Intelligent Route All Capabilities

# Intelligent Route: Complete Static Routing
#
# Demonstrates every candidate capability and selection input handled by
# intelligent_route today:
#
# - inference_model and mcp_tool candidate matching;
# - producer-rendered order for locality and freshness decisions; and
# - configured order as the deterministic request-path selection rule.
#
# Queue depth, cache utilization, latency, admission, and cost are not read on
# the request path. A control-plane producer can use those signals to decide
# candidate eligibility and order before writing this configuration. The
# request path preserves that order rather than recomputing those scores.
#
# The mcp filter classifies MCP tools/call requests and continues past non-MCP
# bodies. model_to_header promotes the model from inference requests and strips
# any spoofed client copy. intelligent_route gives MCP metadata precedence when
# both inputs are present.
#
listeners:
  - name: proxy
    address: "0.0.0.0:8080"
    filter_chains:
      - main

filter_chains:
  - name: main
    filters:
      - filter: mcp
        on_invalid: continue

      # Promote the JSON body's `model` field into X-Model and strip any
      # client-supplied X-Model so the routed model cannot be spoofed.
      - filter: model_to_header
        header: X-Model

      - filter: intelligent_route
        local_site: site-a
        model_header: X-Model
        candidates:
          # The producer placed the eligible local model first.
          - kind: inference_model
            name: granite-3.3-8b
            site: site-a
            cluster: models-site-a
            fresh: true

          - kind: inference_model
            name: granite-3.3-8b
            site: site-b
            cluster: models-site-b
            fresh: true

          # The producer placed the fresh remote model before the stale local
          # model.
          - kind: inference_model
            name: llama-3.2-8b
            site: site-b
            cluster: models-site-b
            fresh: true

          - kind: inference_model
            name: llama-3.2-8b
            site: site-a
            cluster: models-site-a
            fresh: false

          # MCP tool candidates preserve the same producer-rendered order.
          - kind: mcp_tool
            name: weather-lookup
            site: site-b
            cluster: tools-site-b
            fresh: true

          - kind: mcp_tool
            name: weather-lookup
            site: site-a
            cluster: tools-site-a
            fresh: false

          # Equal score: the first configured local candidate wins.
          - kind: mcp_tool
            name: code-search
            site: site-a
            cluster: tools-site-a
            fresh: true

          - kind: mcp_tool
            name: code-search
            site: site-a
            cluster: tools-site-a-secondary
            fresh: true

      - filter: load_balancer
        clusters:
          - name: models-site-a
            endpoints:
              - "127.0.0.1:8001"
          - name: models-site-b
            endpoints:
              - "127.0.0.1:8002"
          - name: tools-site-a
            endpoints:
              - "127.0.0.1:8003"
          - name: tools-site-a-secondary
            endpoints:
              - "127.0.0.1:8004"
          - name: tools-site-b
            endpoints:
              - "127.0.0.1:8005"

admin:
  address: "127.0.0.1:9901"
shutdown_timeout_secs: 5

insecure_options:
  allow_private_endpoints: true # example proxies to local backends