config.yaml

version: v0.3

listeners:
  - name: http-8899
    address: 0.0.0.0
    port: 8899
    timeout: 300s

providers:
  defaults:
    default_model: qwen3-8b
    default_reasoning_effort: medium
    reasoning_families:
      qwen3:
        type: chat_template_kwargs
        parameter: enable_thinking
      gpt:
        type: reasoning_effort
        parameter: reasoning.effort
  models:
    - name: qwen3-8b
      reasoning_family: qwen3
      provider_model_id: qwen3-8b-instruct
      api_format: openai
      pricing:
        currency: USD
        prompt_per_1m: 0.18
        cached_input_per_1m: 0.04
        cache_write_per_1m: 0.225
        completion_per_1m: 0.42
      external_model_ids:
        openai: qwen3-8b
        analytics: qwen3-routing
      backend_refs:
        - name: local-primary
          endpoint: 127.0.0.1:8000
          protocol: http
          weight: 80
          type: chat
          api_key_env: VLLM_SR_PRIMARY_API_KEY
        - name: remote-secondary
          base_url: https://api.example.com/v1
          provider: openai
          auth_header: Authorization
          auth_prefix: Bearer
          extra_headers:
            X-Tenant: reference-config
          api_version: "2025-03-01"
          chat_path: /chat/completions
          api_key: static-demo-key
          weight: 20
    - name: qwen3-32b
      reasoning_family: gpt
      provider_model_id: qwen3-32b-instruct
      api_format: anthropic
      pricing:
        currency: USD
        prompt_per_1m: 1.2
        cached_input_per_1m: 0.3
        completion_per_1m: 3.4
      external_model_ids:
        internal: qwen3-32b
      backend_refs:
        - name: large-primary
          endpoint: 127.0.0.1:8001
          protocol: http
          weight: 100
    - name: llava-omni
      provider_model_id: llava-omni-preview
      api_format: openai
      pricing:
        currency: USD
        prompt_per_1m: 2.5
        cached_input_per_1m: 0.5
        completion_per_1m: 4.0
      external_model_ids:
        openai: llava-omni-preview
      backend_refs:
        - name: omni-primary
          endpoint: 127.0.0.1:8002
          protocol: http
          weight: 100
    - name: sdxl-image
      provider_model_id: sdxl-image
      api_format: openai
      pricing:
        currency: USD
        prompt_per_1m: 4.0
        completion_per_1m: 4.0
      external_model_ids:
        openai: gpt-image-1
      backend_refs:
        - name: image-primary
          endpoint: 127.0.0.1:8003
          protocol: http
          weight: 100

routing:
  modelCards:
    - name: qwen3-8b
      param_size: 8B
      context_window_size: 131072
      description: General-purpose text model for the default routing tier.
      capabilities: [chat, reasoning, tools]
      quality_score: 0.83
      modality: ar
      tags: [default, fast]
    - name: qwen3-32b
      param_size: 32B
      context_window_size: 131072
      description: Higher-quality text model for complex and verified requests.
      capabilities: [chat, reasoning, long-context]
      loras:
        - name: computer-science-expert
          description: Adapter specialized for advanced computer science and software design prompts.
      quality_score: 0.96
      modality: ar
      tags: [premium, analysis]
    - name: llava-omni
      param_size: 13B
      context_window_size: 65536
      description: Omni model that can satisfy both text and multimodal orchestration paths.
      capabilities: [chat, vision, multimodal]
      quality_score: 0.88
      modality: omni
      tags: [multimodal, omni]
    - name: sdxl-image
      param_size: 3B
      context_window_size: 8192
      description: Diffusion-style image generation backend used by image routes.
      capabilities: [image_generation]
      quality_score: 0.79
      modality: diffusion
      tags: [image, diffusion]

  signals:
    keywords:
      - name: code_keywords
        operator: OR
        method: bm25
        keywords: ["code", "function", "debug", "algorithm", "refactor"]
        case_sensitive: false
        bm25_threshold: 0.1
      - name: machine_learning
        operator: OR
        method: bm25
        keywords: ["machine learning", "model training", "gradient", "neural network", "classifier"]
        case_sensitive: false
        bm25_threshold: 0.1
      - name: urgent_keywords
        operator: OR
        method: ngram
        keywords: ["urgent", "immediate", "asap", "emergency"]
        case_sensitive: false
        ngram_threshold: 0.4
        ngram_arity: 3
      - name: fuzzy_sensitive_keywords
        operator: OR
        method: fuzzy
        keywords: ["social security", "credit card", "password", "api key"]
        case_sensitive: false
        fuzzy_match: true
        fuzzy_threshold: 82

    embeddings:
      - name: technical_support
        threshold: 0.75
        aggregation_method: max
        candidates:
          - how to configure the system
          - installation guide
          - troubleshooting steps
          - error message explanation
      - name: account_management
        threshold: 0.72
        aggregation_method: any
        candidates:
          - password reset
          - account settings
          - billing information
          - subscription management
      # query_modality declares which modality of incoming request payload the
      # embedding rule's query is computed from. Defaults to "text" when
      # omitted. Setting it to "image" or "audio" requires
      # global.model_catalog.embeddings.semantic.embedding_config.model_type=multimodal
      # so the query and candidate embeddings land in the same shared space;
      # see website/docs/tutorials/signal/learned/embedding.md for the worked
      # multimodal example.
      - name: clinical_query_intent
        query_modality: text
        threshold: 0.72
        aggregation_method: max
        candidates:
          - patient symptoms or clinical question
          - medication interaction or dosage question
          - test result interpretation request

    domains:
      - name: business
        description: Business and management related queries.
        mmlu_categories: [business]
        model_scores:
          - model: qwen3-32b
            score: 0.96
            use_reasoning: true
      - name: law
        description: Legal questions and policy topics.
        mmlu_categories: [law]
      - name: psychology
        description: Psychology and mental-health oriented prompts.
        mmlu_categories: [psychology]
      - name: health
        description: Health and medical information requests.
        mmlu_categories: [health]
      - name: "computer science"
        description: Computer science and engineering prompts.
        mmlu_categories: ["computer science"]
      - name: other
        description: General fallback traffic.
        mmlu_categories: [other]

    fact_check:
      - name: needs_fact_check
        description: Prompts with factual claims that should be verified.
      - name: no_fact_check_needed
        description: Creative prompts that do not need factual verification.

    user_feedbacks:
      - name: wrong_answer
        description: The user indicates the previous answer was incorrect.
      - name: need_clarification
        description: The user asks for a clearer or more detailed answer.

    reasks:
      - name: likely_dissatisfied
        description: The current user turn closely repeats the immediately previous user turn.
        threshold: 0.8
        lookback_turns: 1
      - name: persistently_dissatisfied
        description: The current user turn repeats the last two user turns in a row.
        threshold: 0.8
        lookback_turns: 2

    preferences:
      - name: terse_answers
        description: Users who prefer short, direct responses.
        examples:
          - keep it concise
          - bullet points only
          - answer in one paragraph
        threshold: 0.7

    language:
      - name: zh
        description: Chinese-language requests.
        # threshold: minimum lingua-go confidence required (0 = built-in default 0.3)
        threshold: 0.0
      - name: es
        description: Spanish-language requests.

