTrackingClient{
-		Backend:      inner,
-		registry:     registry,
-		nodeID:       nodeID,
-		modelName:    modelName,
-		replicaIndex: replicaIndex,
+		ControlBackend: inner,
+		inner:          inner,
+		registry:       registry,
+		nodeID:         nodeID,
+		modelName:      modelName,
+		replicaIndex:   replicaIndex,
 	}
 }
 
@@ -91,154 +106,162 @@ func (c *InFlightTrackingClient) reconcile(err error) error {
 
 func (c *InFlightTrackingClient) Predict(ctx context.Context, in *pb.PredictOptions, opts ...ggrpc.CallOption) (*pb.Reply, error) {
 	defer c.track(ctx)()
-	reply, err := c.Backend.Predict(ctx, in, opts...)
+	reply, err := c.inner.Predict(ctx, in, opts...)
 	return reply, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) PredictStream(ctx context.Context, in *pb.PredictOptions, f func(reply *pb.Reply), opts ...ggrpc.CallOption) error {
 	defer c.track(ctx)()
-	return c.reconcile(c.Backend.PredictStream(ctx, in, f, opts...))
+	return c.reconcile(c.inner.PredictStream(ctx, in, f, opts...))
 }
 
 func (c *InFlightTrackingClient) Embeddings(ctx context.Context, in *pb.PredictOptions, opts ...ggrpc.CallOption) (*pb.EmbeddingResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Embeddings(ctx, in, opts...)
+	res, err := c.inner.Embeddings(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) GenerateImage(ctx context.Context, in *pb.GenerateImageRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.GenerateImage(ctx, in, opts...)
+	res, err := c.inner.GenerateImage(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) GenerateVideo(ctx context.Context, in *pb.GenerateVideoRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.GenerateVideo(ctx, in, opts...)
+	res, err := c.inner.GenerateVideo(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) TTS(ctx context.Context, in *pb.TTSRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.TTS(ctx, in, opts...)
+	res, err := c.inner.TTS(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) TTSStream(ctx context.Context, in *pb.TTSRequest, f func(reply *pb.Reply), opts ...ggrpc.CallOption) error {
 	defer c.track(ctx)()
-	return c.reconcile(c.Backend.TTSStream(ctx, in, f, opts...))
+	return c.reconcile(c.inner.TTSStream(ctx, in, f, opts...))
 }
 
 func (c *InFlightTrackingClient) SoundGeneration(ctx context.Context, in *pb.SoundGenerationRequest, opts ...ggrpc.CallOption) (*pb.Result, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.SoundGeneration(ctx, in, opts...)
+	res, err := c.inner.SoundGeneration(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioTranscription(ctx context.Context, in *pb.TranscriptRequest, opts ...ggrpc.CallOption) (*pb.TranscriptResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.AudioTranscription(ctx, in, opts...)
+	res, err := c.inner.AudioTranscription(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioTranscriptionStream(ctx context.Context, in *pb.TranscriptRequest, f func(chunk *pb.TranscriptStreamResponse), opts ...ggrpc.CallOption) error {
 	defer c.track(ctx)()
-	return c.reconcile(c.Backend.AudioTranscriptionStream(ctx, in, f, opts...))
+	return c.reconcile(c.inner.AudioTranscriptionStream(ctx, in, f, opts...))
 }
 
