vllm-sr 0.3.0.dev20260427140612

vLLM Semantic Router - Intelligent routing for Mixture-of-Models

vLLM Semantic Router

Intelligent Router for Mixture-of-Models (MoM).

GitHub: https://github.com/vllm-project/semantic-router

Quick Start

Installation

# Install from PyPI
pip install vllm-sr

# Or install from source (development)
cd src/vllm-sr
pip install -e .

Usage

# Start the router (includes dashboard, simulator sidecar, and first-run setup)
HF_TOKEN=hf_xxx vllm-sr serve

# Start an isolated second local stack on offset host ports
VLLM_SR_STACK_NAME=lane-b VLLM_SR_PORT_OFFSET=200 HF_TOKEN=hf_xxx vllm-sr serve

# Open the dashboard
# http://localhost:8700
# second stack example: http://localhost:8900

# Optional: open the dashboard in your browser
vllm-sr dashboard

# View logs
vllm-sr logs router
vllm-sr logs envoy
vllm-sr logs dashboard
vllm-sr logs simulator

# Evaluate how signals fire for a prompt (requires: vllm-sr serve)
# Single prompt — readable summary (default)
vllm-sr eval --prompt "Explain inflation vs recession in plain English."
# decision: economics
# used signals: 3
#   - domain:economics
#   - keyword:inflation
#   - embedding:price_movement
# matched signals: 3
#   - domains:economics
#   - keywords:inflation
#   - embeddings:price_movement
# unmatched signals: 3
# signal confidences:
#   - domain:economics: 0.95
#   - keyword:inflation: 0.87
#   - embedding:price_movement: 0.82
# routing: economics

# Single prompt — full JSON payload
vllm-sr eval --prompt "Explain inflation vs recession in plain English." --json

# Multi-turn messages array (OpenAI chat format) — readable summary
vllm-sr eval --messages '[{"role":"system","content":"You are a careful tutor."},{"role":"user","content":"Explain inflation vs recession in plain English."}]'

# Multi-turn messages array — full JSON payload
vllm-sr eval --messages '[{"role":"system","content":"You are a careful tutor."},{"role":"user","content":"Explain inflation vs recession in plain English."}]' --json

# Override endpoint (e.g. remote stack or non-default port)
vllm-sr eval --prompt "hello" --endpoint http://localhost:8080

# Common errors:
#   Router not started:
#     ERROR - Router is not running at http://localhost:8080/api/v1/eval. Start the router with 'vllm-sr serve' and retry.
#   Wrong port (hitting a proxy instead of the router API):
#     ERROR - Router returned 403 from http://localhost:8080/api/v1/eval. This looks like a proxy or gateway —
#             check that --endpoint points directly to the router API port (default: 8080), not to Envoy or another proxy.
#   Invalid request body (400):
#     ERROR - Router returned 400 INVALID_INPUT: text cannot be empty
#   Service unavailable (503):
#     ERROR - Router returned 503 SERVICE_UNAVAILABLE: classifier not ready

# Check status
vllm-sr status

# Send a one-shot chat completion through Envoy (default model: MoM)
vllm-sr chat "hello"
vllm-sr chat --json "hello"
# Remote or port-forwarded stack
vllm-sr chat --base-url http://localhost:8080 "hello"

# Stop
vllm-sr stop

Kubernetes Deployment

The same CLI deploys to Kubernetes via Helm:

# Deploy to Kubernetes (uses your existing config.yaml)
HF_TOKEN=hf_xxx vllm-sr serve --target k8s --profile dev --config config.yaml

# Deploy to a specific namespace and context
HF_TOKEN=hf_xxx vllm-sr serve --target k8s --namespace production --context prod-cluster

# Check status / logs / stop
vllm-sr status --target k8s
vllm-sr logs router --target k8s -f
vllm-sr stop --target k8s

# Chat completion against a port-forwarded or ingress URL (requires --base-url)
vllm-sr chat --base-url http://localhost:8080 "hello"

Credential handling: Sensitive environment variables (HF_TOKEN, OPENAI_API_KEY, ANTHROPIC_API_KEY) are automatically stored in a Kubernetes Secret (vllm-sr-env-secrets) and mounted via envFrom. They never appear as plain-text values in Helm overrides or the Deployment spec. Non-sensitive variables (HF_ENDPOINT, HF_HOME, etc.) are passed as standard env entries.

The secret is created before helm upgrade --install and cleaned up by vllm-sr stop --target k8s.