    context:
      - name: long_context
        min_tokens: 32K
        max_tokens: 256K
        description: Requests that need larger effective context windows.

    structure:
      - name: many_questions
        description: Prompts that contain many explicit questions.
        feature:
          type: count
          source:
            type: regex
            pattern: '[?？]'
        predicate:
          gte: 4
      - name: at_most_one_question
        description: Prompts with one or fewer explicit questions.
        feature:
          type: count
          source:
            type: regex
            pattern: '[?？]'
        predicate:
          lte: 1
      - name: numbered_steps
        description: Prompts that contain numbered list items such as "1. ..."
        feature:
          type: exists
          source:
            type: regex
            pattern: '(?m)^\s*\d+\.\s+'
      - name: first_then_flow
        description: Prompts that express an ordered workflow.
        feature:
          type: sequence
          source:
            type: sequence
            case_sensitive: false
            sequences:
              - ["first", "then"]
              - ["first", "next", "finally"]
              - ["首先", "然后"]
              - ["先", "再"]
      - name: constraint_dense
        description: Prompts with dense explicit constraints relative to multilingual text units.
        feature:
          type: density
          source:
            type: keyword_set
            case_sensitive: false
            keywords:
              - under
              - at most
              - at least
              - within
              - no more than
              - 不超过
              - 至少
              - 最多
        predicate:
          gt: 0.08
      - name: format_directive_dense
        description: Prompts with dense output-format directives relative to multilingual text units.
        feature:
          type: density
          source:
            type: keyword_set
            keywords: ["table", "bullet", "json", "markdown", "表格", "列表", "JSON"]
        predicate:
          gt: 0.08
      - name: low_question_density
        description: Prompts with very low question density relative to multilingual text units.
        feature:
          type: density
          source:
            type: regex
            pattern: '[?？]'
        predicate:
          lt: 0.05

    complexity:
      - name: needs_reasoning
        threshold: 0.75
        description: Escalate multi-step reasoning or synthesis-heavy prompts.
        hard:
          candidates:
            - solve this step by step
            - compare multiple tradeoffs
            - analyze the root cause
          image_candidates:
            - generate a detailed architecture diagram
        easy:
          candidates:
            - answer briefly
            - quick summary
            - simple rewrite
          image_candidates:
            - make a simple icon
        composer:
          operator: OR
          conditions:
            - type: context
              name: long_context
            - type: domain
              name: "computer science"

    modality:
      - name: AR
        description: Text-only autoregressive requests.
      - name: DIFFUSION
        description: Image-generation requests.
      - name: BOTH
        description: Requests that need both text and image generation.

    role_bindings:
      - name: admin
        description: Requests from platform administrators.
        role: admin
        subjects:
          - kind: Group
            name: platform-admins
      - name: premium_user
        description: Requests from paid end users.
        role: premium_user
        subjects:
          - kind: Group
            name: premium-tier

    jailbreak:
      - name: prompt_injection
        method: hybrid
        threshold: 0.8
        include_history: true
        description: Detect common prompt-injection or jailbreak attempts.
        jailbreak_patterns:
          - ignore previous instructions
          - reveal the hidden prompt
          - jailbreak mode
        benign_patterns:
          - explain the policy
          - summarize the safety rules

    pii:
      - name: restricted_pii
        threshold: 0.85
        include_history: true
        pii_types_allowed:
          - EMAIL_ADDRESS
        description: Sensitive prompts where only low-risk identifiers may pass through.

    kb:
      - name: privacy_policy
        kb: privacy_kb
        target:
          kind: group
          value: privacy_policy
        match: best
      - name: security_containment
        kb: privacy_kb
        target:
          kind: group
          value: security_containment
        match: best
      - name: frontier_reasoning
        kb: privacy_kb
        target:
          kind: group
          value: frontier_reasoning
        match: best
      - name: local_standard
        kb: privacy_kb
        target:
          kind: group
          value: local_standard
        match: best

    conversation:
      - name: multi_turn_user
        description: At least two user messages.
        feature:
          type: count
          source:
            type: message
            role: user
        predicate:
          gte: 2
      - name: has_tools
        description: Request defines at least one tool.
        feature:
          type: count
          source:
            type: tool_definition
        predicate:
          gte: 1
      - name: active_tool_use
        description: Tool-result continuation inside an agent loop.
        feature:
          type: exists
          source:
            type: active_tool_loop
      - name: has_developer_msg
        description: Request contains a developer message.
        feature:
          type: exists
          source:
            type: message
            role: developer
      - name: heavy_non_user
        description: Non-user messages exceed threshold.
        feature:
          type: count
          source:
            type: message
            role: non_user
        predicate:
          gt: 5
          lte: 99

    events:
      - name: critical_payment_event
        description: Critical payment or transaction events that need incident-grade routing.
        event_types:
          - payment_failed
          - transaction_declined
        severities:
          - critical
          - high
        action_codes:
          - TXN_DECLINE
        temporal: true

  projections:
    partitions:
      - name: support_intents
        semantics: exclusive
        temperature: 0.3
        members:
          - technical_support
          - account_management
        default: technical_support
    scores:
      - name: request_difficulty
        method: weighted_sum
        inputs:
          - type: embedding
            name: technical_support
            weight: 0.18
            value_source: confidence
            match: 1.0
            miss: 0.0
          - type: embedding
            name: account_management
            weight: -0.08
            value_source: confidence
          - type: complexity
            name: needs_reasoning:medium
            weight: 0.18
          - type: complexity
            name: needs_reasoning:hard
            weight: 0.36
          - type: context
            name: long_context
            weight: 0.18
          - type: structure
            name: first_then_flow
            weight: 0.12
      - name: privacy_kb_bias
        method: weighted_sum
        inputs:
          - type: kb_metric
            kb: privacy_kb
            metric: private_vs_public
            weight: 1.0
            value_source: score
    mappings:
      - name: request_band
        source: request_difficulty
        method: threshold_bands
        calibration:
          method: sigmoid_distance
          slope: 10.0
        outputs:
          - name: support_fast
            lte: 0.20
          - name: support_balanced
            gt: 0.20
            lt: 0.45
          - name: support_escalated
            gte: 0.45

  decisions:
    - name: static_business_route
      description: Static fallback for standard business traffic.
      priority: 200
      tier: 2
      output_contract: Preserve any explicit user-requested output format exactly.
      rules:
        operator: AND
        conditions:
          - type: domain
            name: business
      modelRefs:
        - model: qwen3-8b
          use_reasoning: false
          reasoning_description: Prefer fast, direct business answers.
      candidateIterations:
        - variable: candidate
          source: models
          models:
            - model: qwen3-8b
              lora_name: business-adapter
              weight: 1.0
              quality_score: 0.83
              use_reasoning: false
              reasoning_description: Candidate iteration reference model for DSL authoring coverage.
              reasoning_effort: low
          outputs:
            - type: model
              value: candidate
      algorithm:
        type: static
      plugins:
        - type: system_prompt
          configuration:
            enabled: true
            mode: insert
            system_prompt: You are a senior business consultant. Give practical recommendations and clear tradeoffs.
        - type: semantic-cache
          configuration:
            enabled: true
            similarity_threshold: 0.82
            ttl_seconds: 7200
        - type: tools
          configuration:
            enabled: true
            mode: filtered
            semantic_selection: true
            allow_tools:
              - search_web
            block_tools:
              - exec_cmd
        # Disabled sample: add mode only (filter fields belong under mode: filter; see legal_confidence_route).
        - type: tool_selection
          configuration:
            enabled: false
            mode: add
            tools_db_path: config/tools_db.json
            top_k: 5
            similarity_threshold: 0.30
            strategy: default
            fallback_to_empty: false
            advanced_filtering:
              enabled: true
              candidate_pool_size: 30
              min_combined_score: 0.4
      emits:
        - kind: retention
          retention:
            drop: false
            ttl_turns: 4
            keep_current_model: true
            prefer_prefix_retention: true