 func (c *InFlightTrackingClient) Detect(ctx context.Context, in *pb.DetectOptions, opts ...ggrpc.CallOption) (*pb.DetectResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Detect(ctx, in, opts...)
+	res, err := c.inner.Detect(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) Depth(ctx context.Context, in *pb.DepthRequest, opts ...ggrpc.CallOption) (*pb.DepthResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Depth(ctx, in, opts...)
+	res, err := c.inner.Depth(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) Rerank(ctx context.Context, in *pb.RerankRequest, opts ...ggrpc.CallOption) (*pb.RerankResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Rerank(ctx, in, opts...)
+	res, err := c.inner.Rerank(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) VAD(ctx context.Context, in *pb.VADRequest, opts ...ggrpc.CallOption) (*pb.VADResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.VAD(ctx, in, opts...)
+	res, err := c.inner.VAD(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) Diarize(ctx context.Context, in *pb.DiarizeRequest, opts ...ggrpc.CallOption) (*pb.DiarizeResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Diarize(ctx, in, opts...)
+	res, err := c.inner.Diarize(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) FaceVerify(ctx context.Context, in *pb.FaceVerifyRequest, opts ...ggrpc.CallOption) (*pb.FaceVerifyResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.FaceVerify(ctx, in, opts...)
+	res, err := c.inner.FaceVerify(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) FaceAnalyze(ctx context.Context, in *pb.FaceAnalyzeRequest, opts ...ggrpc.CallOption) (*pb.FaceAnalyzeResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.FaceAnalyze(ctx, in, opts...)
+	res, err := c.inner.FaceAnalyze(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) VoiceVerify(ctx context.Context, in *pb.VoiceVerifyRequest, opts ...ggrpc.CallOption) (*pb.VoiceVerifyResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.VoiceVerify(ctx, in, opts...)
+	res, err := c.inner.VoiceVerify(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) VoiceAnalyze(ctx context.Context, in *pb.VoiceAnalyzeRequest, opts ...ggrpc.CallOption) (*pb.VoiceAnalyzeResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.VoiceAnalyze(ctx, in, opts...)
+	res, err := c.inner.VoiceAnalyze(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) VoiceEmbed(ctx context.Context, in *pb.VoiceEmbedRequest, opts ...ggrpc.CallOption) (*pb.VoiceEmbedResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.VoiceEmbed(ctx, in, opts...)
+	res, err := c.inner.VoiceEmbed(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) TokenClassify(ctx context.Context, in *pb.TokenClassifyRequest, opts ...ggrpc.CallOption) (*pb.TokenClassifyResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.TokenClassify(ctx, in, opts...)
+	res, err := c.inner.TokenClassify(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) Score(ctx context.Context, in *pb.ScoreRequest, opts ...ggrpc.CallOption) (*pb.ScoreResponse, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.Score(ctx, in, opts...)
+	res, err := c.inner.Score(ctx, in, opts...)
+	return res, c.reconcile(err)
+}
+
+func (c *InFlightTrackingClient) SoundDetection(ctx context.Context, in *pb.SoundDetectionRequest, opts ...ggrpc.CallOption) (*pb.SoundDetectionResponse, error) {
+	defer c.track(ctx)()
+	res, err := c.inner.SoundDetection(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioEncode(ctx context.Context, in *pb.AudioEncodeRequest, opts ...ggrpc.CallOption) (*pb.AudioEncodeResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.AudioEncode(ctx, in, opts...)
+	res, err := c.inner.AudioEncode(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioDecode(ctx context.Context, in *pb.AudioDecodeRequest, opts ...ggrpc.CallOption) (*pb.AudioDecodeResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.AudioDecode(ctx, in, opts...)
+	res, err := c.inner.AudioDecode(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
 func (c *InFlightTrackingClient) AudioTransform(ctx context.Context, in *pb.AudioTransformRequest, opts ...ggrpc.CallOption) (*pb.AudioTransformResult, error) {
 	defer c.track(ctx)()
-	res, err := c.Backend.AudioTransform(ctx, in, opts...)
+	res, err := c.inner.AudioTransform(ctx, in, opts...)
 	return res, c.reconcile(err)
 }
 
-// AudioTransformStream, AudioToAudioStream and Forward are deliberately left as
-// embedded passthrough: they return a stream client and the inference spans the
-// stream's lifetime, not the constructor call. Wrapping the constructor with
-// track() would increment and immediately decrement (and fire onFirstComplete)
-// before any audio flows. Tracking those correctly needs the done() func tied to
-// stream close, which the current Backend interface doesn't surface here.
+// AudioTransformStream, AudioToAudioStream and Forward live in grpc.ControlBackend
+// and are passed through via the embedded field, NOT tracked: they return a stream
+// client and the inference spans the stream's lifetime, not the constructor call.
+// Wrapping the constructor with track() would increment and immediately decrement
+// (and fire onFirstComplete) before any audio flows. Tracking those correctly needs
+// the done() func tied to stream close, which the Backend interface doesn't surface
+// here. If they ever need tracking, move them to grpc.InferenceBackend - the build
+// will then force an explicit wrapper here.