If you start in an empty directory, vllm-sr serve bootstraps a minimal workspace and opens the dashboard in setup mode. Configure your first model there, then activate routing.

Local dashboard state is persisted under .vllm-sr/dashboard-data/ and bind-mounted into the container at /app/data. User accounts, evaluation history, and ML pipeline artifacts survive vllm-sr stop followed by a new vllm-sr serve as long as that workspace directory is kept.

The fleet simulator sidecar is started on the same runtime network by default. The dashboard backend proxies it at /api/fleet-sim/*, and the dashboard exposes its workflows under the Fleet Sim top-bar dropdown.

To run parallel local stacks from the same machine or multiple worktrees, set VLLM_SR_STACK_NAME and VLLM_SR_PORT_OFFSET before vllm-sr serve, vllm-sr status, vllm-sr dashboard, and vllm-sr stop. The stack name isolates container and network names, and the port offset shifts the published host ports while keeping internal container ports unchanged.

Advanced YAML-first setup

# Validate a hand-authored canonical config before serving
vllm-sr validate config.yaml

vllm-sr init was removed in v0.3. Author config.yaml directly using the canonical version/listeners/providers/routing/global layout, migrate an older file with vllm-sr config migrate --config old-config.yaml, or import supported OpenClaw model providers with vllm-sr config import --from openclaw. Router-wide defaults come from the router itself and can be overridden under global:.

Features

• Router: Intelligent request routing based on intent classification
• Envoy Proxy: High-performance proxy with ext_proc integration
• Dashboard: Web UI for monitoring and testing (http://localhost:8700)
• Metrics: Prometheus metrics endpoint (http://localhost:9190/metrics)

Endpoints

After running vllm-sr serve, the following endpoints are available:

Endpoint	Port	Description
Dashboard	8700	Web UI for monitoring and Playground
API	8888*	Chat completions API (configurable in config.yaml)
Metrics	9190	Prometheus metrics
gRPC	50051	Router gRPC (internal)
Jaeger UI	16686	Distributed tracing UI
Grafana (embedded)	8700	Dashboards at /embedded/grafana
Prometheus UI	9090	Metrics storage and querying

*Default port, configurable via listeners in config.yaml

Observability

vllm-sr serve automatically starts the observability stack:

• Jaeger: Distributed tracing embedded at http://localhost:8700/embedded/jaeger (also available directly at http://localhost:16686)
• Grafana: Pre-configured dashboards embedded at http://localhost:8700/embedded/grafana
• Prometheus: Metrics collection at http://localhost:9090

Note: Grafana is optimized for embedded access through the dashboard. For the best experience, use http://localhost:8700/embedded/grafana where anonymous authentication is pre-configured.

Tracing is enabled by default. Traces are visible in Jaeger under the vllm-sr service name.

Configuration

Plugin Configuration

The CLI supports configuring plugins in your routing decisions. Plugins are per-decision behaviors that customize request handling (security, caching, customization, debugging).

Supported Plugin Types:

• semantic-cache - Cache similar requests for performance
• memory - Retrieve and store route-local conversation memory
• system_prompt - Inject custom system prompts
• header_mutation - Add/modify HTTP headers
• hallucination - Detect hallucinations in responses
• router_replay - Record routing decisions for debugging
• rag - Inject retrieved knowledge into prompts
• image_gen - Hand a matched route off to an image generation backend
• fast_response - Return a route-local response immediately
• request_params - Sanitize or cap request body parameters before forwarding
• response_jailbreak - Screen model output before returning it
• tools - Restrict or curate tool access per route

Plugin Examples:

Each example shows the plugin list inside a canonical routing.decisions[] entry.

1. semantic-cache - Cache similar requests:

routing:
  decisions:
    - name: "cached-route"
      plugins:
        - type: "semantic-cache"
          configuration:
            enabled: true
            similarity_threshold: 0.92  # 0.0-1.0, higher = more strict
            ttl_seconds: 3600  # Optional: cache TTL in seconds

2. fast_response - Return a route-local response:

routing:
  decisions:
    - name: "guarded-route"
      plugins:
        - type: "fast_response"
          configuration:
            message: "This request was blocked by the matched route policy."