    - name: legal_confidence_route
      description: Confidence-routed legal and factual traffic.
      priority: 190
      rules:
        operator: OR
        conditions:
          - type: domain
            name: law
          - type: fact_check
            name: needs_fact_check
      modelRefs:
        - model: qwen3-8b
          use_reasoning: true
          reasoning_effort: medium
          weight: 0.4
        - model: qwen3-32b
          use_reasoning: true
          reasoning_description: Escalate when confidence is weak.
          reasoning_effort: high
          weight: 0.6
      algorithm:
        type: confidence
        on_error: fallback
        confidence:
          confidence_method: hybrid
          threshold: 0.72
          hybrid_weights:
            logprob_weight: 0.65
            margin_weight: 0.35
          on_error: skip
          escalation_order: small_to_large
          cost_quality_tradeoff: 0.45
          token_filter: alpha_numeric
          verifier_server_url: ""
          verifier_timeout_seconds: 0
      plugins:
        - type: hallucination
          configuration:
            enabled: true
            use_nli: true
            hallucination_action: header
            unverified_factual_action: header
            include_hallucination_details: true
        # Disabled sample: filter mode only (reference config union coverage for ToolSelectionPluginConfig filter fields).
        - type: tool_selection
          configuration:
            enabled: false
            mode: filter
            relevance_threshold: 0.25
            preserve_count: 2

    - name: health_ratings_route
      description: Ratings-based route for health requests that need fact verification.
      priority: 180
      rules:
        operator: AND
        conditions:
          - type: domain
            name: health
          - type: fact_check
            name: needs_fact_check
      modelRefs:
        - model: qwen3-8b
          use_reasoning: true
          reasoning_effort: medium
        - model: qwen3-32b
          use_reasoning: true
          reasoning_effort: high
      algorithm:
        type: ratings
        ratings:
          max_concurrent: 3
          on_error: skip
      adaptations:
        mode: bypass
      plugins:
        - type: router_replay
          configuration:
            enabled: true
            max_records: 10000
            capture_request_body: true
            capture_response_body: true
            max_body_bytes: 4096
            # 0 = no truncation (the documented default). Set to a positive
            # byte cap to independently limit structured tool-trace fields
            # (prompt, tool_definitions, and each tool-call's arguments/output)
            # without having to shrink max_body_bytes.
            max_tool_trace_bytes: 0
            # Cap the number of tool-trace steps retained per record. Long
            # agent sessions (many sequential tool calls) can otherwise grow
            # the trace unbounded and OOM the router (#1835). 0 = no cap.
            max_tool_trace_steps: 100

    - name: computer-science-remom-route
      description: ReMoM route for complex computer science prompts.
      priority: 170
      output_contract_spec:
        type: reference_selection
        reference:
          source: candidate_responses
          id_format: index
        extract:
          mode: exact
          sources: [content]
        postprocess:
          - type: dereference_selected_reference
      rules:
        operator: AND
        conditions:
          - type: domain
            name: "computer science"
          - type: complexity
            name: needs_reasoning:hard
      modelRefs:
        - model: qwen3-32b
          use_reasoning: true
          reasoning_description: Use the strongest text model for synthesis.
          reasoning_effort: high
          lora_name: computer-science-expert
          weight: 0.5
        - model: qwen3-8b
          use_reasoning: true
          reasoning_effort: medium
          weight: 0.2
        - model: llava-omni
          use_reasoning: true
          reasoning_effort: high
          weight: 0.3
      algorithm:
        type: remom
        remom:
          breadth_schedule: [3, 2]
          model_distribution: round_robin
          temperature: 0.7
          include_reasoning: true
          compaction_strategy: last_n_tokens
          compaction_tokens: 512
          synthesis_template: Fuse the strongest technical findings into a final answer.
          synthesis_model: qwen3-32b
          max_concurrent: 3
          round_timeout_seconds: 120
          min_successful_responses: 2
          on_error: skip
          shuffle_seed: 7
          include_intermediate_responses: true
          max_responses_per_round: 2
      plugins:
        - type: memory
          configuration:
            enabled: true
            retrieval_limit: 6
            similarity_threshold: 0.72
            auto_store: true
            hybrid_search: true
            hybrid_mode: rerank
            reflection:
              enabled: true
              algorithm: recency_semantic
              max_inject_tokens: 384
              recency_decay_days: 14
              dedup_threshold: 0.9
              block_patterns:
                - api key
                - password

    - name: deliberation-fusion-route
      description: Fusion route for prompts that benefit from panel deliberation.
      priority: 165
      rules:
        operator: AND
        conditions:
          - type: domain
            name: business
          - type: complexity
            name: needs_reasoning:hard
      modelRefs:
        - model: qwen3-8b
          use_reasoning: true
          reasoning_effort: medium
          weight: 0.5
        - model: qwen3-32b
          use_reasoning: true
          reasoning_effort: high
          weight: 0.5
      algorithm:
        type: fusion
        fusion:
          model: qwen3-32b
          analysis_models:
            - qwen3-8b
            - qwen3-32b
          analysis_overrides:
            - model: qwen3-8b
              temperature: 0.2
              max_completion_tokens: 384
            - model: qwen3-32b
              temperature: 0.15
              max_completion_tokens: 512
          max_concurrent: 2
          max_completion_tokens: 512
          round_timeout_seconds: 90
          min_successful_responses: 1
          temperature: 0.2
          include_analysis: true
          include_intermediate_responses: true
          on_error: skip
          analysis_template: ""
          synthesis_template: ""
          judge_prompt_version: fusion-v1
          # Grounding-aware synthesis: score panel responses for faithfulness
          # (hallucination detector against context, else cross-model NLI) and
          # use the scores to guide the judge. Disabled by default.
          grounding:
            enabled: false
            reference: hybrid          # hybrid | context | panel
            policy: weight             # weight | annotate | filter (see below)
            min_score: 0.0             # filter policy only: drop below this (0-1)
            min_keep: 1                # filter policy only: keep at least this many
            nli_contradiction_penalty: 1.0
            on_error: skip             # skip (fall back to plain fusion) | fail
            # policy controls how the scores are used:
            #   weight   (default) keep every response; tell the judge to weight
            #            each answer by its score (a correct lone dissenter is kept).
            #   annotate keep every response; pass the scores to the judge as notes.
            #   filter   hard-drop responses below min_score (keep min_keep).
            # filter regresses quality on contested factual questions because it
            # deletes the correct dissenter; see bench/grounded_fusion/FINDINGS.md.