core/services/nodes/inflight_test.go

@@ -408,6 +408,13 @@ var _ = Describe("InFlightTrackingClient", func() {
 				return err
 			})
 		})
+
+		It("SoundDetection", func() {
+			assertTracked(func() error {
+				_, err := client.SoundDetection(context.Background(), &pb.SoundDetectionRequest{})
+				return err
+			})
+		})
 	})
 
 	Describe("stale model reload (self-heal)", func() {

docs/content/features/middleware.md

@@ -185,6 +185,13 @@ It is persisted through `POST /api/settings` and read live, so a change takes
 effect on the next request without a restart. A default that names a model no
 longer loaded still appears (marked *not loaded*) so it can be toggled off.
 
+The default set can also be supplied out-of-band with the
+`LOCALAI_PII_DEFAULT_DETECTORS` environment variable (comma-separated model
+names, e.g. `privacy-filter-nemotron,secret-filter`). When set it takes
+precedence over the value persisted via the UI (env > file), which is the
+right behaviour for immutable container deployments that pin filtering policy
+at boot rather than via the admin UI.
+
 This is what makes `cloud-proxy` / MITM redaction work out of the box: those
 backends default to PII-enabled but ship no detector list, so without a
 default detector the filter runs with nothing to scan. Set one here and

gallery/index.yaml

@@ -1,4 +1,54 @@
 ---
+- name: "lfm2.5-1.2b-instruct"
+  url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
+  urls:
+    - https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF
+  description: |
+    Try LFM • Docs • LEAP • Discord
+
+    # LFM2.5-1.2B-Instruct
+
+    LFM2.5 is a new family of hybrid models designed for **on-device deployment**. It builds on the LFM2 architecture with extended pre-training and reinforcement learning.
+
+      - **Best-in-class performance**: A 1.2B model rivaling much larger models, bringing high-quality AI to your pocket.
+      - **Fast edge inference**: 239 tok/s decode on AMD CPU, 82 tok/s on mobile NPU. Runs under 1GB of memory with day-one support for llama.cpp, MLX, and vLLM.
+      - **Scaled training**: Extended pre-training from 10T to 28T tokens and large-scale multi-stage reinforcement learning.
+
+    Find more information about LFM2.5 in our blog post.
+
+    ## 🗒️ Model Details
+
+    LFM2.5-1.2B-Instruct is a general-purpose text-only model with the following features:
+
+    ...
+  license: "other"
+  tags:
+    - llm
+    - gguf
+  icon: https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/dxnYF2fuLpulismtFSGFi.png
+  overrides:
+    backend: llama-cpp
+    function:
+      automatic_tool_parsing_fallback: true
+      grammar:
+        disable: true
+    known_usecases:
+      - chat
+    options:
+      - use_jinja:true
+    parameters:
+      min_p: 0.15
+      model: llama-cpp/models/LFM2.5-1.2B-Instruct-GGUF/LFM2.5-1.2B-Instruct-Q4_K_M.gguf
+      repeat_penalty: 1.05
+      temperature: 0.1
+      top_k: 50
+      top_p: 0.1
+    template:
+      use_tokenizer_template: true
+  files:
+    - filename: llama-cpp/models/LFM2.5-1.2B-Instruct-GGUF/LFM2.5-1.2B-Instruct-Q4_K_M.gguf
+      sha256: b1b3de114215d9507409a662a501a631095a479a419584e8a2ded6304b19b4f5
+      uri: https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF/resolve/main/LFM2.5-1.2B-Instruct-Q4_K_M.gguf
 - name: "qwopus3.6-27b-coder-compat-mtp"
   url: "github:mudler/LocalAI/gallery/virtual.yaml@master"
   urls:

pkg/grpc/backend.go

@@ -41,11 +41,34 @@ func buildClient(address string, parallel bool, wd WatchDog, enableWatchDog bool
 	}
 }
 
