Token Rate Limit Hierarchy

# Hierarchical token budgets
#
# Optional org, then team, then user budgets on one token_rate_limit
# admission (ai#125). Each level is its own hard deny. A user who is
# under their own budget is still rejected when the org or team budget
# cannot take the same estimate.
#
# The rule list includes a catch-all (no match) and reserved_tokens, so
# every request that reaches the filter is estimated and checked against
# all three levels. A config that omits the catch-all, or that uses a
# body estimate with no fallback_estimate, is rejected at load time.
#
# Org and team values are headers. This filter hashes whatever value it
# receives. It does not prove who wrote the header: AuthenticatedIdentity
# carries the user subject and no org or team claim. A trusted component
# in front of this proxy must strip a client-supplied x-org-id or
# x-team-id and set its own. A client that can still set those headers
# can choose another bucket and can fill max_keys. basic_auth below
# only establishes the user level.
#
# Usage:
#   cargo run -p praxis-ai-proxy --features basic-auth-filter,token-rate-limit-filter -- -c examples/configs/token-rate-limit-hierarchy.yaml
#   curl -i http://localhost:8080/v1/chat/completions \
#     -u subject-a:example-password \
#     -H "x-org-id: acme" \
#     -H "x-team-id: engineering" \
#     -d '{"model":"gpt-4","messages":[{"role":"user","content":"hi"}]}'

listeners:
  - name: default
    address: "127.0.0.1:8080"
    filter_chains:
      - main

filter_chains:
  - name: main
    filters:
      - filter: basic_auth
        strip_authorization: true
        credentials:
          - username: subject-a
            password: "example-password"

      - filter: token_rate_limit
        hierarchy:
          - level: org
            identity_header: x-org-id
            algorithm: sliding_window
            window: 1h
            capacity: 5000000
          - level: team
            identity_header: x-team-id
            algorithm: sliding_window
            window: 1h
            capacity: 1000000
          - level: user
            identity: authenticated_subject
            algorithm: sliding_window
            window: 1h
            capacity: 100000
        rules:
          - name: default
            algorithm: sliding_window
            window: 1h
            capacity: 10000000
            reserved_tokens: 500

      - filter: router
        routes:
          - path: "/v1/chat/completions"
            cluster: backend

      - filter: token_count
        provider: openai

      - filter: access_log

      - filter: load_balancer
        clusters:
          - name: backend
            endpoints:
              - "127.0.0.1:3000"

insecure_options:
  allow_private_endpoints: true # example proxies to a local backend