    - name: router-flow-workflow-route
      description: Router Flow route for bounded static or dynamic micro-agent workflows.
      priority: 164
      rules:
        operator: AND
        conditions:
          - type: domain
            name: business
          - type: complexity
            name: needs_reasoning:hard
      modelRefs:
        - model: qwen3-8b
          use_reasoning: true
          reasoning_effort: medium
          weight: 0.5
        - model: qwen3-32b
          use_reasoning: true
          reasoning_effort: high
          weight: 0.5
      algorithm:
        type: workflows
        workflows:
          mode: dynamic
          template: micro_agent
          planner:
            model: qwen3-32b
          max_steps: 6
          max_parallel: 2
          max_completion_tokens: 1024
          round_timeout_seconds: 90
          min_successful_responses: 1
          temperature: 0.2
          include_intermediate_responses: true
          on_error: fail

    - name: terse_feedback_route
      description: Terse-answer route that allows Router Learning adaptation to adjust feedback-driven routing.
      priority: 160
      rules:
        operator: AND
        conditions:
          - type: preference
            name: terse_answers
      modelRefs:
        - model: qwen3-8b
          use_reasoning: false
        - model: qwen3-32b
          use_reasoning: false
      adaptations:
        adaptation:
          mode: apply
      plugins:
        - type: fast_response
          configuration:
            message: The primary model is saturated, so a lightweight response was returned immediately.
        - type: request_params
          configuration:
            blocked_params:
              - logprobs
              - top_logprobs
            max_tokens_limit: 8192
            max_n: 4
            strip_unknown: false

    - name: support_router_dc_route
      description: Embedding-driven support route that escalates only when the derived projection band says the request is not a simple support query.
      priority: 150
      rules:
        operator: AND
        conditions:
          - type: embedding
            name: technical_support
          - type: projection
            name: support_escalated
      modelRefs:
        - model: qwen3-8b
          use_reasoning: false
        - model: qwen3-32b
          use_reasoning: true
      algorithm:
        type: router_dc
        router_dc:
          temperature: 0.2
          dimension_size: 384
          min_similarity: 0.7
          use_query_contrastive: true
          use_model_contrastive: true
          require_descriptions: true
          use_capabilities: true
      plugins:
        - type: header_mutation
          configuration:
            add:
              - name: X-Tenant-Tier
                value: premium
            update:
              - name: X-Route-Source
                value: semantic-router
            delete:
              - X-Debug-Trace

    - name: urgent_automix_route
      description: Cost-aware urgent handling with AutoMix.
      priority: 145
      rules:
        operator: AND
        conditions:
          - type: keyword
            name: urgent_keywords
      modelRefs:
        - model: qwen3-8b
          use_reasoning: false
        - model: qwen3-32b
          use_reasoning: true
      algorithm:
        type: automix
        automix:
          verification_threshold: 0.78
          max_escalations: 2
          cost_aware_routing: true
          cost_quality_tradeoff: 0.35
          discount_factor: 0.9
          use_logprob_verification: true
      plugins:
        - type: response_jailbreak
          configuration:
            enabled: true
            threshold: 0.85
            action: header

    - name: safe_hybrid_route
      description: Composite route combining AND, OR, and NOT with Hybrid selection.
      priority: 140
      rules:
        operator: AND
        conditions:
          - type: domain
            name: business
          - operator: OR
            conditions:
              - type: keyword
                name: urgent_keywords
              - type: complexity
                name: needs_reasoning:hard
          - operator: NOT
            conditions:
              - type: jailbreak
                name: prompt_injection
      modelRefs:
        - model: qwen3-8b
          use_reasoning: true
        - model: qwen3-32b
          use_reasoning: true
      algorithm:
        type: hybrid
        hybrid:
          experience_weight: 0.4
          router_dc_weight: 0.3
          automix_weight: 0.2
          cost_weight: 0.1
          quality_gap_threshold: 0.08
          normalize_scores: true
    - name: spanish_adaptive_route
      description: Spanish-language route that observes Router Learning adaptation with Milvus RAG.
      priority: 135
      rules:
        operator: AND
        conditions:
          - type: language
            name: es
      modelRefs:
        - model: qwen3-8b
          use_reasoning: false
        - model: qwen3-32b
          use_reasoning: true
      adaptations:
        adaptation:
          mode: observe
        protection:
          stability_weight: 1.2
          switch_margin: 0.04
      plugins:
        - type: rag
          configuration:
            enabled: true
            backend: milvus
            similarity_threshold: 0.78
            top_k: 5
            max_context_length: 12000
            injection_mode: tool_role
            on_failure: warn
            cache_results: true
            cache_ttl_seconds: 600
            min_confidence_threshold: 0.55
            backend_config:
              collection: docs
              reuse_cache_connection: true
              content_field: content
              metadata_field: metadata
              filter_expression: locale == "es"

    - name: admin_adaptive_route
      description: Admin route with Router Learning adaptation and external API RAG.
      priority: 130
      rules:
        operator: AND
        conditions:
          - type: authz
            name: admin
      modelRefs:
        - model: qwen3-32b
          use_reasoning: true
      adaptations:
        adaptation:
          mode: apply
      plugins:
        - type: rag
          configuration:
            enabled: true
            backend: external_api
            similarity_threshold: 0.7
            top_k: 6
            max_context_length: 8000
            injection_mode: system_prompt
            on_failure: block
            cache_results: false
            min_confidence_threshold: 0.5
            backend_config:
              endpoint: https://search.example.com/query
              api_key: search-api-key
              auth_header: X-API-Key
              request_format: custom
              request_template: "{\"query\":\"{{.Query}}\",\"top_k\":{{.TopK}},\"threshold\":{{.Threshold}}}"
              timeout_seconds: 15
              headers:
                X-Environment: admin

    - name: long_context_latency_route
      description: Long-context route with latency-aware policy and MCP-backed RAG.
      priority: 125
      rules:
        operator: AND
        conditions:
          - type: context
            name: long_context
      modelRefs:
        - model: qwen3-32b
          use_reasoning: true
      algorithm:
        type: latency_aware
        latency_aware:
          tpot_percentile: 90
          ttft_percentile: 95
          description: Balance throughput and first-token latency for long prompts.
      plugins:
        - type: rag
          configuration:
            enabled: true
            backend: mcp
            similarity_threshold: 0.74
            top_k: 4
            max_context_length: 10000
            injection_mode: tool_role
            on_failure: skip
            cache_results: false
            min_confidence_threshold: 0.5
            backend_config:
              server_name: docs-server
              tool_name: retrieve_docs
              tool_arguments:
                query: ${user_content}
                top_k: ${top_k}
              timeout_seconds: 20