+// Backend is the full client surface of a model backend. It is deliberately
+// composed of two sub-interfaces so that wrappers can get a COMPILE-TIME
+// guarantee about which methods they must account for:
+//
+//   - InferenceBackend - methods that each perform one discrete inference call
+//     (the call begins on entry and ends on return). A wrapper that does
+//     per-call accounting - e.g. the distributed router's in-flight tracker,
+//     core/services/nodes.InFlightTrackingClient - embeds only ControlBackend
+//     and implements every InferenceBackend method explicitly. Adding a method
+//     to InferenceBackend therefore breaks that wrapper's build until it is
+//     implemented: inference can't be added without an accounting decision.
+//   - ControlBackend - everything that is NOT a discrete inference call:
+//     lifecycle/control-plane operations and the streaming constructors whose
+//     work spans the returned stream rather than the constructor call. These
+//     are safe to pass through untracked.
+//
+// Keep the two sets disjoint; every backend method belongs to exactly one.
 type Backend interface {
-	IsBusy() bool
-	HealthCheck(ctx context.Context) (bool, error)
+	InferenceBackend
+	ControlBackend
+}
+
+// InferenceBackend is the subset of Backend whose methods each map to a single
+// inference call. Wrappers that account for in-flight work must implement these
+// explicitly (see Backend). Do NOT add methods that return a stream client or
+// that are control-plane only - those belong in ControlBackend.
+type InferenceBackend interface {
 	Embeddings(ctx context.Context, in *pb.PredictOptions, opts ...grpc.CallOption) (*pb.EmbeddingResult, error)
-	LoadModel(ctx context.Context, in *pb.ModelOptions, opts ...grpc.CallOption) (*pb.Result, error)
 	PredictStream(ctx context.Context, in *pb.PredictOptions, f func(reply *pb.Reply), opts ...grpc.CallOption) error
 	Predict(ctx context.Context, in *pb.PredictOptions, opts ...grpc.CallOption) (*pb.Reply, error)
 	GenerateImage(ctx context.Context, in *pb.GenerateImageRequest, opts ...grpc.CallOption) (*pb.Result, error)
@@ -53,6 +76,8 @@ type Backend interface {
 	TTS(ctx context.Context, in *pb.TTSRequest, opts ...grpc.CallOption) (*pb.Result, error)
 	TTSStream(ctx context.Context, in *pb.TTSRequest, f func(reply *pb.Reply), opts ...grpc.CallOption) error
 	SoundGeneration(ctx context.Context, in *pb.SoundGenerationRequest, opts ...grpc.CallOption) (*pb.Result, error)
+	AudioTranscription(ctx context.Context, in *pb.TranscriptRequest, opts ...grpc.CallOption) (*pb.TranscriptResult, error)
+	AudioTranscriptionStream(ctx context.Context, in *pb.TranscriptRequest, f func(chunk *pb.TranscriptStreamResponse), opts ...grpc.CallOption) error
 	Detect(ctx context.Context, in *pb.DetectOptions, opts ...grpc.CallOption) (*pb.DetectResponse, error)
 	Depth(ctx context.Context, in *pb.DepthRequest, opts ...grpc.CallOption) (*pb.DepthResponse, error)
 	FaceVerify(ctx context.Context, in *pb.FaceVerifyRequest, opts ...grpc.CallOption) (*pb.FaceVerifyResponse, error)
@@ -60,8 +85,25 @@ type Backend interface {
 	VoiceVerify(ctx context.Context, in *pb.VoiceVerifyRequest, opts ...grpc.CallOption) (*pb.VoiceVerifyResponse, error)
 	VoiceAnalyze(ctx context.Context, in *pb.VoiceAnalyzeRequest, opts ...grpc.CallOption) (*pb.VoiceAnalyzeResponse, error)
 	VoiceEmbed(ctx context.Context, in *pb.VoiceEmbedRequest, opts ...grpc.CallOption) (*pb.VoiceEmbedResponse, error)
-	AudioTranscription(ctx context.Context, in *pb.TranscriptRequest, opts ...grpc.CallOption) (*pb.TranscriptResult, error)
-	AudioTranscriptionStream(ctx context.Context, in *pb.TranscriptRequest, f func(chunk *pb.TranscriptStreamResponse), opts ...grpc.CallOption) error
+	Rerank(ctx context.Context, in *pb.RerankRequest, opts ...grpc.CallOption) (*pb.RerankResult, error)
+	TokenClassify(ctx context.Context, in *pb.TokenClassifyRequest, opts ...grpc.CallOption) (*pb.TokenClassifyResponse, error)
+	Score(ctx context.Context, in *pb.ScoreRequest, opts ...grpc.CallOption) (*pb.ScoreResponse, error)
+	VAD(ctx context.Context, in *pb.VADRequest, opts ...grpc.CallOption) (*pb.VADResponse, error)
+	Diarize(ctx context.Context, in *pb.DiarizeRequest, opts ...grpc.CallOption) (*pb.DiarizeResponse, error)
+	SoundDetection(ctx context.Context, in *pb.SoundDetectionRequest, opts ...grpc.CallOption) (*pb.SoundDetectionResponse, error)
+	AudioEncode(ctx context.Context, in *pb.AudioEncodeRequest, opts ...grpc.CallOption) (*pb.AudioEncodeResult, error)
+	AudioDecode(ctx context.Context, in *pb.AudioDecodeRequest, opts ...grpc.CallOption) (*pb.AudioDecodeResult, error)
+	AudioTransform(ctx context.Context, in *pb.AudioTransformRequest, opts ...grpc.CallOption) (*pb.AudioTransformResult, error)
+}
+
+// ControlBackend is the subset of Backend that is NOT per-call inference:
+// lifecycle/control-plane operations and the streaming constructors whose work
+// spans the returned stream rather than the constructor call. In-flight-tracking
+// wrappers embed this directly and pass it through untracked (see Backend).
+type ControlBackend interface {
+	IsBusy() bool
+	HealthCheck(ctx context.Context) (bool, error)
+	LoadModel(ctx context.Context, in *pb.ModelOptions, opts ...grpc.CallOption) (*pb.Result, error)
 	TokenizeString(ctx context.Context, in *pb.PredictOptions, opts ...grpc.CallOption) (*pb.TokenizationResponse, error)
 	Status(ctx context.Context) (*pb.StatusResponse, error)
 