3. system_prompt - Inject custom instructions:

routing:
  decisions:
    - name: "persona-route"
      plugins:
        - type: "system_prompt"
          configuration:
            enabled: true
            system_prompt: "You are a helpful assistant."
            mode: "replace"  # "replace" (default) or "insert" (prepend)

4. header_mutation - Modify HTTP headers:

routing:
  decisions:
    - name: "header-route"
      plugins:
        - type: "header_mutation"
          configuration:
            add:
              - name: "X-Custom-Header"
                value: "custom-value"
            update:
              - name: "User-Agent"
                value: "SemanticRouter/1.0"
            delete:
              - "X-Old-Header"

5. hallucination - Detect hallucinations:

routing:
  decisions:
    - name: "fact-check-route"
      plugins:
        - type: "hallucination"
          configuration:
            enabled: true
            use_nli: false  # Optional: use NLI for detailed analysis
            hallucination_action: "header"  # "header", "body", or "none"

6. router_replay - Record decisions for debugging:

routing:
  decisions:
    - name: "debug-route"
      plugins:
        - type: "router_replay"
          configuration:
            enabled: true
            max_records: 10000  # Optional: max records in memory (default: 10000)
            capture_request_body: true  # Optional: capture request payloads (default: true)
            capture_response_body: true  # Optional: capture response payloads (default: true)
            max_body_bytes: 4096  # Optional: max bytes to capture (default: 4096)

7. memory - Retrieve route-local memory:

routing:
  decisions:
    - name: "memory-route"
      plugins:
        - type: "memory"
          configuration:
            enabled: true
            retrieval_limit: 5
            similarity_threshold: 0.75
            auto_store: true

8. rag - Inject retrieved context:

routing:
  decisions:
    - name: "knowledge-route"
      plugins:
        - type: "rag"
          configuration:
            enabled: true
            backend: "milvus"
            top_k: 5
            similarity_threshold: 0.8

9. tools - Restrict available tools:

routing:
  decisions:
    - name: "tool-route"
      plugins:
        - type: "tools"
          configuration:
            enabled: true
            mode: "filtered"
            allow_tools: ["search_web"]
            block_tools: ["exec_cmd"]

10. image_gen - Route to an image backend:

routing:
  decisions:
    - name: "image-route"
      plugins:
        - type: "image_gen"
          configuration:
            enabled: true
            backend: "vllm_omni"
            backend_config:
              base_url: "http://image-router:8005"

11. request_params - Cap or strip request parameters:

routing:
  decisions:
    - name: "budget-route"
      plugins:
        - type: "request_params"
          configuration:
            blocked_params: ["logprobs", "top_logprobs"]
            max_tokens_limit: 512
            max_n: 1
            strip_unknown: true

12. response_jailbreak - Screen generated output:

routing:
  decisions:
    - name: "safety-route"
      plugins:
        - type: "response_jailbreak"
          configuration:
            enabled: true
            threshold: 0.8
            action: "header"

Router replay records are exposed through:

• GET /v1/router_replay?limit=20&offset=0&search=req-123&decision=foo&model=bar&cache_status=cached - List recent records with pagination metadata. Default page size is 20; larger limit values are capped at 100.
• GET /v1/router_replay/aggregate?search=req-123&decision=foo&model=bar&cache_status=cached - Return summary and chart aggregates for the filtered replay set.
• GET /v1/router_replay/{id} - Fetch a single replay record.

If a replay page would exceed the ext-proc gRPC message budget, the router returns 413 Payload Too Large instead of failing the stream.

Validation Rules:

• Plugin Type: Must be one of: semantic-cache, memory, system_prompt, header_mutation, hallucination, router_replay, rag, image_gen, fast_response, request_params, response_jailbreak, tools
• enabled: Must be a boolean (required for most plugins)
• similarity_threshold/min_confidence_threshold: Must be a float between 0.0 and 1.0
• max_records/max_body_bytes: Must be a positive integer
• ttl_seconds: Must be a non-negative integer
• system_prompt: Must be a string (if provided)
• mode: Must be "replace" or "insert" (if provided)
• injection_mode: Must be tool_role or system_prompt (if provided)
• on_failure: Must be skip, block, or warn (if provided)
• action: Must be block, header, or none (if provided)

CLI Commands:

# Validate configuration (including plugins)
vllm-sr validate

# Migrate older configs to the canonical contract
vllm-sr config migrate --config old-config.yaml

# Import supported OpenClaw model providers into canonical config.yaml
vllm-sr config import --from openclaw --source openclaw.json --target config.yaml

File Descriptor Limits

The CLI automatically sets file descriptor limits to 65,536 for Envoy proxy. To customize:

export VLLM_SR_NOFILE_LIMIT=100000  # Optional (min: 8192)
vllm-sr serve

License

Apache 2.0