    # multi_factor example: composes quality/latency/cost/load into one
    # weighted score with optional SLO ceilings. Issue #37.
    - name: multi_factor_route
      description: Compose quality, latency, cost, and load signals into a single weighted score.
      priority: 124
      rules:
        operator: AND
        conditions:
          - type: domain
            name: other
      modelRefs:
        - model: qwen3-8b
          quality_score: 0.72
        - model: qwen3-32b
          quality_score: 0.94
      algorithm:
        type: multi_factor
        multi_factor:
          weights:
            quality: 0.4
            latency: 0.2
            cost: 0.2
            load: 0.2
          slo:
            max_tpot_ms: 200
            max_ttft_ms: 800
            max_cost_per_1m: 5.0
            max_inflight: 50
          latency_percentile: 95
          on_no_candidates: cheapest

    - name: fallback_knn_route
      description: KNN fallback route for general traffic with OpenAI vector-store RAG.
      priority: 120
      rules:
        operator: AND
        conditions:
          - type: domain
            name: other
      modelRefs:
        - model: qwen3-8b
          use_reasoning: false
      algorithm:
        type: knn
      plugins:
        - type: rag
          configuration:
            enabled: true
            backend: openai
            similarity_threshold: 0.73
            top_k: 8
            max_context_length: 9000
            injection_mode: tool_role
            on_failure: warn
            cache_results: true
            cache_ttl_seconds: 300
            min_confidence_threshold: 0.35
            backend_config:
              vector_store_id: vs_docs_prod
              base_url: https://api.openai.com
              api_key: openai-file-search-key
              max_num_results: 12
              file_ids: [file_a, file_b]
              filter:
                doc_type: manual
              timeout_seconds: 20
              workflow_mode: tool_based

    - name: clarification_kmeans_route
      description: KMeans route for clarification requests with local vector-store RAG.
      priority: 115
      rules:
        operator: AND
        conditions:
          - type: user_feedback
            name: need_clarification
      modelRefs:
        - model: qwen3-8b
          use_reasoning: false
        - model: qwen3-32b
          use_reasoning: true
      algorithm:
        type: kmeans
      plugins:
        - type: rag
          configuration:
            enabled: true
            backend: vectorstore
            similarity_threshold: 0.69
            top_k: 5
            max_context_length: 7000
            injection_mode: tool_role
            on_failure: skip
            cache_results: true
            cache_ttl_seconds: 900
            min_confidence_threshold: 0.25
            backend_config:
              vector_store_id: local-vs
              file_ids: [kb-001, kb-002]

    - name: repeat_question_escalation
      description: Escalate when the user repeats the same question after an unsatisfying answer.
      priority: 114
      rules:
        operator: AND
        conditions:
          - type: reask
            name: likely_dissatisfied
      modelRefs:
        - model: qwen3-32b
          use_reasoning: true
      algorithm:
        type: static
      plugins:
        - type: header_mutation
          configuration:
            add:
              - name: x-escalation-reason
                value: repeated-question

    - name: safe_only_svm_route
      description: SVM route that only matches when prompt-injection signals are absent.
      priority: 110
      rules:
        operator: NOT
        conditions:
          - type: jailbreak
            name: prompt_injection
      modelRefs:
        - model: qwen3-8b
          use_reasoning: false
      algorithm:
        type: svm
      plugins:
        - type: rag
          configuration:
            enabled: true
            backend: hybrid
            similarity_threshold: 0.76
            top_k: 6
            max_context_length: 9000
            injection_mode: tool_role
            on_failure: warn
            cache_results: false
            min_confidence_threshold: 0.45
            backend_config:
              primary: milvus
              fallback: external_api
              strategy: parallel
              primary_config:
                collection: docs
                reuse_cache_connection: true
              fallback_config:
                endpoint: https://search.example.com/query
                request_format: pinecone

    - name: mlp_experimental_route
      description: Experimental MLP-based model selection route.
      priority: 103
      rules:
        operator: AND
        conditions:
          - type: keyword
            name: machine_learning
      modelRefs:
        - model: qwen3-8b
          use_reasoning: false
      algorithm:
        type: mlp

    - name: image_generation_route
      description: Dedicated image-generation route using the OpenAI image backend.
      priority: 105
      rules:
        operator: AND
        conditions:
          - type: modality
            name: DIFFUSION
      modelRefs:
        - model: sdxl-image
          use_reasoning: false
      plugins:
        - type: image_gen
          configuration:
            enabled: true
            backend: openai
            backend_config:
              api_key: openai-image-key
              base_url: https://api.openai.com/v1
              model: gpt-image-1
              quality: hd
              style: vivid
            modality_detection:
              method: hybrid
              classifier:
                model_path: models/mmbert32k-modality-router-merged
                use_cpu: true
              keywords:
                - generate an image
                - draw
                - create a picture
              both_keywords:
                - explain and illustrate
                - describe with an image
              confidence_threshold: 0.7
              lower_threshold_ratio: 0.7
            default_width: 1024
            default_height: 1024
            max_inference_steps: 30
            timeout_seconds: 60

    - name: multimodal_route
      description: BOTH-modality route requiring text and image-capable execution.
      priority: 100
      rules:
        operator: AND
        conditions:
          - type: modality
            name: BOTH
      modelRefs:
        - model: qwen3-8b
          use_reasoning: true
        - model: sdxl-image
          use_reasoning: false
      plugins:
        - type: image_gen
          configuration:
            enabled: true
            backend: vllm_omni
            backend_config:
              base_url: http://image-router:8005
              model: Qwen/Qwen-Image
              num_inference_steps: 28
              cfg_scale: 4.5
              seed: 42
            default_width: 1024
            default_height: 1024
            max_inference_steps: 28
            timeout_seconds: 90

entrypoints:
  - model_names: ["vllm-sr/exhaustive-default"]
    recipe: default
  - model_names: ["vllm-sr/exhaustive-privacy"]
    recipe: privacy-first

recipes:
  - name: privacy-first
    description: Keeps privacy-sensitive traffic on the local default-tier model.
    routing:
      signals:
        keywords:
          - name: privacy_recipe_keywords
            operator: OR
            method: bm25
            keywords: ["ssn", "passport number", "credit card number"]
            case_sensitive: false
            bm25_threshold: 0.1
      decisions:
        - name: privacy_recipe_route
          description: Route privacy-sensitive prompts to the local model.
          priority: 100
          rules:
            operator: AND
            conditions:
              - type: keyword
                name: privacy_recipe_keywords
          modelRefs:
            - model: qwen3-8b
              use_reasoning: false
          algorithm:
            type: static