@@ -70,24 +112,11 @@ type Backend interface {
 	StoresGet(ctx context.Context, in *pb.StoresGetOptions, opts ...grpc.CallOption) (*pb.StoresGetResult, error)
 	StoresFind(ctx context.Context, in *pb.StoresFindOptions, opts ...grpc.CallOption) (*pb.StoresFindResult, error)
 
-	Rerank(ctx context.Context, in *pb.RerankRequest, opts ...grpc.CallOption) (*pb.RerankResult, error)
-
-	TokenClassify(ctx context.Context, in *pb.TokenClassifyRequest, opts ...grpc.CallOption) (*pb.TokenClassifyResponse, error)
-
-	Score(ctx context.Context, in *pb.ScoreRequest, opts ...grpc.CallOption) (*pb.ScoreResponse, error)
-
 	GetTokenMetrics(ctx context.Context, in *pb.MetricsRequest, opts ...grpc.CallOption) (*pb.MetricsResponse, error)
 
-	VAD(ctx context.Context, in *pb.VADRequest, opts ...grpc.CallOption) (*pb.VADResponse, error)
-
-	Diarize(ctx context.Context, in *pb.DiarizeRequest, opts ...grpc.CallOption) (*pb.DiarizeResponse, error)
-
-	SoundDetection(ctx context.Context, in *pb.SoundDetectionRequest, opts ...grpc.CallOption) (*pb.SoundDetectionResponse, error)
-
-	AudioEncode(ctx context.Context, in *pb.AudioEncodeRequest, opts ...grpc.CallOption) (*pb.AudioEncodeResult, error)
-	AudioDecode(ctx context.Context, in *pb.AudioDecodeRequest, opts ...grpc.CallOption) (*pb.AudioDecodeResult, error)
-
-	AudioTransform(ctx context.Context, in *pb.AudioTransformRequest, opts ...grpc.CallOption) (*pb.AudioTransformResult, error)
+	// Streaming constructors: these return a stream client immediately; the
+	// actual inference spans the stream's lifetime, not this call, so they are
+	// NOT tracked as a single in-flight unit.
 	AudioTransformStream(ctx context.Context, opts ...grpc.CallOption) (AudioTransformStreamClient, error)
 	AudioToAudioStream(ctx context.Context, opts ...grpc.CallOption) (AudioToAudioStreamClient, error)