global:
  router:
    config_source: file
    strategy: priority
    auto_model_name: exhaustive-reference
    auto_model_names:
      - vllm-sr/auto
      - auto
      - exhaustive-reference
    include_config_models_in_list: true
    clear_route_cache: false
    streamed_body:
      enabled: true
      max_bytes: 1048576
      timeout_sec: 15
    skip_processing:
      enabled: false
    model_selection:
      method: hybrid
      enabled: true
      router_dc:
        temperature: 0.2
        dimension_size: 384
        min_similarity: 0.7
        use_query_contrastive: true
        use_model_contrastive: true
        require_descriptions: true
        use_capabilities: true
      automix:
        verification_threshold: 0.8
        max_escalations: 2
        cost_aware_routing: true
        cost_quality_tradeoff: 0.4
        discount_factor: 0.9
        use_logprob_verification: true
      hybrid:
        experience_weight: 0.4
        router_dc_weight: 0.25
        automix_weight: 0.25
        cost_weight: 0.1
        quality_gap_threshold: 0.08
        normalize_scores: true
      momentum:
        enabled: false
        attack: 0.7
        release: 0.3
        threshold: 0.6
      ml:
        models_path: models/model-selection
        embedding_dim: 384
        knn:
          k: 5
          pretrained_path: models/model-selection/knn.bin
        kmeans:
          num_clusters: 8
          efficiency_weight: 0.35
          pretrained_path: models/model-selection/kmeans.bin
        svm:
          kernel: rbf
          gamma: 0.12
          pretrained_path: models/model-selection/svm.bin
        mlp:
          device: cpu
          pretrained_path: models/model-selection/mlp.pt
    learning:
      enabled: true
      adaptation:
        enabled: true
        candidate_set: decision
        strategy: routing_sampling
      protection:
        enabled: true
        scope: conversation
        identity:
          headers:
            session: x-session-id
            conversation: x-conversation-id
        tuning:
          idle_timeout_seconds: 300
          min_turns_before_switch: 1
          switch_margin: 0.05
          stability_weight: 1.0

  services:
    api:
      batch_classification:
        max_batch_size: 100
        concurrency_threshold: 5
        max_concurrency: 8
        metrics:
          sample_rate: 1.0
          batch_size_ranges:
            - min: 1
              max: 4
              label: tiny
            - min: 5
              max: 16
              label: standard
          duration_buckets: [0.01, 0.05, 0.1, 0.5, 1.0]
          size_buckets: [128, 512, 2048, 8192]
          enabled: true
          detailed_goroutine_tracking: true
          high_resolution_timing: true
    response_api:
      enabled: true
      store_backend: redis
      ttl_seconds: 86400
      max_responses: 2000
      redis:
        address: redis:6379
        password: redis-secret
        db: 1
        key_prefix: "vsr:responses:"
        cluster_mode: false
        cluster_addresses: [redis-a:6379, redis-b:6379]
        pool_size: 10
        min_idle_conns: 2
        max_retries: 3
        dial_timeout: 5
        read_timeout: 5
        write_timeout: 5
        tls_enabled: false
        tls_cert_path: certs/redis.crt
        tls_key_path: certs/redis.key
        tls_ca_path: certs/ca.crt
        config_path: config/backends/response-api-redis.yaml
    observability:
      tracing:
        enabled: true
        provider: opentelemetry
        exporter:
          type: otlp
          endpoint: jaeger:4317
          insecure: true
        sampling:
          type: probabilistic
          rate: 0.5
        resource:
          service_name: vllm-sr
          service_version: v0.3.0
          deployment_environment: development
      metrics:
        enabled: true
        windowed_metrics:
          enabled: true
          time_windows: [1m, 5m, 1h]
          update_interval: 30s
          model_metrics: true
          queue_depth_estimation: true
          max_models: 50
    authz:
      fail_open: false
      identity:
        user_id_header: x-user-id
        user_groups_header: x-user-groups
      providers:
        - type: header
          headers:
            X-Plan: premium
            X-Role: admin
    ratelimit:
      fail_open: false
      providers:
        - type: redis
          address: redis:6379
          domain: api
          rules:
            - name: premium-per-minute
              match:
                group: premium-tier
                model: qwen3-32b
              requests_per_unit: 120
              tokens_per_unit: 300000
              unit: minute
            - name: user-per-hour
              match:
                user: demo-user
              requests_per_unit: 1000
              tokens_per_unit: 2000000
              unit: hour
    management_api:
      bind_address: 127.0.0.1
      port: 8080
      remote_exposure: false
      auth:
        mode: disabled
        tokens:
          - env: VSR_MGMT_TOKEN
            role: admin
        roles:
          viewer:
            - health.read
            - ready.read
            - docs.read
            - metrics.read
            - classify.invoke
            - config.read
            - data.read
          operator:
            - health.read
            - ready.read
            - docs.read
            - metrics.read
            - classify.invoke
            - config.read
            - config.write
            - learning.ingest
            - data.read
            - data.write
          admin:
            - "*"
    startup_status:
      store_backend: file
    router_replay:
      enabled: true
      store_backend: postgres
      ttl_seconds: 2592000
      async_writes: true
      redis:
        address: redis:6379
        db: 2
        password: redis-secret
        use_tls: false
        tls_skip_verify: false
        max_retries: 3
        pool_size: 20
        key_prefix: "vsr:replay:"
      postgres:
        host: postgres
        port: 5432
        database: vsr
        user: router
        password: router-secret
        ssl_mode: disable
        max_open_conns: 20
        max_idle_conns: 10
        conn_max_lifetime: 300
        table_name: router_replay
      milvus:
        address: milvus:19530
        username: root
        password: milvus-secret
        collection_name: router_replay
        consistency_level: Bounded
        shard_num: 2
      qdrant:
        host: qdrant
        port: 6334
        api_key: ""
        use_tls: false
        collection_name: router_replay

  stores:
    semantic_cache:
      backend_type: milvus
      enabled: true
      similarity_threshold: 0.86
      max_entries: 50000
      ttl_seconds: 7200
      eviction_policy: lru
      embedding_model: mmbert
      redis:
        connection:
          host: redis
          port: 6379
          database: 0
          password: redis-secret
          timeout: 5
          tls:
            enabled: false
            cert_file: certs/redis.crt
            key_file: certs/redis.key
            ca_file: certs/ca.crt
        index:
          name: semantic_cache_idx
          prefix: "vsr:cache:"
          vector_field:
            name: embedding
            dimension: 768
            metric_type: COSINE
          index_type: HNSW
          params:
            M: 16
            efConstruction: 200
        search:
          topk: 10
        development:
          drop_index_on_startup: false
          auto_create_index: true
          verbose_errors: true
        logging:
          level: info
          enable_query_log: true
          enable_metrics: true
      milvus:
        connection:
          host: milvus
          port: 19530
          database: default
          timeout: 30
          auth:
            enabled: true
            username: root
            password: milvus-secret
          tls:
            enabled: false
            cert_file: certs/milvus.crt
            key_file: certs/milvus.key
            ca_file: certs/ca.crt
        collection:
          name: semantic_cache
          description: Semantic cache vectors.
          vector_field:
            name: embedding
            dimension: 768
            metric_type: COSINE
          index:
            type: HNSW
            params:
              M: 16
              efConstruction: 200
        search:
          params:
            ef: 64
          topk: 10
          consistency_level: Bounded
        performance:
          connection_pool:
            max_connections: 20
            max_idle_connections: 10
            acquire_timeout: 10
          batch:
            insert_batch_size: 128
            timeout: 30
        data_management:
          ttl:
            enabled: true
            timestamp_field: created_at
            cleanup_interval: 3600
          compaction:
            enabled: true
            interval: 86400
        logging:
          level: info
          enable_query_log: true
          enable_metrics: true
        development:
          drop_collection_on_startup: false
          auto_create_collection: true
          verbose_errors: true
      valkey:
        connection:
          host: valkey
          port: 6379
          database: 0
          password: valkey-secret
          timeout: 5
          tls:
            enabled: false
            cert_file: certs/valkey.crt
            key_file: certs/valkey.key
            ca_file: certs/ca.crt
        index:
          name: semantic_cache_idx
          prefix: "vsr:cache:"
          vector_field:
            name: embedding
            dimension: 768
            metric_type: COSINE
          index_type: HNSW
          params:
            M: 16
            efConstruction: 200
        search:
          topk: 10
        development:
          drop_index_on_startup: false
          auto_create_index: true
          verbose_errors: true
        logging:
          level: info
          enable_query_log: true
          enable_metrics: true
      qdrant:
        host: qdrant
        port: 6334
        api_key: ""
        use_tls: false
        connect_timeout: 10
        collection_name: semantic_cache
    memory:
      enabled: true
      backend: milvus
      auto_store: true
      disabled_routes: []
      disabled_models: []
      valkey:
        host: valkey
        port: 6379
        database: 0
        password: valkey-secret
        timeout: 10
        collection_prefix: "mem:"
        index_name: mem_idx
        dimension: 384
        metric_type: COSINE
        index_m: 16
        index_ef_construction: 200
        tls_enabled: false
        tls_ca_path: ""
        tls_insecure_skip_verify: false
      milvus:
        address: milvus:19530
        collection: agentic_memory
        dimension: 384
        num_partitions: 32
      qdrant:
        host: qdrant
        port: 6334
        api_key: ""
        use_tls: false
        connect_timeout: 10
        collection: agentic_memory
        dimension: 384
      redis_cache:
        enabled: false
        address: "localhost:6379"
        ttl_seconds: 300
        db: 0
        key_prefix: "memory_cache:"
        password: ""
      embedding_model: bert
      extraction_batch_size: 16
      default_retrieval_limit: 10
      default_similarity_threshold: 0.74
      hybrid_search: true
      hybrid_mode: rerank
      adaptive_threshold: true
      quality_scoring:
        initial_strength_days: 30
        prune_threshold: 0.15
        max_memories_per_user: 200
      reflection:
        enabled: true
        algorithm: recency_semantic
        max_inject_tokens: 512
        recency_decay_days: 14
        dedup_threshold: 0.9
        block_patterns:
          - api key
          - password
    vector_store:
      enabled: true
      backend_type: llama_stack
      file_storage_dir: /var/lib/vsr/data
      max_file_size_mb: 100
      embedding_model: multimodal
      embedding_dimension: 384
      ingestion_workers: 4
      ingestion_drain_timeout_seconds: 30
      supported_formats: [.txt, .md, .json, .csv, .html, .pdf]
      milvus:
        connection:
          host: milvus
          port: 19530
          database: vector_store
          timeout: 30
          auth:
            enabled: true
            username: root
            password: milvus-secret
          tls:
            enabled: false
            cert_file: certs/milvus.crt
            key_file: certs/milvus.key
            ca_file: certs/ca.crt
        collection:
          name: vector_store
          description: Uploaded document embeddings.
          vector_field:
            name: embedding
            dimension: 384
            metric_type: COSINE
          index:
            type: HNSW
            params:
              M: 16
              efConstruction: 200
        search:
          params:
            ef: 64
          topk: 10
          consistency_level: Strong
        performance:
          connection_pool:
            max_connections: 20
            max_idle_connections: 10
            acquire_timeout: 10
          batch:
            insert_batch_size: 64
            timeout: 30
        data_management:
          ttl:
            enabled: true
            timestamp_field: created_at
            cleanup_interval: 3600
          compaction:
            enabled: true
            interval: 86400
        logging:
          level: info
          enable_query_log: true
          enable_metrics: true
        development:
          drop_collection_on_startup: false
          auto_create_collection: true
          verbose_errors: true
      memory:
        max_entries_per_store: 5000
      llama_stack:
        endpoint: http://llama-stack:8321
        auth_token: llama-stack-token
        embedding_model: all-MiniLM-L6-v2
        request_timeout_seconds: 30
        search_type: hybrid
      valkey:
        host: valkey
        port: 6379
        database: 0
        password: valkey-secret
        connect_timeout: 5
        collection_prefix: "vsr_vs_"
        metric_type: COSINE
        index_m: 16
        index_ef_construction: 200
      qdrant:
        host: qdrant
        port: 6334
        api_key: ""
        use_tls: false
        connect_timeout: 10
        collection_prefix: "vsr_vs_"
      metadata_store: memory
      metadata_postgres:
        host: postgres
        port: 5432
        database: vsr
        user: router
        password: router-secret
        ssl_mode: disable
        max_open_conns: 25
        max_idle_conns: 5
        conn_max_lifetime: 300
        table_name: vector_store_registry

  integrations:
    tools:
      enabled: true
      top_k: 5
      similarity_threshold: 0.35
      tools_db_path: deploy/examples/runtime/tools/tools_db.json
      fallback_to_empty: true
      advanced_filtering:
        enabled: true
        retrieval_strategy: weighted
        candidate_pool_size: 50
        min_lexical_overlap: 1
        min_combined_score: 0.42
        weights:
          embed: 0.5
          lexical: 0.2
          tag: 0.1
          name: 0.1
          category: 0.1
        use_category_filter: true
        category_confidence_threshold: 0.7
        allow_tools: [docs.search, tickets.lookup]
        block_tools: [admin.delete]
        hybrid_history:
          history_horizon: 8
          min_history_steps: 1
          history_confidence_threshold: 0
          weight_semantic: 1
          weight_history_transition: 1
          weight_decision_prior: 1
          repetition_penalty_strength: 0
    looper:
      endpoint: http://localhost:8899/v1/chat/completions
      grpc_max_msg_size_mb: 16
      max_response_bytes_mb: 32
      timeout_seconds: 120
      retry_count: 2
      headers:
        X-Trace-Source: looper
      remom:
        model_names:
          - vllm-sr/remom
      fusion:
        model_names:
          - vllm-sr/fusion
      flow:
        model_names:
          - vllm-sr/flow
        state:
          store_backend: file
          ttl_seconds: 1800
          file:
            directory: .vllm-sr/flow-state
          redis:
            address: localhost:6379
            db: 0
            password: ""
            use_tls: false
            tls_skip_verify: false
            max_retries: 3
            pool_size: 10
            key_prefix: "vllm-sr:flow:state:"

  model_catalog:
    embeddings:
      semantic:
        qwen3_model_path: models/mom-embedding-pro
        gemma_model_path: models/mom-embedding-flash
        mmbert_model_path: models/mmbert-embed-32k-2d-matryoshka
        multimodal_model_path: models/mom-embedding-multimodal
        bert_model_path: models/mom-embedding-light
        use_cpu: true
        embedding_config:
          backend: candle
          model_type: mmbert
          preload_embeddings: true
          target_dimension: 768
          target_layer: 22
          enable_soft_matching: true
          top_k: 1
          min_score_threshold: 0.45
          prototype_scoring:
            enabled: true
            cluster_similarity_threshold: 0.9
            max_prototypes: 8
            best_weight: 0.75
            top_m: 2
            margin_threshold: 0.0
        endpoint:
          base_url: http://embedding-service:8000/v1
          model: BAAI/bge-m3
          api_key_env: EMBEDDING_API_KEY
          timeout_seconds: 5
          max_retries: 2
          dimensions: 1024
    system:
      prompt_guard: models/mmbert32k-jailbreak-detector-merged
      domain_classifier: models/mmbert32k-intent-classifier-merged
      pii_classifier: models/mmbert32k-pii-detector-merged
      fact_check_classifier: models/mmbert32k-factcheck-classifier-merged
      hallucination_detector: models/mom-halugate-detector
      hallucination_explainer: models/mom-halugate-explainer
      feedback_detector: models/mmbert32k-feedback-detector-merged
    external:
      - llm_provider: openai
        model_role: guardrail
        llm_endpoint:
          address: openai-proxy
          port: 443
          protocol: https
          name: openai-guardrail
          use_chat_template: true
          prompt_template: "{{ .Prompt }}"
        llm_model_name: gpt-4.1-mini
        llm_timeout_seconds: 30
        parser_type: json
        threshold: 0.65
        access_key: external-guardrail-key
        max_tokens: 2048
        temperature: 0.2
    kbs:
      - name: privacy_kb
        source:
          path: knowledge_bases/privacy/
          manifest: labels.json
        threshold: 0.55
        prototype_scoring:
          enabled: true
          cluster_similarity_threshold: 0.9
          max_prototypes: 8
          best_weight: 0.75
          top_m: 2
          margin_threshold: 0.0
        label_thresholds:
          prompt_injection: 0.7
        groups:
          security_containment:
            - prompt_injection
            - credential_exfiltration
            - jailbreak_role
            - system_prompt_extraction
          privacy_policy:
            - proprietary_code
            - internal_document
            - pii
            - business_strategy
            - trade_secret_ip
            - operational_infrastructure
            - locality_directive
            - customer_data
          frontier_reasoning:
            - architecture_analysis
            - root_cause_analysis
            - multi_step_tradeoffs
          local_standard:
            - generic_coding
            - simple_task
            - general_knowledge
          private:
            - prompt_injection
            - credential_exfiltration
            - jailbreak_role
            - system_prompt_extraction
            - proprietary_code
            - internal_document
            - pii
            - business_strategy
            - trade_secret_ip
            - operational_infrastructure
            - locality_directive
            - customer_data
          public:
            - architecture_analysis
            - root_cause_analysis
            - multi_step_tradeoffs
            - generic_coding
            - simple_task
            - general_knowledge
        metrics:
          - name: private_vs_public
            type: group_margin
            positive_group: private
            negative_group: public
      - name: mmlu_kb
        source:
          path: knowledge_bases/mmlu/
          manifest: labels.json
        threshold: 0.25
        groups:
          escalate:
            - biology
            - business
            - computer science
            - economics
            - history
            - math
            - other
            - philosophy
            - psychology
          keep:
            - chemistry
            - engineering
            - health
            - law
            - physics
        metrics:
          - name: escalate_vs_keep
            type: group_margin
            positive_group: escalate
            negative_group: keep
    modules:
      prompt_compression:
        enabled: true
        # Built-in profile for scoring/preserve defaults: default, coding, medical,
        # security, or multi_turn. The multi-turn alias is normalized to multi_turn.
        # Unknown profiles fail config validation. Explicit weights override the profile.
        profile: default
        max_tokens: 4096
        min_length: 512
        skip_signals: [jailbreak, pii]
        textrank_weight: 0.4
        position_weight: 0.2
        tfidf_weight: 0.3
        novelty_weight: 0.1
        position_depth: 0.1
        preserve_first_n: 3
        preserve_last_n: 2
      prompt_guard:
        enabled: true
        model_ref: prompt_guard
        model_id: models/mmbert32k-jailbreak-detector-merged
        threshold: 0.7
        use_cpu: true
        use_modernbert: false
        use_mmbert_32k: true
        jailbreak_mapping_path: models/mmbert32k-jailbreak-detector-merged/jailbreak_type_mapping.json
        use_vllm: false
      classifier:
        domain:
          model_ref: domain_classifier
          model_id: models/mmbert32k-intent-classifier-merged
          threshold: 0.5
          use_cpu: true
          use_modernbert: false
          use_mmbert_32k: true
          category_mapping_path: models/mmbert32k-intent-classifier-merged/category_mapping.json
          fallback_category: other
        mcp:
          enabled: true
          transport_type: stdio
          command: python
          args: [-m, mcp_classifier]
          env:
            MCP_MODE: local
          url: http://localhost:9010
          tool_name: classify
          threshold: 0.6
          timeout_seconds: 15
        pii:
          model_ref: pii_classifier
          model_id: models/mmbert32k-pii-detector-merged
          threshold: 0.9
          use_cpu: true
          use_mmbert_32k: true
          pii_mapping_path: models/mmbert32k-pii-detector-merged/pii_type_mapping.json
        preference:
          use_contrastive: true
          embedding_model: models/preference-encoder
          prototype_scoring:
            enabled: true
            cluster_similarity_threshold: 0.9
            max_prototypes: 8
            best_weight: 0.75
            top_m: 2
            margin_threshold: 0.0
      complexity:
        prototype_scoring:
          enabled: true
          cluster_similarity_threshold: 0.9
          max_prototypes: 6
          best_weight: 0.75
          top_m: 2
          margin_threshold: 0.0
      hallucination_mitigation:
        enabled: true
        on_hallucination_detected: annotate
        fact_check:
          model_ref: fact_check_classifier
          model_id: models/mmbert32k-factcheck-classifier-merged
          threshold: 0.65
          use_cpu: true
          use_mmbert_32k: true
        detector:
          model_ref: hallucination_detector
          model_id: models/mom-halugate-detector
          threshold: 0.82
          use_cpu: true
          min_span_length: 2
          min_span_confidence: 0.6
          context_window_size: 50
          enable_nli_filtering: true
          nli_entailment_threshold: 0.75
        explainer:
          model_ref: hallucination_explainer
          model_id: models/mom-halugate-explainer
          threshold: 0.9
          use_cpu: true
      feedback_detector:
        enabled: true
        model_ref: feedback_detector
        model_id: models/mmbert32k-feedback-detector-merged
        threshold: 0.7
        use_cpu: true
        use_modernbert: false
        use_mmbert_32k: true
        feedback_mapping_path: models/mmbert32k-feedback-detector-merged/feedback_mapping.json
      modality_detector:
        enabled: true
        prompt_prefixes:
          - generate an image of
          - draw
          - create a picture of
        method: hybrid
        classifier:
          model_path: models/mmbert32k-modality-router-merged
          use_cpu: true
        keywords:
          - image
          - picture
          - illustration
        both_keywords:
          - explain and illustrate
          - describe with an image
        confidence_threshold: 0.7
        lower_threshold_ratio: 0.7